[{"content":"Most of my work runs through AI now, and Pi is basically always open in my terminal.\nIt just feels good to use: fast, no waffling, you say one thing and it moves, rarely spinning in circles. It\u0026rsquo;s noticeably smoother than the native agents. But feel is hard to pin down. I couldn\u0026rsquo;t tell you where exactly it\u0026rsquo;s better, or by how much.\nThen I saw this Databricks benchmark. They paired the major models with the major CLI tools, two by two, and ran the whole matrix against their own multi-million-line production codebase. For the same completed tasks, Pi burned noticeably fewer tokens.\nThe moment the numbers landed, my vague hunch finally had something solid under it.\nThis piece covers three things:\nwhat the benchmark actually shows; why Pi is so cheap for what you get; how to pick the right agent for yourself. 1. What Pi Is First, what Pi is. It\u0026rsquo;s an open-source command-line coding tool, a project started in August 2025, now sitting at 70k+ stars.\nIt does roughly what Claude Code and Codex do. But take the same model and wrap it in a different agent, and the results can swing a lot, because that outer shell decides how much gets sent to the model each turn, which tools the model gets, how it judges whether a task is done, and so on.\nAnd as an independent open-source project, not being locked to any one model is a real advantage too.\n2. The Verdict: Pi \u0026gt; Claude Code/Codex On July 8, 2026, Databricks published an internal benchmark report. Among the authors is Matei Zaharia, the creator of Spark and the company\u0026rsquo;s current CTO. They paired the major models with the major CLI tools, two by two, ran the whole thing against their own multi-million-line production codebase, and plotted an agent-by-model cost-to-quality chart.\nOf the 7 top-scoring model-agent combos on the chart, 4 are Pi. The single highest pass rate, 90%, is also Pi.\nIn other words, on real production code, an open-source project beat the official tools that Anthropic and OpenAI each built specifically for their own models.\nAnd it\u0026rsquo;s much cheaper. On the same Opus 4.8 at high effort, Claude Code averages $1.94 per task and Pi runs about $0.95, more than half off, with the pass rate differing by just 2 points.\nThe benchmark also makes another thing clear: a cheaper model isn\u0026rsquo;t automatically the better deal.\nSonnet 5 has a lower per-token price than Opus 4.8, yet its total cost per task is higher, burning 1.9x the tokens Opus does.\nWhat you should actually compare is \u0026ldquo;how much it costs to finish one task,\u0026rdquo; not \u0026ldquo;the sticker price per million tokens.\u0026rdquo;\n3. How the Data Was Collected Before you trust a score, it\u0026rsquo;s worth seeing how it was produced.\nNo more public test sets The usual references for coding ability today are public benchmarks like SWE-Bench and Terminal-Bench. Databricks says outright that it didn\u0026rsquo;t use them, and the reasoning is solid: the problems are public, the answers are online, and every lab sweeps that content into training.\nThe result is that a chunk of a model\u0026rsquo;s score on those problems comes from having seen them, not from being able to solve them. It\u0026rsquo;s the same as getting the exam questions before the exam.\nThere\u0026rsquo;s a second issue: tasks in public benchmarks don\u0026rsquo;t look much like what engineers actually write day to day.\nThe former is more like doing exercises. The latter is changing things inside a codebase that\u0026rsquo;s been running for years, carrying its own history and its own internal conventions.\nUsing their own merged PRs as the test set Databricks engineers merge thousands of code changes a day. They pulled from that pool with a filter: a change had to be self-contained, not sprawling across too many modules, and it had to come with a solid test suite.\nThen they threw out three categories: anything submitted by bots, anything from service accounts, and anything that was itself AI-generated.\nEvery surviving change got a human pass, boiling down what it was trying to do into a one-line task description, while stripping out any part of the original description that gave away the solution — leave it in, and the problem hands you the answer.\nThe codebase spans more than a dozen languages: Scala backend services, Rust at the systems layer, React plus TypeScript on the frontend, protobuf and gRPC interface definitions, Bazel build configs, all of it.\nGrading uses the test suite the PR originally shipped with. First, does the code compile; if it compiles, run the tests; all tests pass, it counts as passed. No second model was used as a judge.\nThat deserves its own sentence. Plenty of benchmarks these days take the easy road and let GPT or Claude decide whether another model did well, and that scoring method is shaky, because the judge misreads things and has its own preferences. A test suite is binary: it passed or it didn\u0026rsquo;t.\nThey patched the hole where agents peeked at git history The report admits something: early in the runs, some models scored \u0026ldquo;unrealistically high.\u0026rdquo;\nDigging through the execution logs, they found these models were going straight into the git commit history and pulling out the correct answer — every problem came from an already-merged commit, the git history in the workspace wasn\u0026rsquo;t isolated cleanly, so the answer was left sitting in the exam room.\nThey later plugged that hole, and I read this as a good sign.\nA benchmark report willing to write \u0026ldquo;we got it wrong at first\u0026rdquo; is more credible than one that only reports wins. It also tells you the test set was built and fixed as they went.\n4. Why Pi Is So Cheap for What You Get A lean system prompt Before Claude Code does anything, before it\u0026rsquo;s even your turn to speak, it has already sent the model 20,000-plus tokens.\nA token is the smallest unit a model uses to bill and process text. A Chinese character is roughly one to two tokens; a normal line of code, a dozen or so. Twenty-thousand-plus tokens is about a 30,000-to-40,000-word booklet.\nThis number was reverse-engineered by the community, not published officially — someone intercepted the requests and analyzed the full payload Claude Code sends each time. As of the February 2026 version, the system prompt itself is about 23,500 tokens, plus definitions for 27 built-in tools, and those definitions run close to 100,000 characters on their own. Together they form the fixed cost of every conversation.\nPi\u0026rsquo;s side: the system prompt is under 1,000 tokens, with 4 built-in tools — read file, write file, edit file, run command.\n23,500 vs 1,000. 27 tools vs 4.\nSomeone tested an extreme case: say \u0026ldquo;hello\u0026rdquo; through Claude Code, and those two words are worth a handful of tokens, but the actual request burned over 40,000.\nEven with caching, the lean system still wins outright At this point someone will say it\u0026rsquo;s not a problem — everyone has prompt caching now, and within a session everything after the second turn bills at a tenth.\nTrue, but only half true.\nCaching solves the billing problem of \u0026ldquo;resending identical content.\u0026rdquo; What it doesn\u0026rsquo;t solve is this: those 20,000-plus tokens genuinely occupy the model\u0026rsquo;s attention on every single turn.\nDatabricks\u0026rsquo; key finding is this: Pi sends the model about a third of the content the other tools do, per turn. That gap persists across the whole task, and the longer the task, the more files you touch, the denser the back-and-forth, the more it compounds.\nThere\u0026rsquo;s another spot that\u0026rsquo;s easy to miss: subtasks.\nClaude Code lets you hand work off to a separate subagent, but each subagent, on every one of its own turns, has to load the full system prompt and tool definitions again from scratch.\nThey measured one task: done directly, 121,000 tokens; split across two subagents, 513,000 tokens, more than 4x.\nA leaner prompt makes the model smarter This is the point I most want to nail down.\nLonger context makes the model dumber. There\u0026rsquo;s evidence for this.\nThe 2023 Stanford and Berkeley paper Lost in the Middle tested six model families and found that the same piece of information, placed at the start or end of the input, gets used accurately; placed in the middle, accuracy drops by more than 30%. The cause ties back to the model\u0026rsquo;s internal positional encoding — stretch the distance and attention naturally decays.\nIn 2025 Chroma ran a more real-world test across 18 models. The conclusion: as input grows, performance drops 20% to 50%. They specifically noted that a model advertising a 200k-token window already shows clear degradation by 50k tokens. And the decay isn\u0026rsquo;t linear: the more a task needs synthesis and the fuzzier the semantics, the faster it falls off.\nPut the two together: those extra 20,000 tokens you send every turn both charge you on the bill and dilute the model\u0026rsquo;s attention on the part that actually matters.\nSaving tokens and raising quality are, here, the same thing.\nAs models get smarter, the thinner the harness the better Pi\u0026rsquo;s author, Mario Zechner, said something to this effect: in a harness, how much is actually load-bearing, and how much is only there because everyone else does it that way? Pi answers that question by deleting.\nHe also said the tool should be small enough that you can hold every one of its tokens in your head.\nThat\u0026rsquo;s the opposite of how big companies think about products. Over the past year, Claude Code and Codex have kept piling on features — task management, team collaboration, all sorts of subsystems.\nThe system prompt I mentioned earlier grew 28% within two weeks in February 2026, and the growth was mostly the newly added task-management and collaboration modules.\nMore features isn\u0026rsquo;t a bad thing. But every feature has to claim a chunk of the system prompt to tell the model \u0026ldquo;you can also do this,\u0026rdquo; and that chunk gets paid for and takes up attention on every single turn — even when this particular task of yours has no use for it.\nThe conspiracy angle On Hacker News, someone put it more bluntly.\nA top comment: Anthropic can\u0026rsquo;t shake the suspicion of \u0026ldquo;inflating token consumption with its own tool,\u0026rdquo; because the data is right there, and Pi is more than twice as efficient at managing tokens.\nSomeone below added: the incentives simply don\u0026rsquo;t line up. The bigger the context Claude Code carries, the more Anthropic makes.\nI don\u0026rsquo;t want to state this as fact — the main reason big-company system prompts are long is genuinely that they have more features and more scenarios to cover, not necessarily pure profit motive.\nBut one thing is objectively true: neither Anthropic nor OpenAI has any incentive to \u0026ldquo;save users tokens.\u0026rdquo;\nWhich is probably why an independent open-source project does this better.\n5. So Does Pi Just Crush Claude Code/Codex Across the Board? Up to here it reads like a Pi landslide. But we should look at the full data.\nThat Databricks comparison chart has six groups, each one the same model at the same effort level, only the shell swapped. Here\u0026rsquo;s the complete data:\nModel + effort Pi Native tool Delta Cost Opus 4.8　high 85% 87% −2 Pi 2.08x cheaper Opus 4.8　xhigh 90% 88% +2 Pi 1.46x cheaper Opus 4.8　max 82% 89% not labeled Pi 1.20x cheaper GPT-5.5　medium 83% 80% +3 Pi 1.54x cheaper GPT-5.5　high 81% 83% −2 Pi 1.22x cheaper GPT-5.5　xhigh 78% 80% −1 Pi 1.44x cheaper Across the six, Pi wins two and loses four. But three of those four losses are within 2 points, while the cost advantage holds steady between 1.2x and 2.08x.\nDatabricks\u0026rsquo; summary of the chart: \u0026ldquo;significant cost difference, quality held constant.\u0026rdquo;\nStill, I see a few problems in it.\nAt high effort, Pi\u0026rsquo;s advantage vanishes Look at the third row: Opus 4.8 max, 82% vs 89%.\nSame model, only the effort dialed from xhigh to max, and Pi drops from the field-leading 90% to 82%, now losing to the native tool by 7 points.\nThat directly contradicts \u0026ldquo;quality held constant.\u0026rdquo;\nSeven points is a big gap in this data — enough to knock this combo from the top tier (82% to 90%) down to the second tier (71% to 82%).\nAnd this is the one row of the six with no delta labeled. The other five dutifully print \u0026ldquo;−2 pt\u0026rdquo; or \u0026ldquo;+3 pt\u0026rdquo;; only this one is blank. The body text doesn\u0026rsquo;t mention the data point either.\nI don\u0026rsquo;t think this is deliberate concealment — the chart and the numbers are all right there for anyone to count. But a data point that contradicts the core conclusion is sitting there without a word of explanation.\nAs for why, the plausible read is: at max effort, the model\u0026rsquo;s own reasoning gets very long, and Pi\u0026rsquo;s \u0026ldquo;send as little as possible\u0026rdquo; strategy may no longer be enough to keep up.\nAnd the feature gaps that come with minimalism Beyond the benchmark, Pi has some real gaps.\nNo permission system.\nThis is Pi\u0026rsquo;s own project docs: it builds in nothing to restrict file access, processes, network, or secrets, and by default runs with the full permissions of whoever launched it.\nWant boundaries, you wrap it in a container or a sandbox yourself; the docs give three ways to do it.\nBy contrast, Claude Code stops to ask you every time it edits a file or runs a command.\nFor a beginner, that confirmation has value.\nGenuinely fewer features.\nSubagents, MCP extensions, plan mode — most of what Claude Code has, Pi mostly doesn\u0026rsquo;t. The speed and cost you save come at the cost of fewer features.\n6. How to Choose Bottom line first: this isn\u0026rsquo;t a single-choice question. It depends on who you are.\nIf you write clear requirements, use Pi The criteria are concrete. If these describe you, switching to Pi likely pays off immediately:\nYou spell out which file to change, what behavior you want, and how edge cases should be handled Your project has a written spec doc (AGENTS.md or CLAUDE.md) laying out structure, commands, and conventions You\u0026rsquo;re on pay-as-you-go API billing, not a subscription Your task volume is high enough that cost already makes you anxious For this kind of person, the stuff those 20,000-plus tokens in Claude Code provide, you\u0026rsquo;ve already provided yourself. You\u0026rsquo;re paying twice for a manual you don\u0026rsquo;t need.\nThe Databricks data applies directly to this scenario: same work, bill cut by half or more, pass rate wobbling within a few points.\nTwo practical notes: don\u0026rsquo;t mindlessly crank effort to the max — that 82% vs 89% happened at the max tier, and in the Databricks data Pi does best at xhigh; also, Pi has no permission limits, so before running unfamiliar code or letting it execute anything destructive, wrap it in a container first.\nIf you\u0026rsquo;re new to AI, stay on Claude Code Same, with concrete criteria:\nYou\u0026rsquo;re not yet sure how to describe requirements to a tool like this, and often fire off a one-liner like \u0026ldquo;fix this problem for me\u0026rdquo; Your project has no docs written for the AI to read Claude Code is covered by the official subscription You can\u0026rsquo;t do without Subagent/PlanMode/Permission/MCP and the like For this kind of person, those 20,000-plus tokens are insurance, not waste.\nThey hold a large set of \u0026ldquo;here\u0026rsquo;s what to do in this situation\u0026rdquo; fallback rules — which files to look at when a request is vague, which commands to run before editing code, when to stop and ask the user. The terser your prompt, the more these rules fill in for you.\nPi\u0026rsquo;s minimalism has a prerequisite: it assumes you know what you want. That assumption doesn\u0026rsquo;t hold for beginners. In community feedback, \u0026ldquo;unreliable with low-quality prompts\u0026rdquo; is one of the most common complaints about Pi.\nAnd subscription users switching over have to pay separately for the API. The half-off only holds if you were paying by usage to begin with.\nThe middle path If you sit on both sides, split by task:\nWell-defined mechanical work — change an interface, batch-rename, fill in a module following an existing pattern — goes to Pi. This work is the cheapest and needs the least hand-holding. Fuzzy exploratory work — don\u0026rsquo;t know where the bug is, hunting for something in an unfamiliar codebase, need to understand before touching anything — stays with Claude Code, letting its fallback rules help you sharpen the question. Running both side by side doesn\u0026rsquo;t conflict.\nThe Last Word Will Pi stay ahead? I don\u0026rsquo;t know. Anthropic and OpenAI can trim their system prompts anytime; that\u0026rsquo;s an engineering problem, not a capability one.\nBut the thing this data reveals will keep holding:\nHalf of what you pay has nothing to do with the model itself. It has to do with how the agent wrapped around it manages context.\nBefore this benchmark came out, almost everyone\u0026rsquo;s attention was on the model — which one is smarter, which one scores higher.\nWhat the Databricks data shows is that the same model, on the same task, with a different agent around it, can cost twice as much.\n","permalink":"https://ryrenz.com/ai/pi-vs-claude-code-codex/","summary":"Databricks ran every major model paired with every major CLI coding tool against its own multi-million-line codebase. Of the 7 best cost-to-quality combos, 4 were Pi, and the single highest score, 90%, was also Pi. For the same work, Pi\u0026rsquo;s bill is often half of Claude Code\u0026rsquo;s. Here\u0026rsquo;s why the data holds up, why Pi is so cheap, and which one you should actually use.","title":"Pi \u003e Claude Code and Codex? Same Model, Half the Bill"},{"content":"Over the past month, the FDE (Forward Deployed Engineer) role has been everywhere: OpenAI poured $4 billion into a dedicated deployment company, Anthropic embedded engineers directly inside financial giant FIS, and Google Cloud is racing to hire hundreds.\nIn China, things are even more direct. ByteDance posted FDE roles for its Doubao product at 35,000 to 70,000 RMB per month (roughly $4,800 to $9,600) with 15 months of pay. Job listings at Ant Digital Technologies and Zhipu now carry the FDE label too.\nLinkedIn says FDE headcount has grown 42x in two years — three times faster than AI engineering roles.\nSo why the sudden feeding frenzy? Because models are good enough now. The next challenge is making money with them, in real businesses. AI companies have sunk enormous cost into model R\u0026amp;D, and the models themselves are ready. What comes next is plugging them into enterprises, solving actual problems, and generating returns. When people talk about enterprise AI transformation, the ones actually doing the work — wiring it up in the field — are FDEs.\nBut read the job descriptions, and all you get is a list of required skills and a vague sense of what the role involves. After the headlines, you still don\u0026rsquo;t know what this job looks like day to day, or how to break in.\nSo this article skips the conceptual framing and goes straight to practitioners who\u0026rsquo;ve done the work:\nA detailed retrospective from an engineer who spent eight years at Palantir Two interviews with OpenAI\u0026rsquo;s head of FDE A video self-portrait from an engineer who left Google to become an FDE A founder running all over Shenzhen An anonymous long-form post from someone who spent two years embedded in the power industry And the unfiltered complaints on Reddit, Blind, and Xiaohongshu From a practitioner\u0026rsquo;s perspective, five questions:\nWhat does an FDE do every day? How is it different in China vs. overseas? Who\u0026rsquo;s suited for it, and what skills do you need? How do you transition into FDE? Why is it the role closest to entrepreneurship? FDE Overseas: Grueling, Fast, and On-Site Four or Five Iteration Loops Per Week Nobody lays out the day-to-day more clearly than Nabeel Qureshi. He spent nearly eight years as an FDE at Palantir, wrote a widely circulated retrospective in 2024, then filled in the details on a podcast.\nHis week looked like this: Monday, walk into the client\u0026rsquo;s office for a meeting; Monday night, build the thing. Tuesday, show it to the client, collect feedback; Tuesday night, revise. Wednesday, show it again; Wednesday night, revise again. He\u0026rsquo;d run four or five cycles like that in a single week. \u0026ldquo;Six weeks later, you have something the client is willing to pay $20 million for.\u0026rdquo;\nYou Work at the Client\u0026rsquo;s Office, Almost All the Time Nabeel\u0026rsquo;s first major project was Airbus.\nHe moved to Toulouse, France for a year, working four days a week right next to the A350 final assembly line.\nWhat he built, in his own words, was \u0026ldquo;airplane-building task management software\u0026rdquo;: work orders, missing parts, and quality issues pulled together into one interface where you could check things off and see progress at a glance.\nLater, on a project for the National Institutes of Health, he had his own badge and sat side by side with civil servants, biologists, and clinicians every day.\nThe travel intensity was absurd.\n\u0026ldquo;Get a call, book a flight for the next morning, show up in some random city\u0026rdquo; — that was routine at Palantir. The culture was \u0026ldquo;get on the plane first, ask questions later.\u0026rdquo;\nMany FDEs became United Airlines top-tier frequent flyers, clocking over a hundred thousand kilometers a year.\nYou Don\u0026rsquo;t Code That Much — Most of Your Time Goes to Communication and Requirements In Palantir\u0026rsquo;s official 2015 recruiting video, FDEs themselves put the number at: on an ideal day, 40% to 50% coding, 20% handling assorted technical issues, and the rest spent digging through data to figure out what the problem actually is.\nGergely Orosz, the veteran engineer behind The Pragmatic Engineer, offered a more cynical breakdown: the real composition of the job is about 25% coding, 50% integration and data pipeline work, 25% meetings and calming down clients.\nAnu, an engineer who left Google to become an FDE, put it even more bluntly in her self-reported video: \u0026ldquo;Some days, communication is 80% of the job.\u0026rdquo;\nAt Google, she had uninterrupted blocks of coding time, and a product manager shielded her from clients. Now, \u0026ldquo;you\u0026rsquo;re the person sitting in the same room as the client.\u0026rdquo;\nAnother FDE, handling ten clients solo at a startup, wrote a weekly journal entry that began: \u0026ldquo;By the time I looked up, it was 6 p.m. and I hadn\u0026rsquo;t written a single line of code. This happens more often than I expected.\u0026rdquo;\nThe Unsexy Side of FDE Excel and Data Permission Hell The least filtered section of Nabeel\u0026rsquo;s retrospective is about data: \u0026ldquo;Data scattered across formats that are impossible to analyze directly — PDFs, notebooks, Excel files. God, so much Excel.\u0026rdquo;\nWhat was worse was the process of getting the data. \u0026ldquo;The client buys an 8- to 12-week pilot, and we spend the entire 8 to 12 weeks just getting data access. In the final week, we scramble to cobble together something that can be demoed.\u0026rdquo; What blocks the project isn\u0026rsquo;t the tech — it\u0026rsquo;s the internal silos on the client side.\nDebugging Live in Front of the Client Anu described a moment: something broke during a live demo, and she debugged it on the spot, thinking \u0026ldquo;this would be so embarrassing if I can\u0026rsquo;t fix it — they paid for this.\u0026rdquo; This is the biggest difference between FDE and a normal engineering role: there\u0026rsquo;s no \u0026ldquo;fix it in the next iteration\u0026rdquo; option. Today\u0026rsquo;s problem gets solved today.\nA Reddit commenter had an even more vivid version: he went hiking with a Palantir friend, and the friend opened his work laptop to fix a bug at the summit.\nThis Line of Work Eats Pain Palantir president Shyam Sankar has an internal catchphrase, relayed by a former FDE-turned-investor on a podcast: the FDE\u0026rsquo;s job is to \u0026ldquo;eat pain and excrete product.\u0026rdquo;\nA former Palantir FDE who later led FDE recruiting said one of their core screening criteria was \u0026ldquo;grit — a willingness to suffer\u0026rdquo;: \u0026ldquo;Being an FDE is painful. These people have to genuinely believe they can pull off the impossible.\u0026rdquo;\nHer own memory: spending weeks in a small German town with two colleagues, going to the client\u0026rsquo;s factory floor every day. \u0026ldquo;What you see on site and what\u0026rsquo;s in the contract — those are two completely different things.\u0026rdquo;\nA Day in the Life of a Chinese FDE: Running Across Shenzhen, Asking Questions During Water Breaks One Chinese FDE\u0026rsquo;s Schedule The most concrete picture of daily life in China comes from Lawted (a pseudonym from a media interview), based in Shenzhen. A former big-tech programmer, he got into Harvard but decided to defer — choosing to do FDE work as a founder instead. His reasoning: the next year or two is the best window to get into traditional-enterprise AI transformation.\nHis day:\nMorning: visit different companies to research whether their workflows are suitable for AI Afternoon: embed on-site and observe, catching staff during their water breaks to ask questions Evening: organize the day\u0026rsquo;s findings and build systems \u0026ldquo;I used to just sit in the office and wait for people to hand me requirements. Now I\u0026rsquo;m running all over Shenzhen, finding companies, finding bosses.\u0026rdquo;\nHis entry point was classic: a logistics company reached out. Out of their 40 to 50 employees, 20 to 30 spent every day manually pulling tracking numbers and addresses out of PDFs from clients and typing them into Excel.\nHe used AI coding tools to build a demo. What used to take a person several minutes per PDF, the AI parsed in seconds. The client signed a letter of intent on the spot.\nTwo Years of Field Lessons A power-industry practitioner on Zhihu wrote up the complete journey of three enterprise AI projects — the most grounded, no-bullshit firsthand account of AI on-site delivery on the Chinese internet. A few verbatim quotes:\n\u0026ldquo;After three enterprise AI projects, I discovered a brutal fact: 90% of clients have no idea what they want. You ask \u0026lsquo;what do you want to do with AI,\u0026rsquo; and nine out of ten say \u0026lsquo;something like ChatGPT.\u0026rsquo;\u0026rdquo;\nA smart Q\u0026amp;A system took six months from requirements to launch, with three mid-stream changes to the requirements.\nThe client\u0026rsquo;s scanned PDFs had no text layer — just extracting the text took three weeks. \u0026ldquo;The client didn\u0026rsquo;t understand why it was so slow: isn\u0026rsquo;t it just uploading a file?\u0026rdquo;\nHe got smarter over time: \u0026ldquo;During demos, I\u0026rsquo;d deliberately show some wrong answers. Let the client see that AI isn\u0026rsquo;t magic. The biggest enemy of enterprise AI adoption isn\u0026rsquo;t technology — it\u0026rsquo;s expectation management.\u0026rdquo;\nThe most deflating part came after launch: in a unit of over 600 people, fewer than 10 used the system daily.\nLater, when he embedded the AI into the OA approval workflow so it pushed results automatically, usage shot up 5x. His iron rule for pricing: data cleaning is at least 30% of the total cost, and operations at least 20%. These two line items are non-negotiable.\nHe also described the uniquely Chinese delivery environment: the client\u0026rsquo;s internal network isn\u0026rsquo;t connected to the internet. Drivers, CUDA, model weights — tens of gigabytes — all have to be downloaded in advance and carried into the server room on a hard drive.\nSalary and the Bar to Entry in China Lai Juncheng (pseudonym), who handles FDE recruiting at a Shanghai AI company, gave the most honest numbers: at his company, junior FDEs make 20,000 to 30,000 RMB monthly ($2,700 to $4,100). Senior roles are annual-package-based, 400,000 RMB and up. Million-RMB annual packages do exist, but they\u0026rsquo;re concentrated at the very top — not the industry average.\nFor reference at big tech: ByteDance\u0026rsquo;s Doubao FDE roles are listed at 35,000 to 70,000 RMB monthly, 15 months of pay. At the top end, that\u0026rsquo;s about 1.05 million RMB a year ($144,000).\nHis definition of the role is refreshingly stripped-down: in a software company, the FDE is \u0026ldquo;the majordomo who handles everything except coding and admin.\u0026rdquo;\nHe himself came up through implementation engineer, pre-sales, product manager, and project manager — and folded them all into one role.\nWhen hiring, he looks for two things: learning ability and the capacity to see through to the essence of a problem. He doesn\u0026rsquo;t filter by major: \u0026ldquo;People on our team who studied design or niche foreign languages have all done outstanding work.\u0026rdquo;\nBeware of Becoming On-Site Outsourcing To understand the Chinese FDE\u0026rsquo;s predicament, you first need to read an on-site account from 2021. Note: this was traditional IT on-site outsourcing from before the AI wave — not FDE — but it\u0026rsquo;s the pitfall Chinese FDEs fear sliding into most.\nThe engineer wrote: at the client site, you had to avoid the senior leadership at the cafeteria, because one leader\u0026rsquo;s exact words were \u0026ldquo;how come the migrant workers are allowed to eat here too.\u0026rdquo;\nEveryone in the client\u0026rsquo;s IT department, from the director down to the contract workers, wanted to offload some of their own work onto the on-site staff.\nThe office was on the fourth floor. The water dispenser was on the first floor. There was an electric kettle on the fourth floor — but it was for client employees only. He eventually took a 50% pay cut just to escape the on-site role.\nToday, people doing AI deployment on Xiaohongshu are still arguing about the exact same thing.\nOne post title is a straight-up accusation: \u0026ldquo;Aren\u0026rsquo;t 99% of FDEs just outsourced labor and paid hand-holding?\u0026rdquo;\nDoing consulting for clients, building custom workflows, teaching clients how to use AI — \u0026ldquo;let\u0026rsquo;s be honest, it\u0026rsquo;s just repackaging the same old thing in a fancier new term.\u0026rdquo;\nChina vs. Global: Side by Side The Trap for Chinese FDEs Put the firsthand accounts from both sides next to each other, and the differences jump out.\nHere\u0026rsquo;s the overseas benchmark: OpenAI\u0026rsquo;s FDEs built a customer service system for fintech company Klarna, then distilled the solution into an internal framework. They open-sourced it, and it eventually became OpenAI\u0026rsquo;s official Agents SDK — now the standard tool every FDE uses.\nThere\u0026rsquo;s an even better example: a 2,000-person Japanese sales team wanted a slide-generation assistant. The initial slide layouts the model produced were atrocious. The FDE tested 50 different approaches, packaged up the best samples, and handed them to the model training team. Three months later, a new model version shipped, and \u0026ldquo;the slides suddenly got good.\u0026rdquo;\nMeanwhile, here\u0026rsquo;s the self-deprecating take from Chinese practitioners:\n\u0026ldquo;We don\u0026rsquo;t call this new paradigm FDE. We call it: pre-sales + account manager + implementation + tech support + dedicated hand-holding + temporary product manager.\u0026rdquo;\n\u0026ldquo;Overseas, they\u0026rsquo;re just now hyping the idea of using demos to find direction. We\u0026rsquo;ve already demo\u0026rsquo;d ourselves bald.\u0026rdquo;\nThat\u0026rsquo;s the dividing line: does the road you paved at the client site eventually become part of your company\u0026rsquo;s product?\nOverseas, FDE output feeds back into the product and the model. In China, most on-site output stays with that one client — and starts decaying the moment you leave.\nThere\u0026rsquo;s another layer of difference, hidden in where trust comes from.\nA practitioner who\u0026rsquo;s done AI deployment on both sides wrote on Xiaohongshu: in the US, trust is contract-based. Sign the contract, and core data is opened to you. In China, trust is relationship-based. \u0026ldquo;You have to get drunk together first\u0026rdquo; — build personal rapport before you can get access to anything real.\nSo for the same on-site work, Chinese FDEs have one extra layer of labor: infiltrating the organization, building personal connections, punching through silos.\nThe Salary Gap Overseas: LinkedIn data shows FDE headcount grew 42x in two years.\nOn comp: Anu, who jumped from Google, said she didn\u0026rsquo;t take a pay cut. \u0026ldquo;FDE total comp is often higher at the same level — $200,000 is easy to hit.\u0026rdquo;\nA tech recruiter with 18 years in the game said two FDE roles she\u0026rsquo;s filling for the same hedge fund client are both at $350,000, and the client is willing to wait six months for the right person.\nChina: the title just landed. ByteDance, Ant Group, and Zhipu are using \u0026ldquo;FDE\u0026rdquo; directly in their job posts. There isn\u0026rsquo;t even an agreed-upon Chinese translation yet — three different terms are in circulation.\nThe growth is in AI transformation for traditional enterprises. Lawted\u0026rsquo;s math: \u0026ldquo;It used to take a year or two and cost millions of RMB to build an enterprise management system. Now with AI-assisted coding, we can deliver a custom system in two months.\u0026rdquo;\nThe government has stepped in too: Shanghai ran the country\u0026rsquo;s first FDE training program — one month of theory, two months of on-site practice — and is planning to tie it to intermediate professional certifications.\nCommon Misconceptions About FDE Five common misconceptions, each fact-checked with practitioner quotes.\nMisconception 1: FDEs are just pre-sales or product managers who can talk tech.\nYasha (pseudonym), who spent eight years as a PM in the US, then moved to development, then to FDE, put it in absolute terms: \u0026ldquo;An FDE cannot be a people-person product manager who just relays requirements back to programmers. That\u0026rsquo;s too slow and too lossy. You have to have real technical depth — make judgments yourself, solve problems on the spot yourself.\u0026rdquo;\nMisconception 2: FDE is a promotion from engineering.\nAnu made a point of correcting this: FDE is not a level above engineer. It\u0026rsquo;s a completely different job, a different growth curve — a lateral move to a different track. Her performance reviews also shifted from \u0026ldquo;code quality\u0026rdquo; to \u0026ldquo;client outcomes.\u0026rdquo;\nMisconception 3: FDEs mostly write code.\nSee the time breakdowns above: coding takes 25% to 50% of the time. Some days, communication takes 80%.\nMisconception 4: FDEs have to be on-site every day.\nThe counterexample is Ramp: their FDE team visits clients \u0026ldquo;once a quarter at most — everything is over video calls,\u0026rdquo; with one person serving five or six clients simultaneously.\nBeing on-site is a means. Staying close to the client to solve problems — that\u0026rsquo;s the definition.\nMisconception 5: China\u0026rsquo;s on-site delivery equals FDE.\nTwo differences: whether the output feeds back into the product, and whether you charge by outcomes or by headcount-days.\nIf you bill by person-day, whatever you call it, it\u0026rsquo;s outsourcing.\nWho\u0026rsquo;s Cut Out for It — and Who Isn\u0026rsquo;t The Right Fit: Common Threads The person who built Ramp\u0026rsquo;s FDE team has a line: \u0026ldquo;The FDE team is the team that wants to say yes to the client. A lot of engineers won\u0026rsquo;t admit it, but deep down they want to say no — they want to keep building their own thing.\u0026rdquo;\nOpenAI FDE lead Colin Jarvis offered this standard: \u0026ldquo;A relentless pursuit of value is the hallmark of a good FDE. The best FDEs are willing to rip apart what they just built and start over, because the client needs something different.\u0026rdquo;\nAnu\u0026rsquo;s self-deprecating version: \u0026ldquo;Someone like me, who\u0026rsquo;s borderline ADHD — this job actually works to your advantage.\u0026rdquo; People who switch context instantly and enjoy doing eight different things in a day thrive here.\nThe Wrong Fit: A Checklist The FDE handling ten clients solo wrote the most brutal account in his weekly journal: \u0026ldquo;If you need long uninterrupted stretches to produce anything, this role will destroy you.\u0026rdquo; \u0026ldquo;If you just want to build things and not talk to people, this is not your role.\u0026rdquo;\nA former Palantir training lead who transformed over 250 engineers into FDEs has a list of people to steer away:\nPeople allergic to ambiguous requirements People who want to go deep on technical specialization People who can\u0026rsquo;t handle client emotions And people who don\u0026rsquo;t know how to set boundaries That last type will burn themselves down to nothing.\nKnow the Risks Before You Switch One: skill atrophy. A self-identified former Palantir person on Reddit wrote: \u0026ldquo;On a bad project, all you do is drag-and-drop and configuration. Your technical skills atrophy beyond recognition.\u0026rdquo;\nAnother assessment is even harsher: do five years of FDE and then interview for a standard engineering role at a big tech company — you\u0026rsquo;ll likely get down-leveled.\nTwo: getting chained to a client.\nIt sounds like a good problem: you\u0026rsquo;re so good the client can\u0026rsquo;t live without you. But the founder of Dataland (ex-Palantir) pointed out the trap in a roundtable: \u0026ldquo;The client getting addicted to the FDE is even worse than your company getting addicted to the services revenue. You try to pull the person out, the client terminates your contract.\u0026rdquo;\nZoom into the individual level: you build the system, you\u0026rsquo;re supposed to move on to the next project. But the client only trusts you, and demands that you stay. The company can\u0026rsquo;t afford to lose the account, so you stay parked at that one client.\nSlowly, you go from \u0026ldquo;engineer exploring new problems\u0026rdquo; to \u0026ldquo;that one client\u0026rsquo;s dedicated ops person\u0026rdquo; — handling the same system\u0026rsquo;s day-to-day issues, not growing, with only one client on your resume.\nIn China, it gets weirder: clients trust through personal relationships, which means they trust you, not your company. The contract is billed by the day — every day you stay on-site, the company collects another day\u0026rsquo;s fee.\nThe client won\u0026rsquo;t release you. The company doesn\u0026rsquo;t want to swap you out. You can\u0026rsquo;t leave. Eventually, you become exactly what you feared: on-site outsourced labor.\nOpportunity: How to Get In Learn From Successful Transitions Line up the real transition paths from the firsthand accounts:\nEight years as a product manager, then development, then FDE — now being recruited at $400,000 a year (Yasha, US) Google engineer who made a lateral jump to FDE — four or five months later, confident enough to record a comparison video, no comp drop (Anu) Product design consultancy background, joined Palantir as an FDE (new hire from an official interview, four months in) Implementation engineer, pre-sales, product manager, project manager — all rolled into FDE, now runs recruiting (Lai Juncheng, Shanghai) Big-tech programmer who got an AI transformation request from a logistics company and jumped straight into founding a startup (Lawted, Shenzhen) Two common threads.\nOne: nobody starts fresh out of school. Yasha said: \u0026ldquo;A new grad can\u0026rsquo;t do this. You typically need a few years of PM or development experience first. That\u0026rsquo;s also why starting pay isn\u0026rsquo;t low — first year, around $200,000-plus.\u0026rdquo;\nTwo: everyone straddles tech and business. Whatever you\u0026rsquo;re missing, you fill in.\nWhat You\u0026rsquo;re Missing and How to Fill It The former Palantir training lead laid out a transition roadmap:\nMonth one: sit in on client calls, shadow the implementation team, and specifically document the gap between \u0026ldquo;what the client says the problem is\u0026rdquo; and \u0026ldquo;the real root cause\u0026rdquo; Month two: take end-to-end ownership of one client issue Month three: go to a client site and ship a small fix within the same week After that: write a weekly \u0026ldquo;what I learned from the client this week\u0026rdquo; note The skill list is surprisingly basic:\nSolid SQL Can read logs on a Linux command line Comfortable with Docker Python: can handle spreadsheets, make API calls; has a gut feel for dirty-data pitfalls like time zones, encodings, and null values No cutting-edge model techniques required.\nWhat the Interview Actually Tests Palantir has a signature \u0026ldquo;decomposition interview\u0026rdquo;:\nThey give you a vague requirement — something like \u0026ldquo;design a system that lets users share their interests\u0026rdquo; or \u0026ldquo;design a system for assigning cases to analysts based on expertise and availability.\u0026rdquo;\nThis is not the standard system design interview. One candidate ran through the textbook capacity-estimation-and-scalability template and was cut off mid-sentence by the interviewer, told to go back to \u0026ldquo;how the feature itself should be designed.\u0026rdquo;\nOpenAI\u0026rsquo;s process is even more direct: a roughly five-hour take-home assignment. Build something with their API, submit the code, plus a video of yourself explaining it.\nThe reasoning is clear: FDEs present to clients every day, so they test exactly that.\nRamp adds one extra round beyond the standard engineering interview — testing whether you can actually communicate.\nThe Role Closest to Entrepreneurship Let\u0026rsquo;s close with the most overlooked point in this whole thread.\nEveryone is talking about enterprise AI transformation. A lot of people talk about it. Very few actually go down and do it. FDEs are the ones who do: turning \u0026ldquo;transformation\u0026rdquo; from a buzzword into a quality inspection system on a production line, an agent in a customer service backend, an anti-money-laundering flow inside a bank.\nThis position has a byproduct that might be worth more than the paycheck: you live inside your clients\u0026rsquo; real pain points every day. Other people guess what the market needs to start a company. Your market needs walk up to you every morning.\nBob McGrew, a Palantir veteran who later became OpenAI\u0026rsquo;s Chief Research Officer, put it most directly:\nFDE training is, precisely, training to become a founder. At a New York FDE roundtable with four panelists, two were Palantir alumni who had gone on to start their own companies. One of them said it outright: \u0026ldquo;FDE is the best training ground for future founders.\u0026rdquo;\nThe data backs it up. Nabeel\u0026rsquo;s retrospective includes this stat: every Y Combinator batch has more former Palantir founders than former Google founders, even though Google has 50 times more employees than Palantir.\nYasha explained why: \u0026ldquo;FDEs have seen too many successes and too many failures. If you have FDE experience, you know exactly what the market is missing and what product you should build.\u0026rdquo;\nLawted in Shenzhen just flipped the sequence entirely: he entered FDE work directly as a founder. That first logistics client became his company\u0026rsquo;s starting point.\nHow Long Does This Role Last? Box CEO Aaron Levie said: \u0026ldquo;If I were doing college career counseling right now, the first thing I\u0026rsquo;d do is make sure students know this role exists and how to get an offer.\u0026rdquo;\nLai Juncheng, who runs FDE recruiting, was more level-headed: \u0026ldquo;Think of product managers when mobile internet took off ten years ago — white-hot for five to ten years, then gradually cooled down. But during that wave, you got outsized individuals like Zhang Xiaolong and Zhang Yiming emerging.\u0026rdquo;\nHonestly, how long the role lasts doesn\u0026rsquo;t matter that much.\nEnterprise AI transformation as a problem has unlimited runway, and it only gets bigger.\nSomeone who\u0026rsquo;s seen the real pain points and real problems of 100 companies — wherever the next opportunity is, they\u0026rsquo;ll have a sharper edge and a higher probability of seizing it.\n","permalink":"https://ryrenz.com/ai/fde-forward-deployed-engineer/","summary":"Drawing from a Palantir veteran\u0026rsquo;s 8-year retrospective, OpenAI FDE leads, a Shenzhen entrepreneur, and a power-industry deployment engineer, this article answers five questions from a practitioner\u0026rsquo;s perspective: what FDEs do daily, how China differs from overseas, who\u0026rsquo;s cut out for it and who isn\u0026rsquo;t, how to transition in, and why it\u0026rsquo;s the closest role to entrepreneurship.","title":"FDE: What It's Really Like, From Someone Who's Done It — China vs. Global"},{"content":"On June 9, 2026, Anthropic dropped two new models in a single day: Claude Fable 5 and Claude Mythos 5. But only one of them — Fable 5 — is actually available to regular users. Mythos 5 is not yet open to everyone. Add Claude Opus 4.8, released half a month earlier on May 28, and Anthropic now has three flagships on the table.\nSo what is the difference between these three?\nI read through Anthropic\u0026rsquo;s official launch blog posts, the System Card, the related papers, the engineering team\u0026rsquo;s technical discussions, Karpathy\u0026rsquo;s analysis, and a round of tech media coverage. Here are the six differences everyday users actually need to know.\nLet\u0026rsquo;s take them one at a time.\nFirst, get this straight: Fable 5 and Mythos 5 are the same model A lot of people think these are three separate new models sitting side by side. They\u0026rsquo;re not. Technically, the relationship is this: Fable 5 and Mythos 5 share the exact same model weights. The only difference is whether the external safety classifier is turned on. Opus 4.8 is a different, weaker model entirely.\nThis foundational relationship is the prerequisite for understanding the six differences that follow. So let\u0026rsquo;s clarify it first.\nThe opening section of the System Card puts it plainly:\n\u0026ldquo;Fable 5 is being released for general access — it has the same underlying model weights as Mythos 5, but has additional safeguards to prevent misuse for cybersecurity and biology.\u0026rdquo;\nTranslation: Fable 5 and Mythos 5 have identical underlying weights. The only distinction — Fable 5 has a layer of safety classifiers (safeguards) wrapped around it. Mythos 5 strips that layer away in certain domains.\nSo the real relationship between the three looks like this:\nMythos 5: The bare model. Highest capability, with safety guardrails removed in specific domains. Only available to Project Glasswing\u0026rsquo;s cybersecurity defenders, critical infrastructure operators, and a tiny number of biological research institutions. Regular users cannot access it. Fable 5: Same bare model, wrapped with three categories of safety classifiers. This is the version regular users can actually use today. Fully open access. Opus 4.8: A completely different, weaker model, released separately in late May. Its role in this system: when Fable 5\u0026rsquo;s safeguards get triggered and the model refuses to answer, the request is automatically downgraded to Opus 4.8 instead. We also need to sort out the timeline, or it gets confusing. Back in April, Anthropic already released a Mythos Preview, but only for internal testing by Project Glasswing partners. It was never open to the public.\nOn June 9, Anthropic upgraded Mythos to Mythos 5 (still only for the Glasswing crowd), and simultaneously took that same capability set, added safeguards to it, and released it as Fable 5 for everyone.\nSo for regular users, the only genuinely new thing you got on June 9 is Fable 5. Mythos 5 is just the upgrade to the April Preview — and you still can\u0026rsquo;t use it.\nThe names themselves encode this relationship: Mythos means \u0026ldquo;myth\u0026rdquo; and Fable means \u0026ldquo;fable.\u0026rdquo; Etymologically, both refer to \u0026ldquo;a story that is told.\u0026rdquo; Anthropic is effectively saying through its naming: these two are the same thing — one held back, one let loose.\nAn interesting detail: The Verge pressed Anthropic on the versioning. If this is Fable 5 and Mythos 5, where are Fable 1 through 4? Where are Mythos 1 through 4? And how does the \u0026ldquo;5\u0026rdquo; relate to the Mythos Preview from April? Anthropic declined to answer. So where this version number came from remains officially unexplained.\nDifference 1: A leap in long-task performance Anthropic\u0026rsquo;s capability ceiling used to be Opus. Not anymore.\nMythos is officially defined as an entirely new tier above Opus — the Mythos class. VentureBeat put it bluntly: Opus is no longer Anthropic\u0026rsquo;s top commercial capability layer. The Mythos class now sits above it.\nThe benchmarks are in the System Card. Here are the ones that tell the clearest story (bold = best in row):\nGeneral \u0026amp; Coding:\nBenchmark Mythos 5 Fable 5 Opus 4.8 GPT-5.5 SWE-bench Verified 95.5 95 88.6 — SWE-bench Pro 80.3 80 69.2 58.6 Terminal-Bench 2.1 88.0 84.3 82.7 83.4 Humanity\u0026rsquo;s Last Exam (no tools) 59.0 — 49.8 41.4 FrontierCode Diamond — 29.3 13.4 5.7 Look at the FrontierCode Diamond row: Fable 5 scores 29.3. Opus 4.8 scores 13.4. GPT-5.5 scores 5.7.\nswyx, one of the contributors to this benchmark, nailed the insight: on this leaderboard for extended-horizon problems, Opus 4.8 and GPT-5.5 don\u0026rsquo;t improve no matter how much compute you throw at them. Fable and Mythos are the first models whose post-training genuinely harnesses test-time compute to solve ultra-long tasks measured in \u0026ldquo;tens of human-hours equivalent.\u0026rdquo; That\u0026rsquo;s a first.\nBut the real-world cases mentioned in the official blog post tell the story better than any benchmark number:\nPayments company Stripe used Fable 5 to do a full codebase migration on a 50-million-line Ruby codebase — finished in one day. A human team would have taken over two months. Played the card game Slay the Spire with a file-based memory that continuously took notes. Its score improvement was 3x that of Opus 4.8, and it reached the final boss 3x more often. Playing Pokemon FireRed was even wilder: previous Claudes needed a complex scaffolding of maps, navigation tools, and hand-fed info to play at all. Fable 5 beat the game purely by looking at screenshots — pure visual input, zero extra information. It even worked autonomously for over a week on genomics research, assembling data from 138 species and millions of cells, then designing and training a model that outperformed a paper published in Science — while being 100x smaller. These aren\u0026rsquo;t demo stunts. These are real test results named in the launch blog.\nKarpathy (who joined Anthropic in May) gave an even harsher verdict. He called Fable 5 \u0026ldquo;a step-function jump worthy of a major version number, comparable in magnitude to the Claude 4.5 jump last November,\u0026rdquo; especially on long-horizon, hard problems:\n\u0026ldquo;You can give it a lot more ambitious tasks than what you\u0026rsquo;re used to, the model \u0026lsquo;gets it\u0026rsquo; and it will just go.\u0026rdquo;\nClaude Code core engineer Boris Cherny was more down-to-earth: \u0026ldquo;Fable is the best coding model I\u0026rsquo;ve ever used, by a wide margin.\u0026rdquo;\nAnthropic\u0026rsquo;s Head of Growth, Amol Avasare, offered the most intuitive framing in his launch thread: this is \u0026ldquo;the smartest model yet, by a big margin,\u0026rdquo; capable of running continuously for days, and \u0026ldquo;the longer the task, the bigger its lead over other models.\u0026rdquo; He also said this is the first model that feels \u0026ldquo;like a real colleague who can deliver — someone you can actually hand real work to.\u0026rdquo;\nOne sentence to sum up this difference: Opus 4.8 was the previous generation\u0026rsquo;s best. Mythos/Fable 5 built another floor on top of it. The longer and harder the task, the wider the gap.\nDifference 2: Cybersecurity capability with destructive potential — the real reason Mythos is locked up If you only look at general benchmarks, the gap is \u0026ldquo;noticeable but not scary.\u0026rdquo; What truly made Anthropic lock Mythos behind Glasswing and refuse to release it publicly is its offensive cyber capability.\nThe numbers in this area (all from the System Card\u0026rsquo;s cyber section, measuring unguarded Mythos 5):\nExploitBench (41 V8 engine vulnerabilities): Mythos 5 averaged 10.75 capability flags, with a 78% cap rate. Opus 4.8 managed only 5.56 flags, 40%. Firefox 147 exploit development (250 trials): Mythos 5 produced complete, working exploits 88.4% of the time. Opus 4.8: 8.8%. CyberGym (1,507 vulnerability reproduction tasks): Mythos 5 single-shot reproduction rate of 83.8%. Opus 4.8: 78.1%. That Firefox row is the headline: 88.4% versus 8.8%. An order-of-magnitude gap. This is why the April Mythos Preview shook the entire security world the moment it appeared. InfoQ reported at the time that it autonomously discovered \u0026ldquo;thousands of critical vulnerabilities\u0026rdquo; across all major operating systems and browsers. The oldest bug it found was a 27-year-old vulnerability in OpenBSD. Another was a 16-year-old hole in the FFmpeg H.264 decoder. An Anthropic engineer with no security background submitted a request before bed and woke up to a complete, working remote code execution exploit.\nIn Glasswing\u0026rsquo;s first month alone, partner systems yielded over 10,000 high/critical vulnerabilities. Independent security firms verified 90.6% as valid. Mozilla alone found 271 bugs in Firefox — more than 10x what previous testing had uncovered.\nSo here\u0026rsquo;s the difference: Opus 4.8\u0026rsquo;s offensive cyber capability is \u0026ldquo;useful but manageable.\u0026rdquo; Mythos 5\u0026rsquo;s is \u0026ldquo;enough to cause serious harm in the wrong hands.\u0026rdquo; That is the entire reason the Fable 5 safeguards exist.\nDifference 3: Safety safeguards — the only real difference between Fable and Mythos As covered earlier: Fable 5 = Mythos 5 + safeguards. Now let\u0026rsquo;s open up those safeguards and see what they actually block and how they work.\nThree blocked categories (confirmed in the official launch thread, point 4):\nCybersecurity: Full-chain agentic attacks, from reconnaissance to lateral movement to exploitation. Biology \u0026amp; Chemistry: Currently a \u0026ldquo;broad restriction\u0026rdquo; — they\u0026rsquo;d rather block too much than too little. Distillation: Preventing large-scale model capability extraction, especially by authoritarian states. How blocking works — a two-stage detection system:\nStage one: a probe scans Claude\u0026rsquo;s internal activations directly, doing an initial screen of all traffic. Stage two: flagged traffic is handed to an independently trained LLM classifier for review, which then decides whether to block. This classifier is trained using the Constitutional Classifiers method — a \u0026ldquo;constitution\u0026rdquo; written in natural language defines what can and cannot be answered, and synthetic data is generated from that to do the training.\nWhat happens after a block — this is where the Fable/Mythos difference shows up:\nIn Claude clients (web, desktop, mobile, Claude Code): when safeguards are triggered, the request is automatically downgraded to Opus 4.8. The UI tells you \u0026ldquo;this response is from Opus 4.8,\u0026rdquo; and you\u0026rsquo;re charged at Opus 4.8 pricing (not Fable pricing). In the Messages API: no automatic downgrade by default. The request is simply blocked, and a structured rejection reason is returned. Whether to downgrade is up to the developer to implement. One category is special — distillation safeguards are completely invisible to users. No \u0026ldquo;downgraded\u0026rdquo; notification appears. The model is quietly weakened through prompt modifications, steering vectors, or parameter-efficient fine-tuning (PEFT). This is estimated to affect only about 0.03% of traffic, concentrated in fewer than 0.1% of organizations.\nTrigger frequency: Officially, over 95% of Fable sessions in production never trigger any downgrade. The average trigger rate is under 5%. In Anthropic Head of Growth Amol Avasare\u0026rsquo;s plain English: \u0026ldquo;95% of the time, Fable 5 and Mythos 5 are the exact same experience.\u0026rdquo; This loops back to the earlier point: they are the same model underneath. That safeguard layer only stops you once every 20 interactions.\nBut there\u0026rsquo;s a discrepancy in the numbers worth flagging honestly. The System Card also states, in a different section, that in automated abuse audit traffic — deliberately designed to trigger safeguards — the proportion of downgrades to Opus \u0026ldquo;exceeds half of all conversation records.\u0026rdquo; These two numbers don\u0026rsquo;t contradict each other. One is real users\u0026rsquo; full traffic volume; the other is adversarial test traffic designed to step on every landmine. But you need both side by side for the full picture.\nAnd the safeguards are definitely on the sensitive side right now. Karpathy himself acknowledged that the initial release was tuned \u0026ldquo;a bit too trigger-happy.\u0026rdquo;\nReddit already has posts from people getting downgraded to Opus 4.8 while doing math problems, asking \u0026ldquo;why is Fable getting routed away for basic math?\u0026rdquo; Anthropic has acknowledged the false-positive rate is high and says they\u0026rsquo;ll keep tuning.\nSo how solid are these safeguards, really? The official numbers: over 1,000 hours of internal and external bug bounty testing found no \u0026ldquo;universal jailbreak\u0026rdquo; that could fully disable the guardrails. An external organization tested Fable 5 against harmful cyber requests and found it the most resistant of all models evaluated — even 30 known jailbreak techniques couldn\u0026rsquo;t break through. But the official blog also honestly added: the UK AI Safety Institute (UK AISI), within a very short testing window, has already made progress toward a universal jailbreak. The guardrails are strong, but they are not impenetrable.\nDifference 4: Official safety ratings Anthropic has a Responsible Scaling Policy (RSP) that every model release must pass through. The results here are the most authoritative lens for understanding the risk differences among the three.\nFor Mythos 5 (bare model):\nBio (CB-1): Assessed as possessing CB-1 capability — able to provide specific, actionable information to someone with only basic STEM education, saving them significant time they\u0026rsquo;d otherwise need to spend consulting domain experts. Bio (CB-2, novel weapons): Assessed as not crossing the threshold, but the System Card explicitly calls this \u0026ldquo;the most uncertain judgment we\u0026rsquo;ve made for any model\u0026rdquo; and acknowledges Mythos 5 may significantly accelerate a team of trained experts attempting to develop novel biological weapons. In the official words: \u0026ldquo;we think that world-class human expert substitution may now be possible in a few areas.\u0026rdquo; Cyber (Cyber FCF Tier 1): Can provide meaningful technical assistance for known attack techniques, but still requires human intervention for large-scale operations. Has not reached fully autonomous Tier 2. Autonomy: Assessed as not reaching the dangerous threshold of \u0026ldquo;sustained 2x acceleration of AI R\u0026amp;D,\u0026rdquo; a finding confirmed by external METR testing. But the risk is higher than any previous model. Anthropic gave one example that makes this double-edged nature concrete. They tested Mythos 5 on a key step in designing adeno-associated virus (AAV) — a vector normally used to deliver gene therapy, but the same capability, aimed the other way, could design dangerous viruses. Mythos 5, with no specialized training, using pure biological reasoning, outperformed specialized protein-design AI models at predicting how viral capsids assemble. This is both evidence that it can accelerate gene therapy research, and the exact reason the bio safeguards block first and ask questions later.\nFor Fable 5 (with safeguards): in limited evaluations, its resistance to misuse is overall better than any recent publicly released model.\nFor Opus 4.8: as the downgrade target, it\u0026rsquo;s positioned as \u0026ldquo;capable enough, low enough risk.\u0026rdquo; CyberScoop\u0026rsquo;s comparison numbers make this clear — Opus 4.8 without safeguards reproduces about 80% of known open-source vulnerabilities. With safeguards: drops straight to 1%.\nSo Anthropic\u0026rsquo;s own risk ordering is clear: Mythos 5 \u0026gt; Fable 5 \u0026gt; Opus 4.8. And the only reason Fable 5 can be released to the public at all is that layer of safeguards pressing its risk profile back down to near-Opus levels.\nDifference 5: Price — Fable costs twice as much as Opus Pricing is the easiest column to misremember, so let\u0026rsquo;s isolate it.\nFable 5 / Mythos 5: $10/million input tokens, $50/million output tokens. Opus 4.8: Standard mode $5/$25 — half of Fable. Its fast mode (2.5x speed) is what hits $10/$50. April Mythos Preview: $20–25/$100–125, more than double today\u0026rsquo;s Fable/Mythos 5 pricing. In other words: using Fable 5 today costs exactly twice as much per token as standard Opus 4.8. TechCrunch pointed out directly: that high price tag may itself be an invisible abuse barrier, given that many enterprises have already blown past their annual AI budgets. Meanwhile, Fable/Mythos 5 coming in at less than half the April Preview price is Anthropic delivering on its promise to \u0026ldquo;deploy Mythos-class models at scale.\u0026rdquo;\nBut there\u0026rsquo;s a cost more important than the price hike — one that\u0026rsquo;s easiest to miss and matters more for enterprises. Zero-retention data? Gone.\nAll Mythos-class traffic (including Fable 5) carries a mandatory 30-day retention period. Even enterprise customers who previously signed zero-retention agreements with Anthropic — where not a single piece of data is stored — get no exception. Anthropic explained this is a hard requirement for the safety classifiers to function: they need real traffic to identify new jailbreak variants and bring down false-positive rates. They commit that the data will not be used for training and will be almost entirely deleted after 30 days.\nBut for compliance-sensitive industries — finance, legal, healthcare — this is a hard blocker. What you used to pay extra for (\u0026ldquo;my data leaves zero trace\u0026rdquo;) is simply not achievable with Fable/Mythos. GitHub Copilot and Microsoft Foundry both flagged this separately for their customers during integration. TechCrunch went as far as to suggest this could set an industry precedent: want the stronger model? Trade your data retention for it.\nDifference 6: After June 23, your Max plan stops including Fable 5 for free \u0026ldquo;Open to everyone starting today\u0026rdquo; sounds great. The real catch is buried in the subscription plans, and it\u0026rsquo;s worth its own section.\nFable 5 subscription availability comes in two phases:\nJune 9 through June 22: Pro, Max, Team, and per-seat Enterprise plans include Fable 5 for free. Use it as much as you want. Starting June 23: Anthropic pulls Fable 5 from these subscription plans. To keep using it, you\u0026rsquo;ll need to purchase usage credits (pay-as-you-go metered billing) separately. Let me put this as bluntly as possible, because a lot of people will be asking: Yes, after June 23, your Max plan no longer gives you free Fable 5. If you want to keep using it, you pay through usage credits — separate, metered billing. Amol calls this \u0026ldquo;Extra Usage\u0026rdquo; in his thread; same thing.\nAnthropic\u0026rsquo;s Head of Growth, Amol Avasare, was unusually candid about this in his launch-day thread, almost walking through the \u0026ldquo;why\u0026rdquo; sentence by sentence:\n\u0026ldquo;We\u0026rsquo;re giving everyone Fable 5 inside their subscription allowance for two weeks, and then we\u0026rsquo;re taking it away.\u0026rdquo;\nHis stated reasons: demand will be enormous, and compute isn\u0026rsquo;t deployed fast enough yet. \u0026ldquo;We can\u0026rsquo;t guarantee we can meet Pro and Max demand a few weeks from now. We don\u0026rsquo;t want to over-promise, give too much,, then have to slash your rate limits.\u0026rdquo; So these two weeks (through June 22) are \u0026ldquo;the most we can give.\u0026rdquo; If they get lucky, they\u0026rsquo;ll extend week by week — but no promises. After the window closes, Fable 5 is still accessible through Extra Usage; it just no longer counts toward your subscription allowance. Once enough compute is deployed, it\u0026rsquo;ll go back into the standard plan. In his words: \u0026ldquo;we\u0026rsquo;re scrambling to make it happen.\u0026rdquo;\nThe other two are simpler:\nMythos 5: Regular users can\u0026rsquo;t touch it. Only for Glasswing partners and a tiny number of bio research institutions. Future access will roll out gradually through a more systematic trusted-access program. Opus 4.8: Use as normal. And when you trigger a high-risk topic while using Fable and get downgraded, Opus 4.8 is what takes over behind the scenes, billed at Opus pricing. One last practical tip: Fable 5\u0026rsquo;s effort parameter defaults to high. Even at low or medium effort, it often outperforms previous models maxed out.\nClosing When you strip it down, these three flagships are a line Anthropic drew between capability and risk. Mythos is the full-power version reserved for major players. Fable is the public version with safeguards. Opus is the fallback.\nFor regular users, the only one you actually get is Fable 5: included in your subscription today, moving to Extra Usage billing after June 23, with automatic downgrades on sensitive topics.\nWhat\u0026rsquo;s worth a longer look is the timing. Four days before the launch, Anthropic publicly warned that AI is approaching recursive self-improvement and called for the entire industry to hit the brakes. On June 1, they confidentially filed their IPO prospectus — targeting a trillion-dollar valuation.\nWarn of danger while launching the most powerful model ever on a sprint toward IPO. This company\u0026rsquo;s deepest contradictions are now fully on display.\n","permalink":"https://ryrenz.com/ai/mythos-fable-opus-comparison/","summary":"What\u0026rsquo;s the real difference between Mythos 5, Fable 5, and Opus 4.8? After reading Anthropic\u0026rsquo;s official blog posts, the System Card, engineering team discussions, and Karpathy\u0026rsquo;s analysis, here are the six differences everyday users actually need to know.","title":"Mythos, Fable, Opus — The Six Differences That Actually Matter"},{"content":"AI moves too fast. Tech media all sounds the same and rarely has its own take.\nTwitter is too fragmented. You can scroll for a day and remember almost nothing.\nLong-form essays are hard to get through. Official blogs are basically product launches.\nThe only way a normal person can keep up with AI is to listen to the people building it talk — long interviews with industry leaders, internal-feeling conversations with frontline operators.\nWhy podcasts?\nSam Altman is not going to write a 5,000-word blog post telling you OpenAI\u0026rsquo;s real strategy, but he will sit in front of Dwarkesh\u0026rsquo;s microphone for three hours. Moonshot AI\u0026rsquo;s Yang Zhilin is not going to do a quick interview for 36Kr, but he will record a four-hour deep dive with Zhang Xiaojun. Karpathy is not going to write a tech blog explaining why he left Tesla, but he\u0026rsquo;ll say it in one line on No Priors. Written interviews get compressed, edited, rewritten. Podcasts are raw material. What you hear is what they\u0026rsquo;re actually thinking.\nPodcasts also fit the AI era perfectly. Commuting, washing dishes, running, cooking — these hours are already being wasted. Listening to a podcast turns them into free information. Five hours of weekly commute listening to three No Priors episodes will leave your read on Silicon Valley AI investing sharper than 90% of people who only read tech news.\nBelow are the five Chinese and five English podcasts I think are most worth your time.\nChinese Top 5 1. Zhang Xiaojun\u0026rsquo;s Business Talks (张小珺商业访谈录) Host: Zhang Xiaojun, former Caixin journalist. Produced by Language Is the World Studio.\nWhat she does is simple. She gets people like OpenAI\u0026rsquo;s Yao Shunyu, Anthropic\u0026rsquo;s Yao Shunyu, Moonshot AI\u0026rsquo;s Yang Zhilin, Manus\u0026rsquo;s Xiao Hong, and Luo Fuli in front of a microphone and talks to them for three to seven hours.\nWhy does this matter? Because no other channel can pull this off. A press release gives you five questions, 200 words each, all PR-polished. Talk for 200 minutes and every layer of polish wears through.\nHer journalism background is the key. The three Asian Publishing Awards weren\u0026rsquo;t an accident. Her follow-up pacing draws real material out of guests. Plenty of Chinese hosts do long AI interviews. Only she does them at the level of investigative journalism.\nEpisodes to start with: the four-hour interview with Yao Shunyu, the 3.5-hour interview with Luo Fuli.\n2. Crossing (十字路口) Host: Yang Yuancheng (Koji), partner at ZhenFund and founder of AI Hacker House.\nKoji\u0026rsquo;s position at ZhenFund means this podcast pulls first-hand interviews with early-stage AI founders that no one else can get. ZhenFund backed ByteDance, Zhihu, and Xiaohongshu in their earliest rounds — the volume of early-stage founders they\u0026rsquo;ve seen and the accuracy of their reads are top-tier in Chinese VC. Koji also runs AI Hacker House, so basically every new-generation AI founder passes through his network.\nListening to Crossing feels different from listening to OnBoard!. OnBoard! takes the Silicon Valley investor view and asks \u0026ldquo;can this company win?\u0026rdquo; Crossing takes the domestic founder-community view and asks \u0026ldquo;what are we all doing, where are we stuck?\u0026rdquo;\nThe recent Paperboy episode broke down the paradigm shift in AI agents really well. They even gave a Token Grant to an experimental project called YoYo Agent — being both an investor and a live experiment ground is unique in the Chinese podcast scene.\nEpisodes to start with: the Paperboy founder episode, and the one on Agent Harness.\n3. Silicon Valley 101 (硅谷101) Host: Jane Hongjun, founder and CEO of Silicon Valley 101, former US correspondent for Caijing magazine.\nHongjun\u0026rsquo;s advantage is that she actually lives in Silicon Valley. Five-plus years in, 213 episodes deep. Her interviews land one to two months ahead of what Chinese business media eventually picks up.\nHer guests are not popular science influencers. They are former TPU engineers, AI healthcare founders, NVIDIA insiders — people who can take \u0026ldquo;why did Google build its own TPU\u0026rdquo; all the way down to chip design specifics.\nThe China-US dual perspective is the other rare asset. When DeepSeek-R1 dropped, her episode was the first in Chinese to give a full read of what Silicon Valley engineers were actually saying about it in private group chats.\nShe doesn\u0026rsquo;t only cover AI. If what you want is a tech podcast rather than just an AI podcast, Silicon Valley 101 is the better entry point.\nEpisodes to start with: Episode 228 on whether Google\u0026rsquo;s TPU can shake NVIDIA, told by a former TPU engineer; Episode 224, a deep teardown of OpenClaw.\n4. 42 Chapters (42章经) Host: Qu Kai (Kyle), founder of 42 Capital and 42 Chapters.\nKyle is one of the earliest VCs in the Chinese AI scene to bet on the application layer of large models. In November 2024 he said \u0026ldquo;the earlier you dare to believe in AI, the bigger your potential return\u0026rdquo; — at a time when domestic AI applications hadn\u0026rsquo;t really gotten off the ground. Later, when Manus blew up, he did a whole series of interviews with the Manus founder.\n42 Chapters has a lighter rhythm than Zhang Xiaojun or Crossing. Episodes run 30 to 60 minutes, easy to finish in one commute. Kyle is an early-stage VC himself. The guests are mostly founders actually building products. The conversations focus on what happened in the AI application space this year and what new opportunities have opened up.\nIts edge is \u0026ldquo;judgment plus beginner-friendly.\u0026rdquo; In 2024 Kyle predicted that the foundation model roadmap was basically settled and future differentiation would happen at the application layer. That call kept getting validated through 2025. At the same time, the show is welcoming to people new to public and private markets — you don\u0026rsquo;t need to already know a stack of terms to follow along.\nTwo to four episodes a month, steady release schedule.\nEpisodes to start with: the Manus AI founder interview series, the November 2024 wrap-up on the AI application space.\n5. Slaying the Dragon (屠龙之术) Host: Zhuang Minghao, ex-big-tech, ex-VC.\nThis one is the odd one out in the top 5. It\u0026rsquo;s not an interview show. It\u0026rsquo;s a monologue. Zhuang Minghao alone takes the last three to six months of AI industry developments and compresses them into 70 to 130 slides, then narrates the deck. Listening to one episode is like reading a solid industry research report.\nHis calls are genuinely sharp. He predicted the LLM trajectory before GPT-3 hit. He flagged the limitations of AutoGPT before the agent concept became a buzzword. In Chinese AI commentary circles, a Zhihu user once said \u0026ldquo;this academician\u0026rsquo;s grasp of LLM development is weaker than Zhuang Minghao\u0026rsquo;s\u0026rdquo; — that half-provocative compliment reflects real respect inside the community.\nThe format has trade-offs. The upside is information density: he covers in half an hour what others spend three hours discussing. Updates run roughly once every two to four weeks.\nBut for someone who needs to quickly build a panoramic view of the AI industry, nothing else replaces it. Other podcasts give you a guest\u0026rsquo;s perspective. This one gives you an analyst\u0026rsquo;s.\nEpisodes to start with: Vol. 45, which uses 132 slides to cover the entire AI industry in 2025; Vol. 55, 70 slides on Manus and AI agents.\nBest for: investors, analysts, and operators making strategic calls who need a panoramic view fast.\nEnglish Top 5 1. Dwarkesh Podcast Host: Dwarkesh Patel, independent podcaster, named to TIME\u0026rsquo;s 2024 list of 100 most influential people in AI. Format: very long-form deep interviews, 3 to 5 hours per episode.\nWhy it\u0026rsquo;s worth listening to\nDwarkesh is the most prepared AI podcaster of this generation in English.\nBefore every episode he reads everything his guest has published — papers, internal reports, every public interview from the last five years — and walks in with 80 questions so specific the guest has no room to \u0026ldquo;do the rehearsed take first.\u0026rdquo; He only asks what you haven\u0026rsquo;t already said elsewhere.\nThe result is the ceiling of guest quality and depth in English. Dario Amodei has been on twice in a year (one of them specifically on AI 2027 scenarios). Demis Hassabis recorded right after the Gemini 3 launch. Mira Murati gave Dwarkesh her first long interview after leaving OpenAI to start Thinking Machines Lab. Karpathy and Ilya have both been on.\nOne Dwarkesh episode beats reading ten secondhand pieces about the same guest.\nOne caveat: his pace is fast and the episodes are dense. Listening on a commute can be tiring. Save him for quiet listening time.\nEpisodes to start with: the Carl Shulman episode (nearly six hours, widely considered required listening); the Dario Amodei episode.\n2. Latent Space Hosts: swyx (Shawn Wang, AI developer evangelist) and Alessio Fanelli (CTO in Residence at Decibel Partners). Format: built for AI engineers, covering agent frameworks, inference architecture, open-source model deployment, GPU economics.\nWhy it\u0026rsquo;s worth listening to\nThe deepest AI engineering podcast in English. Dwarkesh asks \u0026ldquo;will AI change the world?\u0026rdquo; Latent Space asks \u0026ldquo;what\u0026rsquo;s the fastest way to add a cache layer to Anthropic\u0026rsquo;s API next week?\u0026rdquo; The second question is closer to the actual day-to-day of an engineer.\nIts unique asset is swyx himself. He runs the \u0026ldquo;AI Engineer\u0026rdquo; conference and knows the CTOs of every AI startup in San Francisco. So guests are almost entirely frontline engineers from OpenAI, Anthropic, Meta, Databricks. Conversations go straight into \u0026ldquo;what model are we running, what token price, what latency, what gotchas.\u0026rdquo;\nEvery year-end they publish the \u0026ldquo;State of AI Engineering\u0026rdquo; industry report, the de facto annual benchmark in the AI engineering field.\nEpisodes to start with: the annual State of AI Engineering special; any episode on agent infrastructure.\n3. The Cognitive Revolution Host: Nathan Labenz, AI applications operator who personally participated in the GPT-4 internal red team. Format: AI applications, safety, and policy interviews, 2 to 3 hours per episode.\nWhy it\u0026rsquo;s worth listening to\nNathan\u0026rsquo;s unique edge is the GPT-4 red team — OpenAI brought him and a small group in to probe for vulnerabilities before the GPT-4 launch. That experience gave him first-hand knowledge of AI safety and capability frontiers, and lets him ask questions other hosts can\u0026rsquo;t.\nBut what really broke the show out was Nathan\u0026rsquo;s episode on using AI to help make medical decisions for his son\u0026rsquo;s cancer treatment. That episode pulled \u0026ldquo;AI applications\u0026rdquo; out of the engineering frame into \u0026ldquo;real-life decisions,\u0026rdquo; and is widely considered the breakthrough episode of the English AI podcast scene.\nHe\u0026rsquo;s done several follow-up episodes on the same theme — concrete decisions about AI in medicine, education, and family life.\nIf you want to understand AI not just as a tool but as a possible life companion, The Cognitive Revolution goes deepest on this in English.\nEpisodes to start with: Nathan\u0026rsquo;s series on his son\u0026rsquo;s cancer; the Bolt.new founder interview.\n4. No Priors Hosts: Sarah Guo (founder of Conviction VC, former Greylock partner) and Elad Gil (serial founder, investor in 40-plus unicorns). Format: VC-perspective AI startup interviews, 30 to 45 minutes per episode.\nWhy it\u0026rsquo;s worth listening to\nSarah Guo is one of the most active AI investors in Silicon Valley. Elad Gil has backed more than 40 unicorns. The interviews they do together are basically listening in on top-tier Silicon Valley VCs in a partner meeting — the questions they ask are at the level of \u0026ldquo;is this company worth a $50 million check?\u0026rdquo;\nThe guests match that level: Jensen Huang, Andrej Karpathy, Fei-Fei Li, Alexandr Wang, Bret Taylor. People at that altitude usually only share methodology in interviews, but Sarah and Elad get them down to specifics — the Jensen episode goes into how NVIDIA actually allocates GPU production internally.\nThe other unique advantage is pacing. 30 to 45 minutes per episode, much shorter than Dwarkesh. One commute fits one episode. If your only listening window is on the road, No Priors has the best density-to-length ratio in English.\nEpisodes to start with: the Jensen Huang interview; the Andrej Karpathy interview.\n5. Machine Learning Street Talk Hosts: Tim Scarfe (PhD in machine learning) and Keith Duggar (PhD in philosophy from MIT). Format: academic-leaning deep technical interviews. The only technical podcast that connects AI with cognitive science, neuroscience, and philosophy of consciousness.\nWhy it\u0026rsquo;s worth listening to\nMLST is the most academically dense show of these ten. With one ML PhD and one philosophy PhD, the conversational style feels closer to a top graduate research group than a typical interview show. Guests include Karl Friston (originator of the free energy principle), Stephen Wolfram, Yudkowsky, Francois Chollet — researcher-level figures.\nIt\u0026rsquo;s also one of the few English-language podcasts that openly takes a \u0026ldquo;LLM skeptic\u0026rdquo; position. Most AI podcasts debate \u0026ldquo;how far can LLMs go?\u0026rdquo; MLST often debates \u0026ldquo;is the LLM path itself wrong?\u0026rdquo; That stance creates a counterweight to the dominant \u0026ldquo;AI inevitably leads to AGI\u0026rdquo; narrative and is valuable for building your own independent judgment.\nNot the right show to start with. But after you\u0026rsquo;ve listened to nine AI podcasts and started feeling that everyone is saying the same thing, MLST will show you another possibility exists.\nEpisodes to start with: the Karl Friston episode on the free energy principle; the Blaise Agüera y Arcas episode on artificial life.\nA few that didn\u0026rsquo;t make the top 5 but are worth knowing These are for more specific audiences.\nChinese:\nPeople\u0026rsquo;s Park Talks AI (人民公园说 AI) — produced by JustSayAI. Three partners across a 12-hour time difference, a \u0026ldquo;companion-style\u0026rdquo; podcast where veterans riff on AI news. The companion product \u0026ldquo;JustSayAI Morning \u0026amp; Evening Briefing\u0026rdquo; is a paid subscription. For people who want to stay current on AI news without straining. 355 Loop (三五环) — hosted by Liu Fei. The main thread is city insiders chatting; AI is one of the topics. Liu Fei\u0026rsquo;s ability to get guests to say true things is rare in Chinese podcasts. For internet product managers and people who want a regular-user view of AI. Indie Hackers China (硬地骇客) — the king of Chinese indie developer podcasts, 100+ episodes in. Conversations cover indie hacking, remote work, side projects, real stories of building businesses to several million dollars in ARR, and a pragmatic take on AI. For people building indie products and wanting to use AI to run a real business. English:\nLex Fridman Podcast — close to 5 million YouTube subscribers. Less than half the content is AI (politics, physics, MMA all share airtime), but the Karpathy, Amodei, and Hassabis episodes hit the depth ceiling of English AI podcasts. Treat it as a one-off, episode-by-episode listen. Hard Fork — from the New York Times, hosted by Kevin Roose and Casey Newton. General tech news podcast with a weekly riff on AI industry events. Light, fast-paced, good for people who just want to stay current without working at it. TWIML AI Podcast — long-running, 700+ episodes. Content leans toward enterprise ML in production. When you need to look up \u0026ldquo;how did company X actually deploy ML?\u0026rdquo;, TWIML is an archive with no real equivalent in English. How to get into English podcasts A lot of people get stuck on the English barrier — can\u0026rsquo;t follow, can\u0026rsquo;t keep up, can\u0026rsquo;t find three hours for one episode. The methods below can multiply the return on English podcasts.\n1. Use them as English listening practice — two birds, one stone The guests on English AI podcasts are almost all native English speakers or researchers, engineers, and CEOs with strong English. Pronunciation is clear, vocabulary is standard, technical terms come up repeatedly. Far better than dedicated English-learning material, because you actually care about the content.\nStart with No Priors (30 to 45 minutes per episode, fast pace, two hosts going back and forth). Much easier to swallow than diving straight into a six-hour Dwarkesh. Skip the parts you can\u0026rsquo;t follow. Catching 60% overall is already a win. Three months in, you\u0026rsquo;ll notice your AI-related English reading speed has come up too.\n2. Use AI to compress English transcripts into Chinese summaries Almost every major English AI podcast publishes a written transcript (Dwarkesh, Latent Space, No Priors all post full text on Substack or their own sites). The flow is:\nFind the transcript (YouTube description, show site, Substack) or extract it with AI Paste the full text into Claude or GPT Prompt: something like \u0026ldquo;summarize this AI podcast transcript in Chinese, structured as 5 key insights, each with a guest quote and a one-line value judgment\u0026rdquo; Ten minutes to read the core of a three-hour episode. The trade-off is you lose the host\u0026rsquo;s follow-up rhythm and the guest\u0026rsquo;s micro-pauses while thinking.\n3. Use NotebookLM to turn English audio into a Chinese podcast Google\u0026rsquo;s NotebookLM now supports generating a Chinese two-host conversational podcast from any uploaded material (including audio transcripts). The flow is:\nGet the English transcript Upload to NotebookLM Generate the audio overview, select \u0026ldquo;Chinese\u0026rdquo; Out comes a 10 to 15 minute Chinese two-person conversation, a translated and lightly reworked version of the source What you hear is not machine-translated captions. It\u0026rsquo;s two AI hosts riffing in natural Chinese on the original material, pacing close to the casual VC chat of OnBoard!. Great for commute time when you want to \u0026ldquo;translate\u0026rdquo; a long English interview into a Chinese digest.\n4. Subscribe to the host\u0026rsquo;s Substack for the long-form summary The standard play for English podcasts now is audio + written distribution. Dwarkesh has 80,000 paid Substack subscribers, and the reason is that his Substack is not just transcripts. It includes hand-written episode summaries, edited key quotes, and the host\u0026rsquo;s own reflection notes.\nJust subscribe to the free Substack email lists for Dwarkesh, Latent Space, and The Cognitive Revolution. Even without listening to a single minute of audio, scanning the weekly emails will keep you on top of 70% of the industry\u0026rsquo;s core information.\n5. YouTube chapters and timestamp-jumping Most English podcasts mirror to YouTube. YouTube\u0026rsquo;s auto-generated English subtitles are good enough now, and combined with native chapter markers (many big podcasters add their own timestamps), you can:\nLook at the chapter list first, pick the 2-3 segments that interest you Jump to that timestamp and listen for 10 minutes Skip everything else A three-hour episode often takes 30 minutes to skim for the essentials. Much better return than listening straight through.\nClosing thought Don\u0026rsquo;t try to follow everything at once. Pick one Chinese and one English podcast that you find most interesting, build the habit, then expand.\nThe information quality and density of podcasts is much higher than tech media. Two hours of a top operator talking on a podcast does more for your understanding than reading ten tech articles or scrolling 100 tweets.\nPodcasts are a badly underrated information source for the AI era. Hope this list helps.\n","permalink":"https://ryrenz.com/ai/ai-podcasts-top-10/","summary":"AI moves too fast for tech media, which all sounds the same. Twitter is too fragmented to remember anything. Long-form articles take forever to get through. Official blogs are just product launches. The only way a normal person can keep up with AI is to listen to the people building it talk. Here are the five Chinese and five English podcasts I think are worth the time, plus five ways to actually finish the English ones if you\u0026rsquo;re not a native speaker.","title":"Top 10 AI Podcasts Worth Your Time: 5 Chinese, 5 English (Plus How to Crack English Ones)"},{"content":"Most tutorials that teach you to build a second brain in Obsidian never warn you about what happens when the vault gets too full: the brain crashes.\nClaude Code plus an Obsidian LLM wiki looks perfect on day one. But as the vault fills with blog posts, articles, papers, journals, docs, video scripts, and audio transcripts, Claude Code\u0026rsquo;s index-based search starts to struggle. Give it six months and you\u0026rsquo;ll watch search accuracy steadily slide.\nQMD, the tool this post is about, is the cleanest way to plug that gap. Once it\u0026rsquo;s installed, no matter how big your vault gets, Claude Code can still pinpoint the right content fast, give you accurate results, and cut a huge chunk of search token cost.\nThree things in this post:\nWhat it is: what a second brain is, plus a ten-minute walkthrough for beginners to set up the Obsidian + Claude Code combo. Why: why the second brain weakens once documents pile up, and why QMD is currently the best tool to fix it. How to do it: how to tell when it\u0026rsquo;s time to add QMD, and a step-by-step install. 1. The Second Brain: Obsidian + Claude Code If you don\u0026rsquo;t have a second brain yet, here\u0026rsquo;s a five-minute build.\nStep 1: download and install Claude Code (Codex, OpenClaw, Hermes, OpenCode, Cursor — any of them work). I\u0026rsquo;ll call all of them Claude Code below.\nStep 2: download and install Obsidian.\nStep 3: open Obsidian and create a new vault.\nStep 4: inside that folder, create a file called Claude.md and paste in Andrej Karpathy\u0026rsquo;s LLM wiki rules from this Gist.\nhttps://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f\nStep 5: open Claude Code in that folder and tell it:\nSet up this knowledge base for me according to the rules.\nIt will build out the whole directory structure on its own:\nraw/: raw material. Drop every article, conversation, and video transcript straight in here. wiki/: compiled notes, organized by topic or concept. index.md: a hand-written entry page that holds your knowledge map. log.md: a log of every change. How does the second brain actually work?\nIt first writes a summary of each article you put into raw, then pulls out the concepts inside it and creates a separate page for each one. Every person mentioned in the article gets a page. Every tool gets a page. Every method gets a page. Then it updates index.md and registers all the new pages there. From that point on, you just paste the wiki/index.md path into any AI, and it\u0026rsquo;s plugged into your knowledge base. Any question gets answered from your vault, not from a random web search. The killer feature: once you start dumping your own thoughts and daily reflections in, the AI starts to actually get you. Ask it to write something and it knows your voice. Ask it to weigh in on a decision and it knows your past. The longer you use it, the better it gets.\nAnd because lookups are tiered — the AI reads the index first to find relevant pages, then reads the actual content — it doesn\u0026rsquo;t have to swallow every note at once. Big token savings.\nAfter that, anything you come across — good articles, your own journals, half-baked ideas, voice memos, whatever — just drop it in raw. Say \u0026ldquo;ingest this for me,\u0026rdquo; and the wiki grows by another 5 to 10 pages.\n2. The wiki + index Setup Has Its Own Ceiling But this setup has a ceiling. Three specific failure modes.\nFailure 1: as the index gets longer, pinpointing accuracy drops.\nOnce the wiki system is running smoothly, wiki pages grow steadily. Under Karpathy\u0026rsquo;s schema, each raw file you ingest usually produces or updates 5 to 10 wiki pages. Run a vault for six months and hitting the hundreds is normal.\nAt that point index.md itself becomes hundreds of lines long. It\u0026rsquo;s the first thing the agent reads on every query, so every query loads those hundreds of lines into context. You don\u0026rsquo;t save tokens, and pinpointing accuracy drops — the more granular your categories, the more the agent gets stuck between seven or eight near-identical concept pages. What used to be a clean \u0026ldquo;read index → read wiki\u0026rdquo; two-hop turns into \u0026ldquo;read index → flip through 3-5 wiki pages, none of them right → go back to the index.\u0026rdquo;\nFailure 2: the index can\u0026rsquo;t help with cross-category semantic search.\nKarpathy\u0026rsquo;s schema splits the wiki into Concept, Entity, Synthesis, Self-analysis, Comparison, and other types, each filed by topic.\nBut your real queries usually cross categories. Say you want to dig up \u0026ldquo;how my mindset on indie hacking shifted\u0026rdquo; — that question might touch a self-analysis page (your reflection from two years ago), a synthesis page (your overall take on indie hacking), and some fragment buried in raw from a conversation with a friend. The index sorts by \u0026ldquo;what type of page this is,\u0026rdquo; not by \u0026ldquo;what concept shows up where.\u0026rdquo; Cross-category search is exactly where it falls down.\nClaude Code\u0026rsquo;s built-in grep doesn\u0026rsquo;t help either, because the literal phrase \u0026ldquo;mindset shift\u0026rdquo; might never appear once.\nFailure 3: the details inside raw vanish.\nThis is the sneakiest one, and the most damaging.\nUnder Karpathy\u0026rsquo;s SOP, ingesting raw produces synthesis pages in the wiki — concepts, entities, comparisons, self-analyses. That\u0026rsquo;s a deliberate trade-off: wiki pages only keep the topical skeleton, not the specific details, direct quotes, examples, and numbers from the raw source.\nA concrete example. You stuffed a 30-part AI course note series into raw, and ingest squeezed it into 2 wiki pages (a course map and a methods list). Those two pages tell the agent \u0026ldquo;here\u0026rsquo;s the overall framework and the core methods of this course,\u0026rdquo; but the debugging trick from lesson 5, the specific example from lesson 8, the verbatim line from lesson 14 — all of that lives in raw and never made it into the wiki.\nRaw might hold hundreds of files, and any one of them could hide some specific sentence you\u0026rsquo;ll want back later. But ingest flattens them all into wiki summaries, so the original text never enters the agent\u0026rsquo;s retrieval path. Unless you remember the filename and open raw directly, those details might as well not exist.\nThe root cause of all three problems isn\u0026rsquo;t \u0026ldquo;the wiki model is wrong.\u0026rdquo; It\u0026rsquo;s that the wiki model is great at compiled knowledge, but the whole vault still needs a layer of full-text semantic search to cover what didn\u0026rsquo;t get compiled into the wiki, the cross-category queries, and the cases where the index itself has bloated past the point of usefulness.\nAnd the fix can\u0026rsquo;t be \u0026ldquo;let the agent read all of raw\u0026rdquo; — that would blow tokens through the roof. What you need is a tool that pinpoints precisely, returns only the relevant snippets, and saves tokens on the side.\nIn 2026, the best tool for this is QMD.\n3. RAG Fills the Gap the Wiki Model Leaves What RAG Is RAG stands for Retrieval-Augmented Generation.\nThe whole idea, in one sentence: before you ask the AI a question, first pull the few most relevant passages from your knowledge base, then let the AI answer using those passages.\nThe LLM isn\u0026rsquo;t answering out of thin air. It\u0026rsquo;s answering based on \u0026ldquo;those passages we just pulled.\u0026rdquo; The facts, the quotes, the details in its answer all come from your vault.\nOne caveat: RAG is not a replacement for wiki + index. They solve different problems. Wiki + index handles \u0026ldquo;compile the things you already know are important into structured knowledge.\u0026rdquo; RAG handles \u0026ldquo;find the most relevant scattered snippets across the whole vault for the current query.\u0026rdquo; One compiles, one searches. They complement each other.\nGoing Further: Hybrid Retrieval Vector search finds content that \u0026ldquo;means the same thing but doesn\u0026rsquo;t share the words.\u0026rdquo; Search for \u0026ldquo;indie hacking mindset shift\u0026rdquo; and it can pull up a podcast note from your raw two years ago that says \u0026ldquo;indie hacking ends in boredom\u0026rdquo; — not one word in common, but the meanings converged.\nSo is vector search enough on its own?\nNo. Vector search has its own blind spot: it can\u0026rsquo;t do exact lookups.\nSearch for \u0026ldquo;qmd v2.1.0\u0026rdquo; and vector search might hand back content for v2.0 or v1.9, because they\u0026rsquo;re semantically too close. But what you wanted was that exact version number.\nThis is where BM25 (an exact-match algorithm based on keyword frequency) crushes it.\nThat\u0026rsquo;s why the only retrieval setup worth trusting is hybrid retrieval: BM25 for exact matches, vector search for semantics, then a small reranker model that resorts both streams by relevance and picks the best few.\nQMD does all of this.\n4. Why QMD Right Architecture, Real Results QMD\u0026rsquo;s retrieval pipeline follows the same playbook as Google search and Anthropic\u0026rsquo;s internal Claude retrieval system.\nA query comes in. A fine-tuned small model does query expansion. Every variant runs BM25 and vector search in parallel. All the results merge via RRF, then go through a qwen3-reranker for final reranking, then get position-weighted to surface the most relevant few.\nThis pipeline has been validated repeatedly in academic IR papers since 2020. QMD just packages it: what used to take tens of thousands of lines of code is now an install.\nFully Local, Free, Simple to Use This is the fundamental dividing line between personal RAG and enterprise RAG.\nEnterprise RAG burns money on two things: hosted vector databases (Pinecone, Weaviate) at hundreds of dollars a month, and embedding API calls (OpenAI text-embedding-3) billed by the token.\nQMD takes both of those local.\nAlibaba\u0026rsquo;s the biggest contributor here. Qwen3 is the foundation that lets Chinese-language users run commercial-grade retrieval locally. Without Qwen3, personal RAG for Chinese content would still be a year out.\nQMD is a daily-use utility at heart. It should be free, like grep or ripgrep, and as much a piece of infrastructure as your OS or your notes app.\nAnd It Saves You a Lot of Tokens A hidden upside of QMD: it slashes the token cost of Claude Code\u0026rsquo;s calls into your second brain.\nThe mechanism is simple. The old way, the agent uses grep to find things, then reads every hit in full — one Q\u0026amp;A easily burns 20,000 tokens. With qmd installed, the agent queries directly and gets back 3 to 5 of the most relevant snippets, a few hundred tokens each, under 1,000 total.\nA real number: Andrew Levine published the data from his 600+ note Obsidian vault — a single query used to burn 15,000 tokens; with qmd, the same query uses 500. 96% savings. He says it took him 5 minutes from download to working setup.\nIf Claude\u0026rsquo;s token bill is hurting (especially if you\u0026rsquo;re paying per-API instead of subscribed), this matters more than free does.\nEndorsed by Both Tobi and Karpathy QMD\u0026rsquo;s author is Tobias Lütke, founder and CEO of Shopify.\nHe went on a posting spree in March 2026 about tuning his query-expansion model before bed. The qmd repo on GitHub is now past 25,000 stars.\nA CEO running a company worth tens of billions, staying up late to write a personal markdown search tool, tells you one simple thing: he got fed up with the same problem. The notes he keeps hit the same retrieval ceiling yours and mine do, and he decided to fix it himself.\nIn the \u0026ldquo;Optional: CLI tools\u0026rdquo; section of his llm-wiki Gist, Karpathy writes:\n\u0026ldquo;A search engine over the wiki pages is the most obvious one — at small scale the index file is enough, but as the wiki grows you want proper search. qmd is a good option.\u0026rdquo;\nPlain English: when the wiki\u0026rsquo;s small, the index file is enough; once the wiki grows, you need real search — and qmd is a good pick.\n5. Installing QMD: One Prompt Does It cd into your second-brain folder, fire up Claude Code, paste this in:\nInstall qmd for me, hook up the current directory for retrieval, my notes are in Chinese.\nnpm install -g @tobilu/qmd — install the qmd CLI itself\nclaude plugin marketplace add tobi/qmd \u0026amp;\u0026amp; claude plugin install qmd@qmd — install the plugin (auto-registers the MCP server + skill)\nqmd collection add . --name brain — register the current folder as the knowledge base\nSwitch to the Qwen3-Embedding-0.6B multilingual model (mandatory for a Chinese vault)\nqmd embed -f — build the vector index\nThe first run downloads about 2.4GB of local models. Wait a few minutes. After that, every embedding runs locally on your machine — no internet, no API spend.\nWhy switch to Qwen3: QMD\u0026rsquo;s default embedding model is optimized for English. For a Chinese vault, switching to Qwen3-Embedding-0.6B is a massive lift in retrieval quality.\nNo English tutorial mentions this step.\nFrom there, Claude Code routes through qmd automatically when it queries your second brain. You do nothing. When you add notes, just run qmd update once in a while for an incremental reindex.\nClosing Should you install it right now? Karpathy gives the bar in the llm-wiki Gist:\n\u0026ldquo;The index pattern works great at roughly 100 files and a few hundred wiki pages. But as the wiki grows, you want proper search — qmd is a good option.\u0026rdquo;\nHold your own second brain up against that:\nAround a hundred files and a few hundred wiki pages: stick with index for now, revisit qmd later. Search results are noticeably weaker, or wiki count is approaching a thousand: add qmd as soon as you can. Better search and a smaller bill at the same time. If you want your second brain to keep working for the long haul, QMD is the single tool most worth your time.\n","permalink":"https://ryrenz.com/ai/second-brain-needs-qmd/","summary":"Most Obsidian-as-second-brain tutorials skip the part where it breaks: once the vault gets big, Claude Code\u0026rsquo;s index-based search starts to fall apart. QMD, a local RAG tool by Shopify CEO Tobi Lütke, fixes exactly that gap — hybrid retrieval (BM25 + vector), fully local, free, and a 96% drop in token use. This post covers when to install it and the one prompt that gets it running.","title":"Your Second Brain Won't Scale Without QMD"},{"content":"\nWhy Shenzhen?\nBecause Shenzhen is at the same time China\u0026rsquo;s cross-border e-commerce capital (number one on Amazon\u0026rsquo;s global city ranking, with 100,000+ sellers concentrated here), the country\u0026rsquo;s number-one foreign-trade city (4.5 trillion RMB in imports and exports in 2024, leading the nation 32 years running), the southern stronghold of the big internet companies (Tencent with 110,000 staff, Huawei\u0026rsquo;s Bantian campus with 40,000), and one of the densest cities in China for restaurants and medical aesthetics (Futian CBD does 3 billion RMB in retail sales per square kilometer, and Nanshan has more medical-aesthetics clinics than any other district nationwide).\nThat means the bosses and workers you\u0026rsquo;d serve in all five of these businesses are stacked right around you, and people here only care about one thing — making money.\nSelling services to ordinary people is hard. Selling to people who want to make money is easy — as long as your return beats their cost, you close the deal.\nThe five businesses below share one thing in common: any ordinary person can do them. No tech background, no capital, no team. Just use AI to multiply what one person can output, then sell that output to the anxious small-business owners who can\u0026rsquo;t use AI and to the anxious workers themselves.\nI\u0026rsquo;ll break each one down four ways: how you make money, where the profit lives, what the barrier is, and how an ordinary person actually starts.\n1. AI-Powered Operations for Local Merchants Bosses aren\u0026rsquo;t paying for AI. They\u0026rsquo;re paying for foot traffic.\nShenzhen restaurant growth in 2024 was a sad 1.3% — far below the national 7.9%. First-tier cities see 10–15% of restaurants close every month. Futian CBD pulls 2 million daily foot traffic, did 241.4 billion RMB in retail (third-highest district in China), and gets over 60% of its consumption at night — and stores in Futian still die. Nanshan opened 170 flagship restaurants in three years; most didn\u0026rsquo;t survive two.\nWhat keeps this crowd of restaurant owners, nail salons, medical aesthetics clinics, pet shops, and gym operators awake at night is one sentence: \u0026ldquo;Revenue dropped again this week.\u0026rdquo; They don\u0026rsquo;t have time to study AI. They don\u0026rsquo;t care which model you use. They care about exactly one thing: can you bring customers through my door?\nThe operations business is just selling foot traffic to bosses.\nHow you make money You handle everything for the boss on Douyin, Xiaohongshu, Video Account (WeChat Channels), and WeChat groups. Monthly delivery looks like this: 3–5 short videos a day (AI-written scripts, AI voiceover, Jianying (CapCut) edits), 1–2 live streams a week, comment and DM replies, plus funneling walk-in customers onto WeChat and into the boss\u0026rsquo;s own WeChat group. The whole \u0026ldquo;add WeChat + group management + repeat purchases from old customers\u0026rdquo; play has its own name in the trade — \u0026ldquo;private traffic.\u0026rdquo;\nThe boss pays ¥3,000–5,000/month, 12-month contract.\nOne person doesn\u0026rsquo;t run one store. AI multiplies single-person output by 5–10x, so you can run 8–15 stores at the same time.\nWhere the profit lives Going by published market rates:\nDouyin local-account operations: ¥3,000–5,000/month (industry baseline) Short-video package (20–30 videos/month): ¥5,000–8,000/month Full technical management: ¥8,000–15,000/month Full Douyin Enterprise Account management (live streams + ad spend included): around ¥80,000/month (ceiling) Costs: one computer, one person, a few tool subscriptions. Total annual cost under ¥5,000.\nShenzhen per-capita disposable income is 81,123 RMB, almost 2x the national average. Bosses here are several notches more willing to \u0026ldquo;pay for traffic\u0026rdquo; than bosses inland. A freelancer running 8 clients can clear ¥20,000+ a month without breaking a sweat. The single-person ceiling is 15–25 clients.\nWhat the barrier is You need a little Jianying (CapCut), a little Douyin and Xiaohongshu posting, the ability to follow what a boss actually says (you have to tell the difference between a beauty boss\u0026rsquo;s pain points and a restaurant boss\u0026rsquo;s pain points), and the willingness to go offline (your first 3 clients basically come from door-to-door visits and showing your face in boss WeChat groups).\nYou don\u0026rsquo;t need any tech, capital, team, or business license.\nWho fits: your mom ran a nail salon in the neighborhood, your dad ran a BBQ stall, you used to work as a medical-aesthetics consultant. People who already speak the insider lingo of local small bosses are the fastest to get traction. Another Shenzhen advantage: 65.9% of residents aren\u0026rsquo;t local hukou holders, and roughly 4 million Hunan natives live in Shenzhen — fellow-provincial boss circles are a built-in trust channel.\nWho shouldn\u0026rsquo;t touch this: pure internet people who\u0026rsquo;ve never dealt with an offline business and only talk to code. These bosses trust people they\u0026rsquo;ve sat across from in person — not people who DM them PDFs.\nHow an ordinary person starts Xiaoshi, a hairstylist from a third-tier city in Sichuan, started livestreaming in her sixth month after opening. In 20 days she sold ¥580,000 in hair packages by herself, with an 82% redemption rate. Lao Shen in Shanghai opened a home-decor operations shop in 2024. His first job was a villa sauna room. His first video pulled in ¥50,000 in sales, and within a week his second hit ¥150,000+.\nThe pattern here is \u0026ldquo;boss anxiety + industrialized content services.\u0026rdquo; Shenzhen has dozens of times the density of Sichuan, so this is even more repeatable.\nPick an industry you actually understand. Lanes worth picking in Shenzhen: Nanshan medical aesthetics (340 clinics, first among all Chinese districts), Futian restaurants (top 3 nationwide on the Must-Eat list), Luohu Dongmen beauty, education and training near Bantian\u0026rsquo;s single-occupant apartments in Longgang, pets in Bao\u0026rsquo;an. Pick one and go. Find your first client. Type \u0026ldquo;Shenzhen restaurant boss Douyin customer acquisition\u0026rdquo; into Douyin\u0026rsquo;s search bar, click the \u0026ldquo;users\u0026rdquo; tab, and look for verified enterprise accounts running group-buying links with a lot of \u0026ldquo;where\u0026rsquo;s the address\u0026rdquo; comments. The algorithm already filtered them for you. Shenzhen\u0026rsquo;s store-探店 activity ties with Guangzhou and Chongqing for top 3 nationally — almost every merchant has a live presence on Douyin. Keep your first DM under 100 characters. \u0026ldquo;I saw your group-buying post for your X Road location. I\u0026rsquo;ve done [data] for the same category. Is what bothers you most ticket size or footfall?\u0026rdquo; Don\u0026rsquo;t lead with a WeChat QR code. First job free. Pick one nearby shop you can visit daily, run it for 30 straight days, and screenshot every metric — footfall, DMs, group-buy redemptions. Build a case deck. Picture-heavy PDF, with the data on page one. Then float in local restaurant boss groups, beauty boss groups, and chamber-of-commerce groups. Hand out cases, not business cards. Price and sign. First 3 clients at cost — ¥2,000–3,000/month. From client #4 on, you charge the industry rate, ¥3,000–5,000/month. Shenzhen bosses are better at paying than inland bosses, but pin four things into the contract: monthly payment + monthly review, no serving competitors, who covers third-party fees, and what plan B looks like. The week you land your first hit, raise the price or expand the SKU. Lao Shen\u0026rsquo;s window between his first ¥50,000 video and his ¥150,000+ video a week later is exactly the renewal-negotiation moment. Either bump the price, or expand from short videos to \u0026ldquo;short videos + live + WeChat group\u0026rdquo; full package. Run an account matrix to dominate search. Aqiang in a second-tier city took 70% of the local \u0026ldquo;whole-house renovation\u0026rdquo; search results with his own accounts, generating 400 leads in one month. Run four account types in parallel: amateur recommender, marketing account, designer IP, sales account. Then split off long-tail keywords like \u0026ldquo;[neighborhood name] + whole-house renovation.\u0026rdquo; Shenzhen has thousands of high-end neighborhoods you can ride for long-tail traffic — Overseas Chinese Town, Xiangmihu, Qianhai Bay, Sea World. One person handles 10–15 accounts. 2. The Cross-Border AI Product-Image Factory Sellers know how to make product images. They just can\u0026rsquo;t keep up with how fast their rivals are pumping them out.\nA third of all Chinese cross-border sellers are in Shenzhen. Shenzhen tops Amazon\u0026rsquo;s global city ranking, with 150,000+ registered cross-border companies. Bantian crams 10,000 cross-border e-commerce companies into 28.51 square kilometers — half of all Shenzhen sellers. Dalang in Longhua has 700 apparel companies in its garment town. Xixiang in Bao\u0026rsquo;an and Huaqiangbei in Futian don\u0026rsquo;t even need explaining for electronics.\nThe real cost of an apparel photoshoot in Shenzhen: ¥100–120 per piece for a domestic model, ¥1,200–1,500 per hour for a foreign model, ¥750–2,000 for one full shoot, with monthly fixed budgets starting at ¥8,000–15,000. That math started collapsing in 2024. A drop-ship seller adding 200 SKUs a week needs a hero shot, alt images, detail-page assets, and short videos for every product — plus separate sets for white, Black, Asian, and Latin models. Traditional human photoshoots at ¥100–500 a shot, with 10+ days of turnaround, simply can\u0026rsquo;t keep up.\nHow you make money You batch-produce AI product images, model shots, scene shots, and short videos for cross-border sellers. They take what you ship and post it straight onto Amazon, Temu, SHEIN, or their own DTC site.\nTwo pricing modes. Pay per image: ¥15–50 each, retail. Monthly bundle: 300–500 images per month for ¥1,500–3,500 flat, on a 3–6 month contract.\nThe core profit logic: tools from ByteDance Seedance, Alibaba, and Kling let you batch-generate images at an actual cost of ¥0.1 per image. You quote ¥15–50 per image. The spread is the business.\nApparel is where the money is. Real-human apparel shoots are expensive, slow, and SKU-heavy, so AI model shots have the highest replacement value. Dalang in Longhua and Xixiang in Bao\u0026rsquo;an — the two cross-border apparel hubs — are the easiest niches to break into.\nWhere the profit lives Price tiers:\nDomestic batch tools (ByteDance Seedance and similar): $0.022 per image (about ¥0.1) Xiaohongshu freelance gigs: ¥15/image Self-run studios with custom work: ¥20–50/image AI product-image tool packages from the Mogujie founders\u0026rsquo; new venture: ¥298 / ¥598 / ¥1,598 (for 2,000 / 6,000 / 20,000 generation credits respectively) Zhubajie hero images: starting at ¥700 Zhubajie detail-page bundles: ¥1,000–2,000 Amazon-grade premium detail pages: $150–4,000 Per-image gross margin is over 99%. One studio can run 300–500 images a day. Hitting six-figure monthly revenue isn\u0026rsquo;t hard.\nReference for the top of this industry: a Shenzhen Nanshan company called Ganzhi Jiebu (registered at Guoxin Investment Building on Gaoxin South 7th Road) makes the AI product-image tools ZMO.AI and Creati. Hillhouse led their $8M Series A and $20M Series B; Creati hit $10M ARR. Shenzhen-listed cross-border brands — Anker, Aukey, Ugreen, Insta360, Hammer Power, Saviet — are all headquartered nearby. Fourteen to seventeen listed companies stack up into one complete supply chain.\nWhat the barrier is You need a little computer literacy (install a few tools, manage files in folders), aesthetic sense (the judgment to tell which image makes a customer click \u0026ldquo;Add to Cart\u0026rdquo; one more time), and the patience to wrestle with prompts (AI image generation means constantly tweaking your Chinese inputs — patience beats technical skill).\nYou don\u0026rsquo;t need English (your clients are Chinese sellers — they know English), cross-border ops experience (sellers know Amazon\u0026rsquo;s rules), professional photo gear, or any fine-arts training.\nWho fits: stay-at-home moms, design students, programmers who want a side gig, internet ops people who got laid off. Anyone with time, taste, and the patience to sit down with a tool. Shenzhen\u0026rsquo;s special edge: you can hop on the subway to Bantian, Longhua, or Bao\u0026rsquo;an and meet sellers face-to-face — that\u0026rsquo;s a layer of trust remote studios don\u0026rsquo;t have.\nWho shouldn\u0026rsquo;t touch this: people who have never done e-commerce and don\u0026rsquo;t even know what a Taobao hero image is. The metric in this business is \u0026ldquo;seller click-through rate,\u0026rdquo; not \u0026ldquo;how pretty the image looks.\u0026rdquo;\nHow an ordinary person starts Find templates to copy. Posts like \u0026ldquo;Zero cost! Upload your clothing photo and generate a model-wearing shot\u0026rdquo; and \u0026ldquo;AI e-commerce model swap side hustle, ¥300+/day.\u0026rdquo; Steal the post structure, the comment-section scripts, the Xianyu listing routing — all of it. Hook them with 5 images for ¥9.9. Push trial pricing to slight loss (cost is around ¥0.5–2 per image). The point isn\u0026rsquo;t profit. It\u0026rsquo;s getting the client\u0026rsquo;s product info and aesthetic preference on file. Pitch the monthly tiers on day three. ¥599 for 30 images, ¥1,499 for 100, ¥3,500 for 300. Slot a \u0026ldquo;dedicated AI model training: ¥999/model\u0026rdquo; into the top tier. Once you\u0026rsquo;ve trained a model face on their brand, switching suppliers becomes a real cost for them. Page one of your case deck is pure data. \u0026ldquo;Served 12 Amazon sellers, 8,000+ images delivered, average click-through on hero shots up 21%, average shoot cost reduction 92%.\u0026rdquo; Page two is the cost comparison: \u0026ldquo;Traditional shoot ¥1,500 vs our studio ¥50; 7–15 days vs 24–48 hours.\u0026rdquo; Cross-border sellers don\u0026rsquo;t care about beauty. They care about these two pages. Cold-start by attacking specific niches. The takeaway from Sister Lin in Guangzhou running plus-size women\u0026rsquo;s wear: chase plus-size, maternity, silver-haired senior wear, and pet apparel — categories where real models are the most expensive and hardest to find. AI replacement value is highest there. In Shenzhen, you can go straight to the Longhua Dalang apparel hub, Xixiang children\u0026rsquo;s wear in Bao\u0026rsquo;an, and Huaqiangbei electronics. Get into the Shenzhen cross-border circles offline. Offline circles have resources you cannot find online. Three-phase scale-up: 0→¥10K over 60 days: pull 5–8 small clients off Xianyu, Xiaohongshu, and forums ¥10K→¥30K over 180 days: convert 3–5 of them to monthly subscriptions and self-host open-source AI models to cut costs 30% ¥30K→¥50K+ over 360 days: either go deep on a single category or upgrade to AI runway videos (¥199 per 5–10 second clip) When DMing a cross-border seller, don\u0026rsquo;t open with \u0026ldquo;I can do AI images for you.\u0026rdquo; Open with: \u0026ldquo;I\u0026rsquo;ll cut your detail-page asset cost from ¥1,500 per SKU to ¥50, with 12 hero shots, 6 scene shots, and 3 detail-page assets shipped in 48 hours.\u0026rdquo;\n3. AI Export Lead Gen (Get Factory Bosses Their Orders) Factory bosses don\u0026rsquo;t lack capacity. They lack orders.\nShenzhen imports and exports hit 4.5 trillion RMB in 2024, leading China 32 years running. The Qianhai district alone did 700 billion RMB in trade, with 121 billion in cross-border e-commerce and 102% year-over-year growth. Qianhai packs in 10,000 sellers and 100 service providers. Shenzhen\u0026rsquo;s 16 first-class national ports are the most of any city in China; 70% of Shenzhen–Hong Kong freight goes by road; cross-border clearance is 3 minutes.\nBut Shenzhen is only one end. The actual factories are in the surrounding belt — 3,000 garment factories in Dongguan Humen, a trillion-RMB electronics industry in Dongguan Songshan Lake, 30,000 lighting factories in Zhongshan Guzhen (half the world\u0026rsquo;s lighting), Zhongshan Xiaolan locks holding 30% of the national market, the 100-billion RMB battery industry in Huizhou Daya Bay. These factory bosses survived for over a decade on Alibaba.com gold showcase listings. Now on-platform traffic keeps getting more expensive, and the average open rate on sales development emails dropped from 3.8% in 2019 to 0.4% in 2024.\nTheir order book runs dry next month. This crowd is the easiest to hit with one sentence: \u0026ldquo;I\u0026rsquo;ll pull 30 real customer inquiry emails for you every month. ¥3,000–5,000/month is no problem.\u0026rdquo;\nHow you make money You run the full pipeline — finding customers, sending emails, following up, collecting inquiries — for factory bosses.\nConcrete delivery: 200 precise overseas buyer leads pulled per month, 50 personalized sales development emails sent per month (not bulk templates), 4 follow-up emails per cycle (same day, day 3, day 7, day 14), and a month-end report on open rate, reply rate, and customer-source analysis.\n¥3,000–15,000/month, 6–12 month contract.\nThe spread is the studio\u0026rsquo;s lifeblood. The same overseas tool stack (Apollo + Smartlead + GPT) — factory bosses can\u0026rsquo;t buy, can\u0026rsquo;t read, and can\u0026rsquo;t configure it. You package it as a service. They pay happily.\nWhere the profit lives First, a definition: a customer inquiry email is one where an overseas buyer reaches out to ask \u0026ldquo;what\u0026rsquo;s your price?\u0026rdquo;, \u0026ldquo;can you send a sample?\u0026rdquo;, or \u0026ldquo;can you customize this?\u0026rdquo; Unlike a cold lead, it already comes with buying intent. It\u0026rsquo;s three to five steps from an order. These emails are the most valuable thing a factory boss can get — a few more per month means the production line keeps running.\nTool costs and pricing:\nYour tool stack: Apollo (overseas leads tool, $59/month) + Smartlead (email automation, $39/month) + ChatGPT Plus ($20/month) = around ¥850/month Your quote to the factory boss: ¥3,000–15,000/month Spread: ¥2,000–14,000/month, pure profit The local benchmark software is OKKI (Xiaoman Tech), headquartered at the China Resources Building in Nanshan, Shenzhen. Acquired by Alibaba in 2021, OKKI held 14.7% of China\u0026rsquo;s export-trade software market in 2024 (number one) with 40,000 paying clients. But that\u0026rsquo;s just a SaaS subscription — the boss still has to operate it himself. What you sell is \u0026ldquo;I\u0026rsquo;ll run this pipeline for you.\u0026rdquo; That\u0026rsquo;s exactly the slot in the market.\nReal client cases: Boss Wang in Qingdao used some export software and pulled his management efficiency up 30%. Xiao He, born in 2001, did ¥3M in fitness-equipment sales with 2 people and 1 AI. A new-energy motorcycle brand pulled 600+ inquiry emails on a DTC site in 3 months. A machinery factory in Chongqing pulled 566 inquiries in 6 months.\nWhat the barrier is You need English that lets you read replies (with GPT helping with tone — CET-4 level is fine), a little computer literacy (configuring accounts on a few overseas tools), some niche category knowledge (pick a vertical to go deep on — electric scooters, pet snacks, hardware fasteners), and a bit of sales instinct (reading what a buyer actually wants between the lines of their inquiry).\nYou don\u0026rsquo;t need: native English, prior export experience, an office, or a team (1 person + 1 AI really did pull ¥3M in sales).\nWho fits: people who worked at a foreign company for 1–2 years (they know email etiquette), CET-4+ English plus some e-commerce background, ex-Taobao customer service or sales people. Shenzhen\u0026rsquo;s extra edge: Hi-Tech Park + Futian CBD concentrate the highest density of fluent-English white-collar workers, and Qianhai packs in 10,000 cross-border trading firms — finding peers to learn from and clients to meet face-to-face is easier than anywhere inland.\nWho shouldn\u0026rsquo;t touch this: people who\u0026rsquo;ve never touched international business and freeze when an English email lands. This business is half tools, half human. The human half is being able to go back and forth with a customer in English.\nHow an ordinary person starts Cap tool spend at ¥850/month. Apollo Basic ($59, searchable across 250M overseas contacts) + Smartlead Basic ($39, automated email sequences) + ChatGPT Plus ($20, email tone polish). Don\u0026rsquo;t pile on anything else early on. Run free for week one to show the boss what you\u0026rsquo;ve got. Free-tier delivery: 100 customer profiles organized (website, decision-maker email, purchasing leads from the last 6 months), 30 emails sent, report delivered on day 7. Day 8, you quote ¥3,000/month. Rewrite four words in the subject line and watch open rates jump from 8% to 35%. Swap \u0026ldquo;We are a professional manufacturer of LED lights\u0026rdquo; for \u0026ldquo;Mike from {{your company}} (LED supplier), 30% cost savings for your lighting operations.\u0026rdquo; One valve factory dropped a buyer\u0026rsquo;s annual downtime number (420 hours) into the subject line and proposed cutting it to 265 hours. Open rate hit 58%. Three months later they signed three million-Euro orders. Four follow-up emails: day 0, day 3, day 7, day 14. Build the sequence in Smartlead and let it run. Hang out in the right circles. Post 1–2 screenshots a week of \u0026ldquo;I used AI to pull X inquiry emails for X industry\u0026rdquo; with real numbers. Factory bosses know each other. The ones who see the data come to you. Three-tier pricing. First client at ¥3,000/month (100 leads + 50 emails + monthly report); run 2–3 months and bank some anonymized cases. Second cohort at ¥5,000–8,000/month. Once you go deep on a vertical, ¥10,000–15,000/month. 4. AI Avatar Short Videos (and Live-Stream Clipping) Some bosses want a personal brand but won\u0026rsquo;t show their face. Live-stream teams want to slice one stream into dozens of short videos. AI avatars eat this gap.\nTwo definitions first. \u0026ldquo;Clipping\u0026rdquo; means taking a top streamer\u0026rsquo;s live broadcast and slicing it into dozens of short videos, each linking to the same product, with the clipper taking commission. \u0026ldquo;Matrix\u0026rdquo; means running multiple accounts at once to dominate search results for the same keyword.\nShenzhen has two homegrown AI avatar leaders worth knowing. One is Shanjian Intelligence, headquartered in Bao\u0026rsquo;an, number one nationally in 2024 in AI avatar applications with over 300 million users. The other is Shiyun Tech (the China entity behind HeyGen), headquartered in Nanshan, with ARR breaking $20M in 2024 and valuation up 6x in 4 months.\nHow you make money Three parallel paths, ranked by barrier and payback speed:\nOutsourced gigs: list \u0026ldquo;AI avatar promo videos, 1-hour delivery\u0026rdquo; on Taobao or Pinduoduo. Charge local restaurants, cosmetics shops, and medical aesthetics clinics ¥50–200 per video. Live-stream clipping: get authorization from a top streamer, slice their live into dozens of videos linked to product, take a cut of sales (30/70, 40/60, 50/50 tiers). Self-run matrix: run one AI avatar across 5–10 accounts, push products you picked yourself (senior health, education shorts, local boss IPs), and rely on platform distribution to move inventory. Where does the money come from? Shenzhen has two especially fat demand pools. One is cross-border e-commerce: 120,000 cross-border sellers, 18 listed companies (most in China) — all natural buyers for \u0026ldquo;AI foreign-model on camera\u0026rdquo; services. Shanjian\u0026rsquo;s public case studies include a cross-border seller with +600% performance. The other is the local boss crowd willing to spend on personal branding: 809 medical aesthetics clinics in Shenzhen (second nationally), 284 in Nanshan + 256 in Futian + 97 in Luohu adding up to 637 in the three CBDs, plus lawyers, insurance brokers, and cross-border bosses. This group wants a personal brand but won\u0026rsquo;t go on camera. Exactly the right client for AI avatar IP ghostwriting.\nWhere the profit lives Price tiers:\nPure voiceover short video gigs: ¥50–150 per video, cost under ¥15 Single AI avatar explainer video (Shenzhen local rates): ¥5,000–9,000 per package Mainstream AI avatar subscription tools: Chanjing ¥298/month, Feiying from ¥120, Shanjian annual pass ¥398 A top domestic AI avatar company\u0026rsquo;s average ticket: ¥350,000, with ¥655M in 2024 revenue, heading for Hong Kong IPO A Hangzhou AI live-stream tool drove ~¥100M in sales for a domestic sportswear brand at 90x+ ROAS Clipping revenue tiers (top livestream agency): 30/70 for clips under ¥50K, 40/60 for ¥50K–100K, 50/50 above ¥100K Single-video gross margin runs 70–90%. One computer, one AI avatar account, 30 voiceover videos a day — that\u0026rsquo;s the simplest assembly-line version.\nOnce a matrix is humming, ¥50K–200K/month for one person is realistic.\nWhat the barrier is You need to know Jianying (CapCut) (open it, cut it, add subtitles), some topic instinct (what hook makes someone finish a Douyin video), and tolerance for platform risk (this business is on regulators\u0026rsquo; tightest leash — accounts get banned overnight).\nYou don\u0026rsquo;t need to go on camera (clone your image once, use it forever), have stage presence (AI writes the script), have advanced editing chops (auto-clipping tools), or pick products (clipping uses what the top streamer already vetted).\nWho fits: people who\u0026rsquo;ve worked as editors at livestream agencies, run Douyin accounts, or have a real instinct for what trends on short video. Shenzhen\u0026rsquo;s bonus: Tencent HQ is here. The Video Account product, ops, and ecosystem teams are all in Shenzhen, so local service providers catch Video Account opportunities one beat faster than anyone outside. Shenzhen Houtuo (in Chegongmiao), Tencent\u0026rsquo;s core local ad agency, has served 50,000 South China companies over 11 years.\nWho shouldn\u0026rsquo;t touch this: people who don\u0026rsquo;t watch short video and don\u0026rsquo;t know the platform rules. Over 60% of clipping-factory editors unbind their authorization within a month. Most can\u0026rsquo;t hack it. One more warning: Tencent Cloud\u0026rsquo;s developer community documented a real case where someone in Shenzhen did medical aesthetics, paid ¥1M for an AI avatar agency package, got banned the moment they went live, and got fully refunded. Sketchy agency packages — dodge them.\nHow an ordinary person starts By resources on hand: outsourced gigs are lowest barrier, clipping is medium, self-run matrix is heaviest.\nPath A: outsourced gigs. List \u0026ldquo;AI avatar promo, 1-hour delivery\u0026rdquo; on Taobao or Pinduoduo, or get onto Zhubajie or Yungongwang at the ¥7-clone entry tier. Charge Shenzhen-local restaurants, cosmetics shops, and medical aesthetics clinics ¥50–200 per video. Keep per-video cost under ¥50 (Chanjing membership at ¥298/month works out to about ¥3.3 per minute). Path B: clipping. Clear the Zhongxiao Er app gate first. Download, hit \u0026ldquo;distribution\u0026rdquo; at the bottom, then \u0026ldquo;training and exam,\u0026rdquo; watch all 4 lessons (about half an hour), pass the online exam, add a streamer\u0026rsquo;s account, get the agency\u0026rsquo;s authorization, sign the contract. Entry requirements: Douyin account with 1,000+ followers, product showcase enabled, ¥500 deposit, and 10+ no-violation posts on your profile. Within 24 hours of getting authorization, you must delete unrelated posts and links, or your authorization auto-terminates. A single streamer can authorize up to 5,000 accounts. Compensation: clipping-factory team leads earn ¥12K, rank-and-file ¥6K–7K. Path C: self-run matrix. Clone your first AI avatar. Shoot a 5-second to 1-minute front-facing talking video — 1080p, solid background, even lighting, no mask or sunglasses. Record 30 seconds to 1 minute of clean audio. The cheapest combo is Chanjing + DeepSeek: DeepSeek writes scripts, Chanjing generates video from the uploaded avatar, Jianying adds a 1.5-second fixed intro. Steal the 30-videos-a-day workflow. 7–9 AM: Doubao scans trending topics, DeepSeek writes 100 ideas, you cut to 20. 9–11 AM: drop into opening-hook templates and write scripts. 11 AM–1 PM: Chanjing or Feiying generates videos. 1–4 PM: second pass in Jianying. 4–6 PM: matrix distribution, 30+ minute intervals between accounts on the same IP. 9–11 PM: catch the senior-bedtime sales window on Video Account and post. Before the matrix hits 10 accounts, do not skip the account-isolation trifecta. Anti-detect browser, dedicated residential IP per account, fully isolated cookies. Hard caps: max 2 new accounts per phone IP per day, max 5 per network IP per day. New Xiaohongshu accounts have a clear go-live trigger after 30 days of warmup. The moment your short-video avatar click rate hits 1.5%, go live immediately — conversion rate gap can be as much as 40x. 5. AI Resume Rewriting and Job-Search Coaching The first four sell to bosses. This one sells to workers.\nPeople rewriting their resume at 2 AM are the fastest payers there are. The cash flow in this business doesn\u0026rsquo;t come from AI. It comes from insomnia. Shenzhen is one of the country\u0026rsquo;s biggest suppliers of that insomnia.\nNanshan Hi-Tech Park holds Tencent\u0026rsquo;s 12,000-person Binhai Building. As of Q2 2025, Tencent had 111,200 employees, and 20,000 were already slated to move from Nanshan to Bao\u0026rsquo;an. Houhai stacks up ByteDance\u0026rsquo;s Houhai Center (77,400 sqm), Alibaba Center (123,300 sqm), and Xiaomi\u0026rsquo;s international HQ (46,000 sqm) in a single row. Huawei\u0026rsquo;s Bantian campus in Longgang holds about 40,000. Futian CBD has the China Merchants Bank tower and the 600-meter, 118-floor Ping An Finance Centre. Happening at the same time: ByteDance Feishu cut 20% of staff; Sangfor laid off 749 people at once and posted a ¥592M net loss in 2024; Huawei\u0026rsquo;s outsourcing-to-full-time conversion rate swings from 20% to 80% across departments. Shenzhen programmer unemployment hit 7.4% in 2024 versus 5.1% nationally.\nJob-seeker applications from the 35–45 bracket jumped 180% year-over-year in 2025. \u0026ldquo;Mid-life anxiety\u0026rdquo; psychological consultations rose 150%. Nanshan secondhand homes average ¥84,000/sqm, with a median transaction price of ¥8.06M and a typical monthly mortgage of ¥15,000–20,000. Two to eight months unemployed means ¥60K–160K in net mortgage outflow.\nAmong all job-search coaching clients in China, Shenzhen contributes the most desperate buyers — the ones willing to pay top dollar, able to pay top dollar, and wanting it done as fast as possible.\nHow you make money You handle the full pipeline for job-seekers: resume rewrites, mock interviews, application assistance.\nWhat you actually sell:\nResume rewriting (entry point): client sends their old resume plus the target JD. You use AI to rewrite it into a \u0026ldquo;quantified results + keyword backfill + ATS-friendly\u0026rdquo; version. Delivery: PDF and Word. Mock interviews: 1–2 Tencent Meeting recordings plus feedback forms, drilled on the client\u0026rsquo;s target role. Job-search coaching package: 4–12 weeks of full-pipeline guidance — resume, interview, role matching, internal referral, salary negotiation. Prices run from ¥99 per resume to ¥50K–60K guaranteed-offer packages. A healthy pace is 2–8 jobs a day. Monthly revenue past ¥50K is achievable.\nWhere the profit lives Price ladder:\nDeepSeek + prompt-engineered resume rewrite: ¥99 per piece, 2–3 jobs a day, low five figures monthly Xiaohongshu influencer packages: ¥99–199 / ¥299–499 / ¥699 (three tiers) Shenzhen local rates: resume rewrite ¥300–888, 1-on-1 consult ¥500/hour, coaching package ¥2,500–3,999, guaranteed-offer package around ¥20,000, internal referral ¥7,000–8,000 Returnee-student guaranteed-offer sprint package: ¥57,800 Returnee-student non-guaranteed package: ¥37,800 AI mock interview, single session: ¥30 Human mock interview, single session: ¥200–400 (6–13x the AI version) ¥99 volume gigs and ¥50K–60K coaching packages exist side by side. Low-tier margin is 70%; high-tier is 90%+.\nWhat the barrier is You need to know an industry\u0026rsquo;s hiring logic (internet engineering, investment banking and consulting, returnee students — pick one and go deep), the ability to wrestle JDs from AI with more clarity than the recruiter wrote them, the patience to listen to people cry (clients DM you at midnight venting about being laid off — you have to take it), and patience (2–3 rounds of revision per resume is normal).\nYou don\u0026rsquo;t need a recruiter background (though Big Tech, investment-bank, or returnee experience can 5x your ticket size), a counseling license (you\u0026rsquo;re rewriting resumes, not doing therapy), an office (Xianyu, Xiaohongshu, and Feishu meeting rooms are enough), or upfront investment (an AI tool and a screen recorder at a few tens of yuan a month gets you started).\nWho fits: ex-recruiters, ex-campus recruiters, people who got laid off from Big Tech in the past 2 years but jumped to a new role, returnee-student recruitment officers with 3+ years of experience, ex-headhunters. Shenzhen has one natural advantage: if you just came out of Tencent, Huawei, ByteDance, China Merchants Bank, or Ping An, you can explain in detail how hiring logic differs across Nanshan Hi-Tech Park, Bantian, and Futian CBD. Clients pay you for that \u0026ldquo;I was inside\u0026rdquo; credibility.\nWho shouldn\u0026rsquo;t touch this: fresh graduates without a full \u0026ldquo;apply → interview → get offer\u0026rdquo; cycle behind them. Clients are most afraid of paying someone whose situation looks like theirs.\nHow an ordinary person starts A ¥50 \u0026ldquo;resume ghostwrite\u0026rdquo; listing on Xianyu cumulatively sold 50,000 units. Lao Hui, a Juejin blogger, originally worked in construction; he crossed over with GPT — half an hour per resume, ¥100 deposited. A writer at Shuying spent ¥180 to test the service and got back a 12-page, 3,164-word revision plan.\nThe underlying pattern is clear: don\u0026rsquo;t fake being a Big Tech recruiter. Just deliver honestly and let AI speed you up.\nPick one persona of three: Big Tech interviewer, \u0026ldquo;been there done that\u0026rdquo; peer, company owner. Xianyu title formula has four parts: brand word + product word + modifier + trending word. Start at ¥50–99. Below ¥50 you get crushed by template-copying farms; above ¥99 with no reviews, you can\u0026rsquo;t hold the price. Run your first job on Lao Hui\u0026rsquo;s four-step playbook. First, sort out \u0026ldquo;rewrite from scratch\u0026rdquo; vs \u0026ldquo;edit existing.\u0026rdquo; Have the client send the target JD too. Feed work history into GPT in segments. Apply template and confirm. Five fixed deliverables: assessment PDF, revised docx + PDF, marginal annotations on every change, one free revision within 24 hours, naming convention {Name}_{Role}_v{N}.pdf. Six-dimension prompt framework. Structure, content sharpening (\u0026ldquo;responsible for sales\u0026rdquo; → \u0026ldquo;Q3 YoY +37%, ¥6.8M → ¥9.3M\u0026rdquo;), STAR (situation, task, action, result), de-colloquializing, stripping images, tables, and multi-column layouts for ATS compatibility, JD keyword backfill (8–12 terms). For internet roles, hammer \u0026ldquo;project + tech stack + business impact.\u0026rdquo; For state-owned finance (CMB, Ping An, Shenzhen Securities Communications and other top Shenzhen financial institutions), swap to \u0026ldquo;compliance + licenses + cross-departmental.\u0026rdquo; For 35+ candidates, switch date granularity to years and push the education section to the bottom. Post on Xiaohongshu for 15–30 days straight. Build titles around \u0026ldquo;number + contrast + time limit\u0026rdquo;: \u0026ldquo;Sent 80 applications with no response. Changed 3 lines and got 5 interviews in a week.\u0026rdquo; \u0026ldquo;Broke down 30 Big Tech JDs. Only 12 high-frequency keywords.\u0026rdquo; Train the search system to flag your account as resume-related. Shenzhen-local search-term traffic is already there: \u0026ldquo;Nanshan layoffs,\u0026rdquo; \u0026ldquo;Huawei outsourcing conversion,\u0026rdquo; \u0026ldquo;ByteDance Feishu layoffs,\u0026rdquo; \u0026ldquo;35 and over Tencent optimization.\u0026rdquo; Five pricing tiers: ¥99: rewrite 1 resume + 24-hour delivery, bank 5–10 reviews ¥199: add ATS keyword match score, cover letter, 2–3 role variants ¥599: add Tencent Meeting recorded walkthrough, mock interview feedback form ¥2,999: 4-week coaching with weekly Feishu reviews ¥5,999: add internal referral and salary negotiation Do not touch auto-application bots. BOSS Zhipin banned almost 20,000 fraudulent accounts in Q3 2025. When accounts get banned, the damage claims all land on the service provider.\nThe compliant version is the \u0026ldquo;application assistance pack\u0026rdquo;: auto-fills forms, hands the client a list of companies, and the client clicks submit themselves.\nClosing The people actually making money in this AI wave are the ones who can ground AI in what merchants and workers actually need.\nMaking money on AI in Shenzhen requires two things:\nThe ability to hear what the boss is actually hurt by The ability to wire the work into a delivery pipeline that runs ","permalink":"https://ryrenz.com/industry/shenzhen-ai-five-businesses/","summary":"Local merchant operations, cross-border AI product-image factories, AI export lead gen, AI avatar short videos, and AI resume rewrites plus job-search coaching. No tech background, no capital, no team needed — just use AI to amplify a single person\u0026rsquo;s output and sell to the anxious small-business owners and white-collar workers packed into Shenzhen. Each business broken down by how you make money, where the profit sits, what the barrier is, and how to start.","title":"Five AI Businesses That Actually Print Money in Shenzhen — A Full Teardown"},{"content":"Ask GPT to write you hard sci-fi in the vein of The Three-Body Problem, or a Harry Potter-style fantasy, or a pulpy web novel. By around the 10,000-word mark it will start talking gibberish — characters break, plot lines tangle, the whole thing collapses.\nMost people quit here and conclude: AI can\u0026rsquo;t write a novel.\nForget the argument for a second. Look at four facts.\nJanuary 2024: Japanese author Rie Kudan won the Akutagawa Prize, Japan\u0026rsquo;s top literary award, for Sympathy Tower Tokyo. She openly admitted about 5% of the book came directly from ChatGPT. The judges called it \u0026ldquo;almost flawless.\u0026rdquo; American indie sci-fi author Joe Vasicek used to spend 6 months to 2 years on a book. His most recent novel — 18 chapters, 80 scenes, 153,254 words — took 5 days of AI generation. American genre author Leanne Leeds has 5 series, 40+ books, a 4.5-star Amazon average. Beta readers can\u0026rsquo;t tell she uses AI. She finishes a book in about 3 weeks. March 2026: Hachette, one of the Big Five publishers, urgently pulled a horror novel called Shy Girl. It launched in November 2025, blew up on BookTok, and racked up over 4,900 ratings on Goodreads. The reason for the pull: an AI detection company called Pangram measured the book at 78.4% AI-generated. The tools they use are mostly ones you\u0026rsquo;ve never heard of. The workflows are barely covered by Chinese-language tech press. The results are already on Amazon, on the Akutagawa stage, and on Hachette\u0026rsquo;s preorder lists.\nChatGPT and Claude Aren\u0026rsquo;t the Right Tools Most people\u0026rsquo;s mental model of \u0026ldquo;writing a novel with AI\u0026rdquo; stops at one action: open ChatGPT, paste a prompt, let it run.\nThat approach can\u0026rsquo;t produce a novel. I\u0026rsquo;ve covered the reason before: the context window of a large model is a FIFO queue. By round 20, the protagonist\u0026rsquo;s personality drifts. By round 50, the worldbuilding collapses. Gemini\u0026rsquo;s 1M-token window sounds huge, but attention tests show that past 50-100k tokens the middle of the context gets ignored — academics call it \u0026ldquo;lost in the middle.\u0026rdquo; The window fits, the model just can\u0026rsquo;t use it.\nAuthors who actually write long-form have a three-tier answer:\nTier 1: Purpose-built front ends for novelists, like Sudowrite and NovelCrafter, which manage context injection for you. Tier 2: Beyond the specialized front ends, use Scrivener or Atticus as the primary writing software, and bring AI in only at specific points — scene expansion, descriptive polish, outline brainstorming. Tier 3: Wire the Claude API into your own workflow (Notion, Obsidian, a custom tool) and treat AI as a callable subtask, not the main writer. The people using these are almost all Western authors. Let\u0026rsquo;s walk through them.\nSudowrite Plus Muse Is the Western Mainstream If you\u0026rsquo;ve only ever used Claude\u0026rsquo;s web UI to write fiction, you wouldn\u0026rsquo;t know: there\u0026rsquo;s a tool built specifically for novelists that calls Claude, GPT, and Gemini all as backends — and has trained its own dedicated novel-writing model.\nIt\u0026rsquo;s called Sudowrite.\nIn one sentence: it breaks \u0026ldquo;writing a novel\u0026rdquo; into dozens of UI components, each backed by a different model call. You\u0026rsquo;re no longer staring at a blank chat box. You\u0026rsquo;re inside a Photoshop-style interface — character and worldbuilding panels on the left, the chapter you\u0026rsquo;re writing in the middle, AI-generated candidate paragraphs on the right.\nSudowrite the company is worth a paragraph on its own:\nFounded 2020 in San Francisco, bootstrapped, with only one $3M round raised by 2025 16-person team, $1.8M ARR in 2025 Investor list reads like a Silicon Valley plus Hollywood crossover: Ev Williams (Twitter, Medium), Sahil Lavingia (Gumroad), Matt Mullenweg (WordPress), John August (Aladdin screenwriter), George Nolfi (director of The Bourne Ultimatum) In March 2025 Sudowrite launched its in-house model, Muse 1.0. Three months later came Muse 1.5. By 2026, Muse 1.5 is the default. It\u0026rsquo;s fine-tuned specifically on published novels — bad at coding, bad at trivia, good at exactly one thing: writing fiction.\nThe company\u0026rsquo;s changelog claims: in internal blind tests, authors preferred Muse 1.5\u0026rsquo;s prose to Claude 3.7 Sonnet\u0026rsquo;s at a 2:1 rate.\nDiscount that number. The blind test was run by Sudowrite, with no third-party verification and no public methodology. But heavy users on blogs and podcasts keep confirming that Muse really is stronger on sensory detail and prose feel.\nAn example. Describing \u0026ldquo;a strange book,\u0026rdquo; a generic LLM gives you something flat like \u0026ldquo;the smell of leather.\u0026rdquo; Muse 1.5 gives you:\n\u0026ldquo;The grind of worn leather against old wood. A thread of metallic blood. The cover carries the loamy smell of soil after a lightning strike.\u0026rdquo;\nThat kind of sensory granularity is what Sudowrite paid millions in fine-tuning to get.\nSudowrite\u0026rsquo;s core features:\nStory Bible: you write down your worldbuilding, characters, and outline; Sudowrite stores it as structured data and injects the relevant pieces every time you write a paragraph. It\u0026rsquo;s an always-on novel-setting encyclopedia. Describe: highlight a noun in your text, say \u0026ldquo;that strange book,\u0026rdquo; and get 3 candidate descriptions from each of the five senses — sight, smell, touch, hearing, taste. Expand: highlight a skeleton sentence like \u0026ldquo;She reached out and opened the book,\u0026rdquo; and it expands into a full paragraph. Brainstorm: drop in a scene or a problem, get 5-10 creative options — no full prose, just sparks. Draft: feed in a chapter outline (200-500 words describing each scene) and generate the full chapter. What Sudowrite Power Users Actually Do Feature lists are abstract. Look at three authors who\u0026rsquo;ve actually shipped books with Sudowrite and how they use it.\nCase 1: Joe Vasicek cut his book cycle from 2 years to 14 days Joe Vasicek is an American indie sci-fi author with 20+ books on KDP, best known for the Captive of the Falconstar series.\nHe publishes every number from his latest project on his blog:\nTitle: The Soulbond and the Sling Chapters: 18 chapters plus prologue and epilogue, 80 scenes Total word count: 153,254 words Pre-writing phase (outline + worldbuilding): 9 days AI generation phase: 5 days Manual revision afterward: no specific day count given, but \u0026ldquo;the largest chunk of total time\u0026rdquo; Sudowrite credits burned: about 770,000 His workflow runs on Claude. He says outright on his blog that Claude\u0026rsquo;s chapter-level output is \u0026ldquo;stunning\u0026rdquo; and makes him \u0026ldquo;feel more like a first-time reader of the novel than its author.\u0026rdquo; He uses Muse 1.5 as a parallel generator — same input fed to both Claude and Muse, then he hand-picks the better paragraphs and stitches them together.\nHis warning to anyone trying AI writing:\n\u0026ldquo;The worst thing you can do is let the AI do all the work. AI doesn\u0026rsquo;t really create. It analyzes patterns in human language and reproduces them.\u0026rdquo;\nAfter every AI-generated chapter, he opens a new document and rewrites it by hand. Never copy-paste. The AI draft is a reference he can consult at any time, nothing more.\nCase 2: Leanne Leeds writes 12-15 books a year at a 4.5-star Amazon average Leanne Leeds is an American genre author with 5 series spanning contemporary supernatural, fantasy, and cozy mystery. She has 40+ KDP titles, a 4.5-star Amazon average, and beta readers who don\u0026rsquo;t notice the AI.\nShe laid out her full workflow on The Creative Penn podcast:\nSkeleton draft: she calls it \u0026ldquo;boring writing.\u0026rdquo; Just action and dialogue. \u0026ldquo;He looked at her, she looked at him, he blinked, she kissed him.\u0026rdquo; No description. Dual-monitor setup: Scrivener on the center screen, Sudowrite plus ChatGPT plus Quillbot on the left. Scene expansion: she selects a skeleton paragraph and uses Sudowrite\u0026rsquo;s Rewrite to expand it. She gives specific instructions like \u0026ldquo;show don\u0026rsquo;t tell\u0026rdquo; or \u0026ldquo;describe the look on his face.\u0026rdquo; Mystery elements: she hands suspect lists and red herrings to ChatGPT. Sentence polish: Quillbot does the final cleanup pass. The key numbers:\nAI text as a share of the final draft: about 10% (she measures it with a text-comparison tool, holds steady between 8 and 11%) Time to finish a book: about 3 weeks Every book includes an AI-assisted creation disclosure on the copyright page Her most-quoted line:\n\u0026ldquo;I use AI to fix the weaknesses in my own writing so that readers get a better experience.\u0026rdquo;\nBy 2026 she\u0026rsquo;s migrated off pure Sudowrite onto a stack of Claude plus Gemini Pro plus NotebookLM. In her latest podcast episode:\n\u0026ldquo;I don\u0026rsquo;t use AI for plot suggestions — I know where each chapter and scene is going. I use AI to express what I\u0026rsquo;m trying to say better than I would alone.\u0026rdquo;\nCase 3: Joanna Penn, the AI-Assisted Artisan on the NYT list Joanna Penn (pen name J.F. Penn) is a British mystery and thriller author whose ARKANE series has hit both the NYT and USA Today bestseller lists. Her 2025 release is Death Valley.\nOn her own podcast, The Creative Penn, she\u0026rsquo;s mapped out her positioning: AI-Assisted Artisan Author. The concrete stack:\nChatGPT for research and outlining Sudowrite for scene description and expansion Midjourney for covers ElevenLabs to clone her own voice for audiobooks Runway ML for book trailers Every prompt runs over 100 words — \u0026ldquo;treat it like an employee, give it everything in your head\u0026rdquo; In early 2026 she announced a strategic pivot from \u0026ldquo;high-volume low-price digital books\u0026rdquo; toward \u0026ldquo;premium physical books and high-margin products,\u0026rdquo; with AI handling research and repetitive work while she focuses on creative decisions.\nThe Second Core Tool: NovelCrafter Plus Codex If Sudowrite is \u0026ldquo;an AI writing tool\u0026rdquo; with AI as the protagonist, NovelCrafter is \u0026ldquo;writing software plus an AI connector\u0026rdquo; — you\u0026rsquo;re the protagonist, AI is what you call.\nCompany background:\nFounded 2023, by Leonie Grabel, based in Hamburg, Germany German GmbH, fully bootstrapped, no VC 10-person team (including 2 office dogs) 157,000 registered authors Pricing $4-20/month The core feature is Codex. In plain English: you build a database of your novel\u0026rsquo;s worldbuilding, where each entry is a character, a location, a faction, or a prop. Each entry has a name, aliases, a description, custom fields, and relationship links to other entries.\nWhile you write, the system scans the paragraph you\u0026rsquo;re working on. If the name or an alias of any Codex entry shows up, all of that entry\u0026rsquo;s information gets dropped into the model\u0026rsquo;s context automatically. It\u0026rsquo;s an always-on, keyword-triggered, auto-responding worldbuilding reference manual.\nThere\u0026rsquo;s also a cascading relationship feature. If character A is linked to location B, B is linked to faction C, and C is linked to characters D and E, then writing A pulls B, C, D, and E into context automatically.\nThis mechanism solves what Sudowrite\u0026rsquo;s Story Bible can\u0026rsquo;t: cross-book series. A fantasy trilogy shares one Codex across all three books, so protagonist personality, worldbuilding, and faction relationships don\u0026rsquo;t degrade between volumes.\nNovelCrafter is compatible with 300+ models. How? It directly integrates 8 providers (OpenAI, Anthropic, Google, Mistral, Groq, Anyscale, local LM Studio, local Ollama), and the 300+ figure comes via the OpenRouter aggregator. Write one chapter with Claude Opus 4.7, another with locally hosted DeepSeek.\n2026 Long-Form Writing Model Benchmarks inkfluenceai.com recently ran a systematic comparison: 70,000-word novel generation tests scored out of 70.\nKey findings:\nClaude Opus 4.7 is the most stable on cross-chapter character consistency — it \u0026ldquo;correctly preserves the voice of a minor character who only appears once in chapter 3\u0026rdquo; GPT-5 is the strongest on plot logic, but loses character detail Gemini 2.5 Pro\u0026rsquo;s 1M context is theoretically the largest, but it scores lowest on writing style Then the cost comparison gets brutal:\nA 70k-word novel via API: Claude Opus 4.7 costs about $32 Same book via Claude Sonnet 4.5: about $5.50 Quality gap: only 4%. Cost gap: 88% Community consensus: use Sonnet for routine chapter generation, save Opus for character description review and the key chapters that need polish.\nThe Scoreboard for AI Writing Lining up everyone who\u0026rsquo;s actually written something serious with AI:\nRie Kudan: January 2024, Sympathy Tower Tokyo wins Japan\u0026rsquo;s Akutagawa Prize. About 5% of the text is from ChatGPT. The jury calls it \u0026ldquo;almost flawless.\u0026rdquo; Japan\u0026rsquo;s 13th Hoshi Shinichi Award: roughly a quarter of submissions carried AI signal. The grand prize, Tower of the Genome, was a human-AI collaboration. The award institutionalized rules: no verbatim copying of AI output, mandatory records of the creation process. It\u0026rsquo;s one of the few global literary prizes that explicitly welcomes AI-assisted submissions. Leanne Leeds: 40+ KDP titles, 4.5-star Amazon average. Beta readers don\u0026rsquo;t notice the AI. Joe Vasicek: 20+ sci-fi titles on KDP. Book cycle compressed from 6-24 months to 1-2 months. Joanna Penn: ARKANE series on the NYT and USA Today bestseller lists. Elisa Shupe: American military veteran, AI Machinations. In April 2024 the U.S. Copyright Office granted limited copyright to an AI-collaborative work for the first time — a landmark case in copyright law. Readers Can\u0026rsquo;t Tell Anymore This is one story, and it needs telling from the start.\nFebruary 2025: an American author named Mia Ballard self-published a horror novel on Amazon KDP called Shy Girl. The cover was a girl\u0026rsquo;s silhouette; the themes were trauma, isolation, childhood shadow.\nThe book caught fire on BookTok (TikTok\u0026rsquo;s reading community). Book reviewers filmed unboxing videos of the physical edition, gave it 5 stars, recommended it enthusiastically. On Goodreads it racked up over 4,900 ratings, averaging 3.52 stars.\nThe buzz reached Hachette, one of the Big Five publishers — meaning it sits next to Penguin Random House, HarperCollins, Simon \u0026amp; Schuster, and Macmillan at the very top of global publishing. Hachette\u0026rsquo;s UK imprint Wildfire is the one that published Fifty Shades of Grey.\nIn March 2025, two Hachette imprints — Orbit Books in the US and Wildfire in the UK — jointly acquired the rights to Shy Girl. A self-published horror novel had walked all the way from KDP to a Big Five preorder catalog.\nIn November 2025 the UK edition shipped about 1,800 copies.\nUp to this point it\u0026rsquo;s a textbook BookTok success story: viral indie hit gets picked up by a major publisher.\nThe turn came in January 2026.\nA YouTube book reviewer named Frankie (channel: Frankie\u0026rsquo;s Shelf) dropped a 2-hour 40-minute analysis video titled \u0026ldquo;i\u0026rsquo;m pretty sure this book is ai slop.\u0026rdquo;\nHer specific evidence:\nRepeated use of the \u0026ldquo;three in a row\u0026rdquo; structure when describing the position of door locks An unusually high frequency of the word \u0026ldquo;sharp\u0026rdquo; Layout choices that were \u0026ldquo;baffling,\u0026rdquo; including jarring paragraph breaks Inconsistent narrative voice across passages — like different people had written them The video went viral on BookTok and X, hitting 1.5M+ views.\nWhat came next was sharper. Thad McIlroy, a publishing industry analyst, submitted the Shy Girl ebook to Pangram, an AI detection company. Pangram is one of the more academically cited AI detection tools today.\nPangram\u0026rsquo;s verdict: 78.4% of the text was AI-generated.\nMcIlroy passed the report to New York Times reporter Alexandra Alter. On March 19, 2026, the NYT ran the investigation.\nWithin a day of being notified, Hachette canceled the U.S. publication and halted UK sales. The statement was carefully worded: after \u0026ldquo;an extensive investigation,\u0026rdquo; they had decided to cancel publication. No direct accusation of the author.\nMia Ballard\u0026rsquo;s response was unexpected. She denied using AI to write the book — claiming that \u0026ldquo;a freelance editor she\u0026rsquo;d hired had used AI tools without her knowledge.\u0026rdquo; Citing ongoing litigation, she declined to comment further.\nThis is publishing\u0026rsquo;s defining event of 2026, because it exposed a loophole no one had thought about: existing publishing contracts only require AI disclosure at the author level, not from editors, proofreaders, literary agents, or anyone else in the production chain. A book might not be written by the author using AI; it might be edited by the editor using AI.\nAnd the 4,900 BookTok readers who gave it 5 stars never noticed. Hachette\u0026rsquo;s acquisition editors never noticed. Not until one YouTube reviewer spent 2 hours and 40 minutes dissecting the text, and one AI detection company put a 78.4% number on it, did the story break.\nWhere the Big Five Stand on AI None of the Big Five publishers has issued a company-wide policy banning AI use by authors. That policy vacuum is one root cause of the Shy Girl incident.\nAmazon KDP\u0026rsquo;s policy is delicate:\nIt distinguishes \u0026ldquo;AI-generated\u0026rdquo; (must be disclosed) from \u0026ldquo;AI-assisted\u0026rdquo; (no disclosure needed) Grammar checks and rewriting suggestions count as assistance Disclosures are not shown to readers on the product page, only kept in Amazon\u0026rsquo;s internal records Enforcement ramped up in 2025-2026, but no names of banned AI-author accounts have been made public — enforcement is mostly quiet delistings U.S. Copyright Office, March 2026: the Supreme Court declined to hear the case on whether AI alone can create a copyright-protected work, upholding the lower court ruling — AI cannot be a copyright holder.\nThe Authors Guild launched a Human Authored certification program, $10 per book for non-members. Their 2025 survey:\n90% of authors believe they should be compensated if their work is used for AI training 96% demand it requires consent Where the Chinese-Language Writing World Stands Start with the counterexamples. The Chinese-speaking world isn\u0026rsquo;t a total void on AI writing, but everything sits at one of two extremes: academic stunts or web-novel assembly lines.\nAcademic experiments:\nShen Yang\u0026rsquo;s team at Tsinghua University\u0026rsquo;s School of Journalism used AI to generate a ~40,000-word sci-fi novella, Land of Machine Memory, in 3 hours (66 prompts, Kafka-style). It won second prize at the 5th Jiangsu Sci-Fi \u0026amp; Popular Science Award, submitted anonymously. After the team came forward, Liu Cixin said it was \u0026ldquo;honestly better than what I\u0026rsquo;d write\u0026rdquo; and felt \u0026ldquo;a huge sense of loss.\u0026rdquo; Wang Feng\u0026rsquo;s team at East China Normal University used AI to write a ~1.1M-word novel, Apostle of Fate, over 6 weeks. Neither was a commercial release. Both were academic demos.\nWeb-novel assembly lines:\nTomato Novel (Fanqie) hit 5,606 new-book debuts in a single day in March 2024, up 13x from 400 in 2023 — pure AI-driven mass production April 2025: Tomato tightened its signing process March 2025: Qidian banned all AI-generated content Jinjiang allows proofreading and polishing, but bans AI-driven plotting No Chinese platform allows AI-labeled works to be displayed or distributed on their own Chinese AI writing tools are almost entirely web-novel focused: Waqu Pinwen and Wawa Writing (Hangzhou Yinli Zhihang Tech) pitch long-form web-novel memory, \u0026ldquo;removing the AI smell,\u0026rdquo; and outline generation. Yuewen\u0026rsquo;s Author Assistant integrates DeepSeek-R1. There is no Chinese equivalent of Sudowrite — no professional tool aimed at serious long-form fiction.\nWhere Chinese literary heavyweights stand:\nMo Yan: AI prose is impressive but \u0026ldquo;lacks real thought or creativity\u0026rdquo; Liu Cixin: read Land of Machine Memory, said \u0026ldquo;honestly better than what I\u0026rsquo;d write,\u0026rdquo; felt a sense of loss — doesn\u0026rsquo;t use AI himself Han Song: AI can be a collaborator that sparks ideas, but the unique human experience and subconscious insight aren\u0026rsquo;t things AI has Mao Dun Literature Prize and Lu Xun Literary Prize: rules don\u0026rsquo;t explicitly mention AI but emphasize originality Science Fiction World magazine: explicitly stopped accepting AI-generated fiction in 2023 The Chinese writing world is a blank space on \u0026ldquo;an indie author using AI assistance to write a quality long-form novel.\u0026rdquo; Not for one single reason:\nMisaligned commercial incentives: industrialized web-novel mass production is far more profitable than serious AI-assisted literary work Model quality: Chinese-language LLMs lag Claude and GPT-5 on long-form coherence and emotional depth Cultural taboo: publicly admitting AI use in China can mean delisting, copyright disputes, social-reputation pressure Tool gap: every domestic tool targets web-novel volume; nothing supports serious long-form workflow Distribution gap: China has no Amazon KDP — no direct self-publish channel to readers Stack those five and Chinese-language AI novel writing only exists at two poles: the bottom rung of the web-novel assembly line and the academic lab. The middle — indie authors using AI to produce quality long-form work — is empty.\nWhat the Truth About AI Writing Looks Like Put all the data together and here\u0026rsquo;s where AI writing actually stands:\n1. AI doesn\u0026rsquo;t turn bad writers into good ones. Joe Vasicek, Leanne Leeds, Joanna Penn, Rie Kudan — the people producing real long-form work were already writers. AI made them 5-10x faster. It didn\u0026rsquo;t raise their ceiling.\n2. AI massively lifts the productivity of good writers. A 6-month-to-2-year book cycle compressed to 1-2 months. A KDP genre author shipping 12-15 books a year, 4.5 stars average. This is happening, not predicted.\n3. AI writing tools keep getting stronger. Sudowrite, NovelCrafter, Claude API plus Scrivener are putting publishable work on shelves every day.\n4. Quality still isn\u0026rsquo;t there. Claude Opus 4.7 scored 63.2 out of 70 in benchmarks — the highest combination today — but no case yet of \u0026ldquo;AI as primary writer produces a lasting work of literature.\u0026rdquo; Rie Kudan\u0026rsquo;s Akutagawa novel was only 5% AI.\n5. Readers already can\u0026rsquo;t tell. Shy Girl had 4,900+ Goodreads ratings at 3.52 stars on average. Hachette\u0026rsquo;s acquisition editors didn\u0026rsquo;t catch it. Not until a YouTube reviewer spent 2 hours and 40 minutes on the text plus an AI detection company posted a 78.4% number did anyone know. AI fiction is already clearing the average reader\u0026rsquo;s taste filter.\n6. The rules of the game are still being written. Hachette pulling Shy Girl exposed the Big Five\u0026rsquo;s policy vacuum. Amazon KDP\u0026rsquo;s distinction between pure AI generation and AI assistance leaves an enormous gray zone. The U.S. Copyright Office\u0026rsquo;s stance shifts monthly.\n7. The Chinese-language scene lags. The cause is a four-way absence: culture, commerce, regulation, and tool ecosystem.\nProblems keep getting solved. AI-written novels keep going on sale. Readers keep handing out 5 stars and can\u0026rsquo;t tell whether or not it\u0026rsquo;s AI.\nI\u0026rsquo;m starting to wait for the first AI novel that lasts.\n","permalink":"https://ryrenz.com/ai/ai-novel-money-makers/","summary":"Rie Kudan won the Akutagawa Prize, Joe Vasicek cut his book cycle from 2 years to 14 days, Leanne Leeds ships 12-15 books a year at a 4.5-star Amazon average, and Hachette pulled a novel after Pangram clocked it at 78.4% AI-generated. The Sudowrite / NovelCrafter / Claude + Scrivener stack is now the Western mainstream, and the Chinese-language world has nothing equivalent.","title":"How Authors Are Making Money Writing Novels with AI"},{"content":"On May 15, 2026, X updated the open-source version of its For You recommendation algorithm on GitHub, in the xai-org/x-algorithm repository.\nThis release includes a full runnable inference pipeline, a pretrained mini Phoenix model, a standalone content moderation service called Grox, and an ad-mixing module.\nA lot of things we used to verify through rumors and guesswork can now be read straight from the source code. I cross-checked the new xai-org/x-algorithm repo against the old twitter/the-algorithm repo and X\u0026rsquo;s official help pages, then pulled out 10 facts that directly affect account growth.\nFor each point, we\u0026rsquo;ll first look at how the algorithm works, then what it means for creators, and finally what you can do about it right now.\n1 The final algorithm score is a weighted formula How the algorithm works:\nDuring ranking, the Phoenix Transformer model predicts the probability of more than a dozen actions for each candidate post: P(like), P(reply), P(repost), P(quote), P(click), P(profile click), P(follow), P(dwell time), P(image expand), P(video view), P(share to DM), P(copy link), P(not interested), P(block), P(mute), and P(report). What this means:\nThe algorithm does not judge whether your post is good in some absolute sense.\nIt predicts whether your target users will take an action after seeing it.\nThe same post may go straight to the top of Zhang San\u0026rsquo;s timeline and never even enter Li Si\u0026rsquo;s recommendation pool.\nWhat to do:\nBefore writing any post, run this through your head: who do I want to see this? What do I want them to do after reading it? Follow me? Repost it? Reply? If you don\u0026rsquo;t know, don\u0026rsquo;t post it. Stop asking, \u0026ldquo;Will this topic go viral?\u0026rdquo; Ask, \u0026ldquo;Will my target reader actually do something because of this post?\u0026rdquo; It is better to have 100 non-target users scroll past and 10 target users take strong actions. 2 Likes barely count anymore. Replies, bookmarks, reposts, and retention are the real score How the algorithm works:\nEach action has a different weight in the weighted formula. The old version (twitter/the-algorithm) published a reference set of weights: like 0.5, repost 1.0, reply 13.5, profile click 12, author engaging in the reply conversation 75, and report -369. The new version does not publish the weights, but the structure has not changed. Shallow actions, such as likes and simple clicks, carry low weight. Deeper actions, such as replies, quotes, shares to DM, profile clicks, follows, and dwell time, carry high weight. Negative actions carry a heavy negative weight. Dwell time appears in the source code as dwell and dwell_time, two separate high-weight fields. Even if a reader does not press any button, if they actually finish reading your post, the algorithm still gives it a high score. What this means:\nThat \u0026ldquo;viral\u0026rdquo; post with 100 likes and zero new followers basically looks like nobody saw it, from the algorithm\u0026rsquo;s point of view. By contrast, a post with only 10 likes that brings in five new followers, 20 bookmarks, and an average dwell time of 30 seconds is what actually grows your account weight.\nDwell time deserves special attention. It is a high-weight signal most creators ignore. It cannot be botted or faked. It only comes from the quality of the content itself. Readers either finish reading or scroll away. There is no middle ground. A like can be casual. A repost can be impulsive. But dwell time means the user voted for you with their actual time.\nWhat to do:\nStop staring at like counts. Open your X Premium Analytics dashboard and look at six numbers: replies, reposts, quotes, profile visits, new followers, and average read time. If any one of these moves, the post worked. Before writing, ask yourself: does this post have a part that makes someone want to screenshot it into a group chat or forward it to a specific friend? If not, don\u0026rsquo;t post it. It is just filler. If you want bookmarks and reposts, a few structures work especially well: checklists (\u0026ldquo;7 traps for creators\u0026rdquo;), SOPs (\u0026ldquo;30-day cold-start plan\u0026rdquo;), comparison tables (\u0026ldquo;old playbook vs. new playbook\u0026rdquo;), evaluation criteria (\u0026ldquo;what qualifies as good content\u0026rdquo;), and mistake lists (\u0026ldquo;5 things not to do\u0026rdquo;). These structures are naturally screenshot-friendly. If you want higher dwell time, use four patterns: long posts that make people scroll, images that make people stop and look, videos whose watch time counts directly toward retention, and clear paragraphing with a good visual rhythm so reading feels easy and people keep moving down the page. These can raise your score without any interaction. They are an underrated hidden scoring channel. Stop ending with \u0026ldquo;like and follow if this helped.\u0026rdquo; End with a specific question people actually want to answer, such as: \u0026ldquo;Which type of content got the most engagement for you in the last 30 days?\u0026rdquo; Once people answer, your account\u0026rsquo;s deep interaction signal goes up. 3 Fighting over titles, openings, and hooks is still the core game How the algorithm works:\nThe Phoenix Transformer takes a user\u0026rsquo;s behavior sequence plus a candidate post as input. From massive amounts of real interaction data, the model learns what kinds of content lead to high-value actions such as follow_author, dwell, share, and quote. What it learns lines up with human intuition: content with a strong opening hook, a clear title, and high information density is more likely to make users stay, click the profile, repost, and bookmark. What this means:\nTitles, openings, and information density matter just as much as they did two years ago. The new algorithm did not make these three things easier, but it did not make them harder either. The only difference is the goal. Before, you were optimizing for likes. Now you are optimizing for profile clicks, follows, reposts, and quotes.\nWhat to do:\nThe first sentence, meaning roughly the first 30 characters of the post, must give a counterintuitive conclusion or a concrete promise right away. No vague openings, no long setup, no question openings. In sentences two and three, tell readers immediately what problem the post is going to solve. If the first three sentences do not make readers understand why the post matters to them, they will scroll away. Nothing later in the post can save it. After writing, reread every paragraph and ask: does this paragraph add a new piece of information? If not, cut it. If a paragraph can lose 30% of its length without losing meaning, cut it. Use this structure for long posts: opening hook → core conclusion up front → three to five concrete pieces of evidence → an action list or a closing hook. This is not a rule. It is the template that has tested best. 4 A fuzzy account profile kills distribution. Don\u0026rsquo;t jump between niches How the algorithm works:\nOut-of-network retrieval relies on Phoenix Retrieval\u0026rsquo;s two-tower model. The User Tower encodes each user\u0026rsquo;s interests, based on what they have clicked, replied to, reposted, and spent time on, into a vector. The Candidate Tower encodes each post into another vector. The system finds the posts closest to a given user in vector space and pushes them to that user. Your account profile is not a standalone vector. It is reflected through the accumulated history of all your past posts. The more focused your content is, the more concentrated this type of post becomes in vector space, and the more likely it is to appear consistently in a certain group\u0026rsquo;s top-K. Stable profile → your posts cluster in a semantic group → the system knows who to show you to. Fuzzy profile → your posts scatter across vector space → you are not close enough to any user group → out-of-network retrieval fails. What this means:\nIf your account profile is fuzzy, the result is not \u0026ldquo;a little less traffic.\u0026rdquo; It means out-of-network distribution never opens up. Only your existing followers see you, and you never break out of that circle. Follower growth, spread, and discovery by new people all go to zero. This is the easiest point to overlook among the 10, and it carries the highest cost.\nWhat to do:\nPick one specific theme as your main axis, such as \u0026ldquo;monetizing AI tools,\u0026rdquo; \u0026ldquo;X growth experiments,\u0026rdquo; or \u0026ldquo;commercializing AI products.\u0026rdquo; Then post only about that theme for 30 days straight. Not \u0026ldquo;mostly this, with occasional other things.\u0026rdquo; Only this. Use one stable term for the theme. Don\u0026rsquo;t keep switching. If you write \u0026ldquo;AI tools\u0026rdquo; this week, \u0026ldquo;AI products\u0026rdquo; next week, and \u0026ldquo;AI apps\u0026rdquo; the week after, the system sees those as different labels. Every switch scrambles your profile. Pick one and stick with it. Who you follow, who you reply to, and who you like all train your profile. Randomly following unrelated accounts or replying in other circles pollutes it. Keep your interactions inside your niche. If you really need to switch niches, give the old one a clean exit first. Delete or archive old content, update your bio, and change your pinned post. Give the profile one full reset. Half-switching only ensures the new profile never forms. Never use your account to \u0026ldquo;test which niche gets better traffic.\u0026rdquo; Every switch dilutes your profile, and the damage is not easily reversed. If you want to test a new niche, use an alt account. 5 Main posts are about quality, not volume. One to three per day is enough How the algorithm works:\nThe logic in home-mixer/scorers/author_diversity_scorer.rs is: sort by weighted score first, then count by author ID, and apply a decay coefficient to later candidates from the same author. The coefficient drops as the rank order goes down. The second post is already discounted. The third is hit harder. The fifth is basically invisible. What this means:\nPosting 10 main posts today will not make you appear 10 times on a user\u0026rsquo;s For You screen. You may appear once, maybe twice at most. The remaining eight posts either get pushed far down or are not shown at all. Worse, every low-quality post that gets no dwell time and no interaction teaches the algorithm, \u0026ldquo;this account produces low-quality content.\u0026rdquo; The more you post like that, the more damage you do.\nWhat to do:\nPost at most one to three main posts per day, and write each one until you are satisfied with it. If you can only write one good post today, publish only that one. Having two fewer main posts will not hurt your account. Too much mediocre content can crater your reach. If you still have capacity, put it all into replies. Write high-quality replies under big accounts or mid-sized accounts in your circle. This does not trigger author-diversity discounting and effectively opens dozens of extra exposure slots. Point 6 covers this in detail. If you make a series on the same topic, change the angle each time: one tutorial, one counterexample, one comparison table, one checklist, one opinion. If every post is the same idea in different packaging, the system will treat it as redundant and merge the score. Don\u0026rsquo;t publish three or four main posts on the same topic within a short window of one or two hours. They compete with each other for candidate-pool positions, and all of them end up with weaker scores. Leave at least two to three hours between main posts. 6 High-quality replies under big accounts are a high-ROI traffic entry point How the algorithm works:\nPhoenix Transformer predicts fields such as P(reply), P(quote), P(profile_click), and P(follow_author). The original author participating in the conversation around your reply, called reply_engaged_by_author in the old version with a weight of 75, has always been one of the high-weight signals. Replies appear on the original post\u0026rsquo;s conversation page, and a separate conversation-ranking system decides which replies rank near the top. In the source code, Grox specifically handles labeling for reply spam and reply ranking. This means your reply exposure does not need to go through the Thunder and Phoenix Retrieval flow used by the main recommendation system. Your exposure is directly tied to the original post. What this means:\nWriting one high-quality reply under a big account\u0026rsquo;s post is like borrowing that account\u0026rsquo;s traffic for free. Even better, replies do not count against your own main-post quota. Five main posts plus 30 excellent replies in one day is like opening 32 separate exposure slots for yourself. This is the cheapest leverage during cold start.\nWhat to do:\nSet a reply KPI for yourself every day, such as 20 to 40 replies. Pick posts in your target circle and reply within 30 minutes after they are published. Early replies get much more exposure. If you show up an hour or two late, that post\u0026rsquo;s traffic has already passed. Your replies must contain real substance. Add a counterexample, a data point, a step, a tool, or a personal experience to the original post. \u0026ldquo;Agree,\u0026rdquo; \u0026ldquo;learned something,\u0026rdquo; \u0026ldquo;+1,\u0026rdquo; and \u0026ldquo;great insight\u0026rdquo; are the same as writing nothing. They will not bring profile clicks. Choose big accounts that actually reply to followers. Accounts with millions of followers that never reply to anyone have low reply value. One reply from the original author injects a direct high-weight signal into your account. The accounts you reply to must be in your niche. Don\u0026rsquo;t jump into a viral post from the next circle over. Replying there is the same as niche-hopping, and it will pollute your profile too. Mix three types of reply targets by ratio: replies to big accounts are a bet on exposure, replies to mid-sized accounts build relationships, and replies to peers train each other\u0026rsquo;s profiles. You cannot ignore any one of the three completely. 7 Generic posting-time charts are useless. Use your own data to find your windows How the algorithm works:\nDuring Transformer inference, the Phoenix Ranking model uses a candidate-isolation attention mask. Each candidate can see only the user context, not other candidates. This makes scoring pointwise and cacheable, independent of who else is in the same batch. But upstream retrieval through Thunder still pulls content from recent time windows, so timing still matters. What this means:\nThe old logic of \u0026ldquo;post at 8 a.m. or 9 p.m. because there is less competition, so you are more likely to get pushed\u0026rdquo; is dead. How many people post at the same moment no longer affects your score. But whether your target readers are online at that time, and whether they can give you early engagement in the first two hours, still decides whether the post can take off.\nWhat to do:\nThrow away all generic posting-time charts like \u0026ldquo;8 a.m. and 9 p.m.,\u0026rdquo; \u0026ldquo;Wednesday at 8 p.m.,\u0026rdquo; or \u0026ldquo;Sunday noon.\u0026rdquo; Those numbers are based on platform-wide averages and have nothing to do with your account\u0026rsquo;s target audience. Open your own X Premium Analytics dashboard and look at the time distribution of likes and replies on your historical posts. Your data tells you exactly when your target readers are active. No data yet? Two weeks is enough. For seven straight days in week one, publish your main post at a different time each day and record the engagement. In week two, lock in the time slot that performed best. Align your time zone with your target audience. If you are in Beijing but your readers are in North America, post from Beijing afternoon to late night, which maps to the North American workday. If your readers are in China\u0026rsquo;s startup circle, post from 8 p.m. to 11 p.m. If your readers are in Europe, post in Beijing afternoon. Long posts can be published off-peak because dwell time accumulates over several hours. Short punchline posts are highly sensitive to burst engagement in the first two hours, so publish them at your audience\u0026rsquo;s online peak. 8 Grox kills edgy bait, clickbait, and AI garbage How the algorithm works:\nThe source code adds a new grox/ module. It runs classifiers, vectorizers, and task-execution engines specifically for spam detection, post classification, PTOS policy enforcement, and safety classification. Grox is a standalone service that labels posts before the main ranking step. Labeled content enters the demotion and removal pipeline in home-mixer/filters, and the signal is written back into the account profile. Template-like AI content, if Grox does not directly mark it as spam, gets demoted through another path: in main ranking, the Phoenix model predicts low P(retention) and high P(not interested), which trains the system that \u0026ldquo;users dislike this kind of content.\u0026rdquo; These are two separate mechanisms. Grox handles hard policy labels. Phoenix handles user-behavior feedback. What this means:\nEven if your edgy bait or clickbait post lucks into 100,000 views, it is a net loss for the account. The small upside from that hit gets eaten several times over by the drop in account trust. Your next post, and the one after that, enter the recommendation pool carrying that negative record. An account is a slow asset. One bad move can take three months to repair.\nWhat to do:\nThe title and content must match. Do not use shock-style, reversal-style, \u0026ldquo;you won\u0026rsquo;t believe this,\u0026rdquo; or \u0026ldquo;99% of people don\u0026rsquo;t know\u0026rdquo; hooks. The algorithm recognizes these hooks better than humans do. Remember this one-line policy boundary: hate speech, false information, and obvious edgy bait are all off-limits. The machine makes the call directly. There is no human safety net. The bet that \u0026ldquo;I write in Chinese, so moderators won\u0026rsquo;t understand\u0026rdquo; has already lost. If AI drafts the post, you must rewrite it yourself. Do three things: replace empty words like \u0026ldquo;empowerment,\u0026rdquo; \u0026ldquo;closed loop,\u0026rdquo; \u0026ldquo;underlying logic,\u0026rdquo; and \u0026ldquo;top-level design\u0026rdquo; with concrete actions; break up the \u0026ldquo;first, second, finally\u0026rdquo; template and use natural transitions; attach a concrete example to every abstract point. Don\u0026rsquo;t play the game of \u0026ldquo;create deliberate misunderstanding → comments melt down → traffic explodes.\u0026rdquo; Every time you publish one of those posts, your account trust score drops. If one or two of them blow up in the short term, don\u0026rsquo;t get smug. Three months later, your traffic curve will show you the bill. 9 You can chase broad traffic for short-term gain, but long-term distribution still requires a vertical niche How the algorithm works:\nAccount profiles accumulate over time. During cold start, the profile is sparse and the algorithm\u0026rsquo;s judgment of you is imprecise, so broad-traffic content can still enter recommendations. Once the profile stabilizes in the middle and later stages, content that deviates from it gets suppressed immediately. The people who reply to you and follow you also train your profile. What your early followers are like determines who you can be pushed to later. What this means:\nAccount building has stages, and each stage has a completely different goal. Early on, it is fine to use broad traffic to build a profile and collect a base audience. In the middle and later stages, you must tighten into a vertical niche. The most dangerous stage is actually after you get your first sponsored posts or ad revenue. Many people relax their content standards once money comes in, their account signals drop, and the previous months of work get wasted.\nWhat to do:\nFrom zero to 1,000 followers, you can chase broad traffic, ride trends, and use controversial topics to build the account profile and assemble a base. But every week, publish at least two or three strong pieces aimed at your target vertical audience, so broad-traffic content does not drag your profile completely off course. From 1,000 to 10,000 followers, start choosing topics by vertical audience. AI, startups, investing, developers, and tool users tend to have much deeper engagement than entertainment or lifestyle circles, which makes them better for building a durable account. After 10,000 followers, go fully vertical and stop chasing broad traffic. The goal at this stage is not follower growth. It is deepening interaction with existing followers. Deep follower engagement \u0026gt; total follower count. Once you start taking sponsored posts or earning ad revenue, run a content-quality review every week. This is the easiest stage to crash. People loosen up once money arrives, and once account signals drop, it takes one to two months to recover. Set your content bar even higher than before you started making money. 10 Search traffic is the underrated second exposure channel outside the timeline How the algorithm works:\nX Search has its own ranking system, completely separate from For You. Top search results combine three types of scores: engagement, health, which looks at reports, blocks, and policy violations, and relevance, which looks at query matching, post age, author network, and content score. X\u0026rsquo;s official search help page emphasizes this directly. What this means:\nFor You traffic comes fast. After 24 hours, a post is basically dead. Search traffic comes slowly, but it can keep running for months or even years. These are two completely independent ranking systems, which means the same post may keep bringing people in through Search after it has died in For You, as long as you considered Search when writing it.\nWhat to do:\nWhen writing a main post, naturally include entity terms your target users will search for: model names (Claude, GPT, Gemini), tool names (Cursor, Vercel, Supabase), platform names (X, Notion), and scenario terms (algorithm, monetization, SOP). These are the keys to the Search entry point. Don\u0026rsquo;t write only in metaphors and memes. Memes can perform well in For You, but they are not searchable. A post made entirely of jokes has zero long-tail traffic. Produce evergreen content on purpose: guides, comparisons, templates, mistake lists, and term explanations. This kind of content can still be found six months later. Its long-term compounding beats trend-chasing by a wide margin. Put at least one searchable keyword combination into every long post, such as \u0026ldquo;X algorithm Phoenix,\u0026rdquo; \u0026ldquo;Claude Code workflow,\u0026rdquo; or \u0026ldquo;Vercel deployment.\u0026rdquo; These combinations help you hold a long-term position for a specific search query. Don\u0026rsquo;t stuff hashtags. X Search looks at natural-language matching, not tag matching. In 2026, hashtags are only classification labels. They do not help search weight. Final note The biggest value of this source-code release is that it maps three years of speculation to real code.\nAfter reading this, do three things immediately:\nRescore your posts from the last 30 days by replies, reposts, quotes, dwell time, profile visits, and new followers. See which layer your content is actually scoring on. If you are only scoring on likes, the content still needs work. Check whether you have switched niches, spammed posts, done mutual-follow games, used edgy bait, touched politics, or posted AI-template formatting in the last 30 days. Avoid these as much as possible going forward. Change your account goal from follower count to \u0026ldquo;followers from the right vertical audience + replies/reposts/bookmarks + dwell time.\u0026rdquo; This is also why high-quality video has been getting so much distribution recently. Dwell time is a major metric on social platforms. ","permalink":"https://ryrenz.com/tech/x-recommendation-algorithm-open-source/","summary":"After X open-sourced the For You recommendation algorithm, creators should stop obsessing over likes and focus on replies, reposts, retention, account profile, reply leverage, and search long tail.","title":"X Open-Sourced Its Latest Recommendation Algorithm Today. Here Are 10 Things Every Creator Needs to Know"},{"content":"1. Where AI Still Falls Short The authors of SWE-Bench just released a set of numbers that should make engineers uncomfortable.\nTheir new hardcore benchmark, ProgramBench, asks AI to rebuild real open-source software projects from scratch. No internet access. No judging by code similarity. Only final behavior counts.\nThe result: Claude Opus 4.7, GPT-5.4, Gemini 3.1 Pro — every frontier model scored a 0% completion rate.\nTo be clear, this does not mean AI cannot write code. It can write plenty of code. It can make individual functions look beautiful.\nBut ask it to build a real project from zero that actually runs, and it stuffs all the logic into one monolithic file. No modularity, no architecture, no long-term plan. In the end, it fails behavioral verification.\nWhat does that tell us?\nCode generation is no longer the bottleneck. The bottleneck is system architecture and engineering practice.\nNow look at another data point.\nLangChain ran an experiment. They used the same gpt-5.2-codex model without changing a single weight. They only optimized the engineering structure around the model. The result: their coding agent went from 52.8 to 66.5 on Terminal-Bench 2.0, jumping from outside the Top 30 into the Top 5.\nOpenAI has shared an even more extreme case. A three-person team used Codex for five months to write roughly one million lines of code and merge about 1,500 PRs. This was not a demo. It was a real software product with internal daily active users and external alpha testers. OpenAI’s team said their focus had shifted from writing code to four things: designing the environment, making intent explicit, building feedback loops, and letting the Agent see, verify, and repair its own work.\nPut these two data points together and the conclusion is obvious:\nRaw models have a 0% completion rate on serious engineering tasks. Add the right engineering infrastructure around the same model, and a three-person team can build production-grade software.\nThe gap between those two outcomes is the Harness.\n2. What Is a Harness? The literal meaning of harness is the gear used to control a horse.\nA large model, whether Opus 4.7 or gpt 5.5, is like a wild horse. If you cannot control it with reins, its speed means nothing.\nHarness Engineering is the craft of building that gear and using it to control the horse.\nHooks, Skills, MCP, CLAUDE.md / AGENTS.md, sub-agents, plugins, tools — you have probably heard of these or used a few of them. Harness is the umbrella term for designing all of them as one system.\nA lot of people use AI like this: they see a new MCP that looks fun and install it; they see a hook example and copy it; they write two random lines in CLAUDE.md and forget about it; when a bug appears, they have no idea which part to fix. The features do not work together, and too much harness can tire out your horse. The effect is limited.\nThe point of the term Harness Engineering is to force you to look at these scattered actions as a system.\nIt gives you a checklist: is your AI workflow complete across five core dimensions? Is your horse always in the best possible condition?\nThe industry has offered a simple formula:\nAgent = Model + Harness\nThat is why some tools feel so smooth while others have strong models but feel brainless in practice. The harness is different.\nSeen in the evolution of AI engineering:\nThe innermost layer is Prompt Engineering, which is about how to instruct the AI. The middle layer is Context Engineering, which is about what information to give the AI and when to give it. The outer layer, Harness Engineering, wraps both of those and adds tool orchestration, state persistence, verification loops, task decomposition, sub-Agents, permission sandboxes, and rollback mechanisms. Together, they form a complete engineering infrastructure. Next, we will use the five dimensions of Harness to look back at your current Claude Code / Codex / OpenClaw / Cursor setup. You will get a clearer sense of whether your Agent needs more gear, or whether it needs less weight so it can run lighter.\n3. The Five Core Dimensions of Harness To turn the scattered idea of Harness into concrete engineering actions, we can break it into five dimensions:\ncontext management, execution capability, task orchestration, feedback loops, architectural guardrails\n3.1 Context Management: A Three-Layer Memory Architecture AI forgets project rules in long conversations because of how the context window works. In each turn, the model sees a flat list of messages. Everything you said earlier is mixed together with the current question. There is no hierarchy between project rules and casual chat. The longer the conversation gets, the more diluted the earlier constraints become.\nOpenAI made this clear in its Codex engineering article: from the Agent’s point of view, knowledge that is not available at runtime does not exist. Rules you said out loud, discussed in Slack, or assumed as team convention are invisible to AI unless they live in the repository as files.\nThe core Harness move at this layer is to turn rules into structured files. In production practice, this usually has three layers.\nThe first layer is AGENTS.md or CLAUDE.md in the project root. Think of it as the project map. It gets loaded at the top of the context in every new conversation. It includes the tech stack, directory structure, things that are forbidden, commands that must be run before committing, and UI style no-go zones. Keep it to about 100 lines.\nThe second layer is detailed rule files split by topic under the docs directory, such as frontend.md, security.md, and api-design.md. The AI sees the pointers in AGENTS.md and reads these files when needed, instead of stuffing all of them into context at once.\nThe third layer is the built-in memory mechanism in tools like Claude Code: lightweight indexes of roughly 150 characters each are always loaded; detailed files are pulled in on demand; raw records are accessed only through search tools like grep. This design balances two needs: the AI can see the big picture, while the context window does not get blown up.\nOpenAI admitted one mistake in its article. At first, they put all their rules into one giant AGENTS.md that ran for thousands of lines. The AI became more likely to ignore key information. They fixed it by switching to a two-layer structure: a map plus detailed documents.\nCore principle: persist context in layers, and avoid consuming so much context that you damage the model’s attention.\n3.2 Execution Capability: Giving the Model Hands and Feet A model by itself can only output text. It can tell you to run npm install in the terminal, but it cannot run the command, inspect the result, or adjust the next step based on the error. Pure output leaves AI unable to close the loop. At every step, a human has to act as the messenger.\nThe Harness work at this layer is to connect the model to a real operating environment. There are three levels, from basic to advanced.\nThe basic layer is terminal plus file system plus browser. The terminal lets AI run commands, install dependencies, execute tests, and inspect logs. The file system lets it read code, modify files, and write intermediate documents. The browser lets it view real pages, click buttons, and verify with screenshots.\nThe advanced layer is MCP. MCP is a standard protocol that lets AI connect to external capabilities. Common targets include databases, search engines, crawlers, design tools, and monitoring systems.\nThe higher layer is Skills. Skills package multi-step workflows into reusable capability bundles: writing a technical tweet, generating a weekly report, scraping competitor data from a website. When the AI sees the matching request, it invokes the whole Skill directly instead of redesigning the steps every time.\nBut you cannot pile on tools forever. Vercel hit a counterexample while building an internal text-to-SQL Agent. At first, they built a stack of specialized tools: schema lookup, query validation, and error recovery. The success rate was 80%. Later, they deleted 80% of the specialized tools and let Claude use basic Unix tools like grep, cat, find, and ls to read files and write SQL by itself. The success rate rose to 100%, speed improved by 3.5x, and token usage dropped by 37%.\nThe reason is simple: the more tools you give the model, the larger its choice space becomes at every step, and the more likely it is to pick the wrong tool or go down the wrong path.\nCore principle: choose every tool deliberately. Fewer and sharper beats more and messier.\n3.3 Task Orchestration: Letting AI Handle Long Tasks The most common failure mode on long tasks is that AI tries to one-shot an entire feature. The model’s context window is limited. Asking it to swallow a requirement like “a list page with search, filters, and pagination” all at once is like asking an engineer to skip the design doc, skip task decomposition, skip iterations, and just grind until the end. Failure is guaranteed.\nAnthropic described this failure mode in detail in its article on long-task Harness: the model tries to write everything in one pass, runs out of context, realizes halfway through that the earlier plan was wrong, then goes back to rewrite earlier code. The more it changes, the more tangled things get.\nThe Harness work at this layer is to structure long tasks.\nStep one is Plan Mode. The AI first outputs an implementation plan: which subtasks to split into, and how each subtask will be implemented. A human confirms the plan before the AI starts editing. This is the brake pedal. It moves the cost of going in the wrong direction up into the planning stage.\nStep two is step-by-step execution. Do one subtask at a time, and verify each one after completion. This prevents the collapse that happens when the context fills up midstream.\nStep three is externalized state. After each feature is completed, ask the AI to write a document, usually called progress.md or plan.md. It should state in detail what has been completed, which technical approach was used, which key architectural decisions were made, which bugs remain unresolved, and what is still left to do. This document is external memory across context windows. When a new conversation starts, the AI reads it and immediately gets back into the work.\nStep four is parallelism. Use sub-agents to run independent subtasks at the same time.\nAnthropic proposed a more advanced solution for long-task engineering called Ralph Loop. It is basically a two-stage relay.\nThe first stage is the Initializer Agent. It runs only once at the beginning of the project. Its job is to set up the development environment, break the full requirement into a feature list, write the first progress.md, and make the initial git commit.\nThe second stage is the Coding Agent. This runs in every new conversation window. Its fixed routine is: read git log to understand the commit history; read progress.md to understand current progress and where the previous round stopped; pick the highest-priority item from the unfinished feature list; complete it; then commit to git, update progress.md, and clearly write what was done in this round and what the next round should do.\nEven if the AI gets interrupted, the model version changes, or the conversation window fills up, the next round can read git log and progress.md and immediately get back into state.\nCore principle: progress.md and git commits are save points for AI. Save state before you trust it with long tasks.\n3.4 Feedback Loops: AI Test-Driven Development Models do not run code. They read the code text and judge whether it looks like it should work. If a piece of code follows a familiar pattern, the variable names line up, and the indentation looks clean, the model marks it as done. But whether code runs has nothing to do with whether it looks runnable. Only running it can prove that. This is why AI often says “fixed” with total confidence while the project is still full of errors.\nThe Harness work at this layer is to move verification from humans to automation. There are three kinds of feedback.\nRule feedback: make the AI automatically run the linter, typecheck, unit tests, and integration tests before every commit. If any one of them fails, the task is not done.\nVisual feedback: for UI tasks, make the AI use tools like Playwright to open the browser, click through the user path, and produce screenshots as evidence of completion.\nLLM review feedback: make another AI independently review the code that was just written, looking for logic holes, architecture problems, and potential bugs.\nAnthropic gave a concrete number in its engineering blog: giving the model a loop to verify its own work can improve output quality by 2x to 3x. This is the single most reliable Harness investment.\nThere is one counterintuitive design point here. Letting the same AI that generated the code review its own work performs much worse than you might expect, because the generator is naturally biased toward defending itself. Anthropic’s experience is to split the generator and evaluator into two independent Agents with different role configurations and different prompts. Real peer review finds problems. This is the same reason traditional software teams say you should not be the only reviewer of your own code.\nCore principle: AI saying it fixed something means nothing. Passing tests is what counts.\n3.5 Architectural Guardrails: Blocking Bad AI Code Before It Lands AI-written code has a hidden problem: it imitates the patterns already present in the repository. Good code gets imitated. Bad code gets imitated too. Each individual AI commit may look reasonable in isolation, but stacked together, the project keeps getting worse.\nOpenAI called out this behavior in its Codex article: Agents copy existing patterns in the repository. If the existing patterns are unstable, inconsistent, or simply bad, AI amplifies them.\nThe Harness work at this layer is to move architectural rules out of documents and into executable code, so bad code gets blocked automatically before it enters the main branch.\nThe most basic layer is pre-commit hooks. Before a git commit, they automatically run a batch of checks. Noncompliant commits are blocked immediately.\nThe second layer is an architecture linter. It checks architectural violations specifically: the UI layer cannot directly access the database layer, module dependencies must flow in one direction, files above a size threshold must be split. This is different from the syntax linter in 3.4. That one checks syntax errors; this one checks architecture errors.\nThe third layer is the CI gate as a backstop. Even if local hooks are bypassed, CI runs the checks again to ensure the main branch always satisfies architectural constraints.\nOpenAI also has a more aggressive practice it calls “garbage collection”: periodically run background Codex tasks to scan the whole codebase, find places that drift away from architectural principles, and automatically open small PRs to pay down technical debt. The logic is that AI creates code faster, so technical debt accumulates faster. Debt cleanup has to be automated too.\nCore principle: account for AI’s limits ahead of time, and set up guardrails.\n4. How Top Teams Build Harnesses 4.1 Anthropic / Claude Code: Textbook Long-Task Engineering Claude Code is Anthropic’s own agent harness for driving Claude. It breaks Harness into 12 independent components. I will not list all of them here. Instead, let’s focus on the two designs most worth copying.\nDesign 1: A three-layer memory architecture\nClaude Code’s memory system has three layers.\nThe top layer is a lightweight index. Each item is about 150 characters and is always loaded into context. Its job is to make sure the AI always knows what files, modules, and core conventions exist in the project. Because each item is short, dozens or even hundreds of them will not blow up the context window.\nThe middle layer is detailed files: README.md, ARCHITECTURE.md, API docs, and design notes for each module. These files are not loaded by default. When the AI sees a lightweight index item that says “see docs/architecture.md,” it reads that file only when needed. After the information enters context, it can also be compressed out when it is no longer needed.\nThe bottom layer is raw records: the full git log, complete conversation history, and complete log files. This layer is never loaded automatically. The AI can only retrieve it through commands like grep and tail. It has the most data, but it does not pollute context.\nThe core idea behind these three layers is simple: context is scarce, so it must be layered by “how often it needs to be read” and “how important it is.” Put what must always be seen at the top. Put what is occasionally needed in the middle. Leave everything else to search.\nDesign 2: Ralph Loop for long tasks across context windows\nFor long tasks that take days and span multiple conversation windows, Anthropic designed a workflow called Ralph Loop. It is basically a two-stage relay.\nThe first stage is the Initializer Agent. It runs only once at the beginning of the project. It does four things: sets up the development environment, breaks the full requirement into a feature list, writes the first progress.md with the todo items, and makes the initial git commit.\nThe second stage is the Coding Agent. It runs in every new conversation window. Its fixed process is:\nRead git log first to see the commit history. Read progress.md to see current progress and the unfinished list. Pick the highest-priority item from the unfinished list. Complete that item, run verification, and commit to git. Update progress.md with what this round did and what the next round should do. Even if the AI gets interrupted, the model version changes, or the conversation window fills up, the next AI only needs to read git log and progress.md to get back into state. The key to this design is not that the AI is smarter. It is that progress.md and git history together create external memory across context windows.\nWhat you can learn from this design:\nYour own project should at least have a progress.md that clearly states what is done, which key architectural decisions were made, which bugs remain unresolved, and what should happen next. Do not leave this stuff only in your head. Treat each completed feature as one git commit. Git history itself is an AI-readable work log. Layer your context. Put what the AI must always see in AGENTS.md, put what it may need to look up in docs, and leave the rest to search. 4.2 OpenAI / Codex: Rebuilding the Development Environment for AI Behind the story of OpenAI’s Codex team using AI to write one million lines of code in five months with three people, the most important factor was not model strength. It was an engineering idea they call Codex legibility.\nCodex legibility roughly means “readable by Codex.” The point is that future codebases must be readable not only by humans, but also by Agents. That means all the development infrastructure originally built for human engineers — logs, monitoring, debugging tools, local environments — has to be rebuilt into forms AI can use directly.\nOpenAI had five concrete practices.\n1. Each git worktree automatically starts an independent app instance. When Codex works on a change, the corresponding branch can automatically spin up an independent development server. The AI can open that instance, operate it, inspect responses, and verify whether the change actually worked as intended. What used to require an engineer to manually run npm run dev is now done by the AI.\n2. Chrome DevTools Protocol is connected to the Agent runtime. This gives AI the same browser-level debugging ability as an engineer: inspect the DOM, listen to network requests, inject JS, take screenshots. Reproducing a UI bug no longer requires an engineer to open the browser by hand. The AI can reproduce and locate it by itself.\n3. Logs, metrics, and traces are exposed for AI queries. OpenAI lets Codex directly access production monitoring data through query languages like LogQL, PromQL, and TraceQL. That means the AI does not need an engineer to manually paste logs into the chat during debugging. It can grep logs, inspect metric anomalies, and trace requests by itself.\n4. Custom linters turn architectural constraints into executable rules. OpenAI encoded which layers cannot call which layers, which names must be consistent, and which patterns are forbidden as automatic linter rules. Any AI-written code that violates those rules gets blocked.\n5. Background garbage collection tasks. This is the most interesting one. OpenAI periodically runs a batch of background Codex tasks to scan the entire codebase for places that drift away from architectural principles, then automatically opens small PRs to fix them. In other words, paying down technical debt is automated too. It no longer depends on engineers manually refactoring code.\nTogether, these five practices turn the whole development environment into an AI-readable, AI-operable, AI-verifiable workbench.\nWhat you can learn from this design:\nIf your project logs are unreadable even to humans, AI will understand them even less. Refactoring logs into a structured format, such as JSON with a fixed schema, is the first step toward making them AI-friendly. Move the architectural principles in your head, in Slack, and in code review comments into linter rules. The real value is not only preventing violations. It also turns architecture into something learnable: AI can read the linter error and learn how to write compliant code. During system design, ask one question: can AI see this state? If not, figure out how to expose it. 4.3 Nous Research / Superpowers: An Open-Source, Reusable Skills Framework Superpowers is an open-source Agent Skills framework from Nous Research. If Claude Code is Anthropic’s internal engineering practice, and Codex is OpenAI’s model for rebuilding the development environment, then Superpowers is the open-source version individual developers can copy and use at home.\nIt includes several common workflows out of the box.\nTDD workflow: AI writes tests before implementation. The work is not done until the tests pass. This turns the feedback loop from 3.4 into the default path. AI must pass tests before it can deliver.\nTwo-stage Code Review: the first stage has the generator agent write code; the second stage has the reviewer agent review it. The two agents use different role configurations and prompts, forcing a real split between generator and evaluator.\nSub-agent collaboration template: it includes a standard flow for task decomposition, parallel execution, and result synthesis, ready to use out of the box.\nHermes Agent, the sister project of Superpowers, has an even more forward-looking direction called Self-Evolution. This system uses DSPy and GEPA to continuously optimize the Harness itself. Put simply, DSPy treats prompts and tool descriptions as optimizable parameters, while GEPA uses a genetic-algorithm-like approach to find better configurations from the success and failure traces in execution records. The whole process does not retrain the model. It only tunes the Harness. In other words, Harness can automatically improve from the tasks it has done before.\nWhat you can learn from this design:\nIf you do not want to design a TDD workflow and two-stage review from scratch, copy the Superpowers templates and start there. Package the few task types you do most often, such as writing a blog post or analyzing a dataset, into Skills. Next time, the AI can call them by itself instead of making you teach the workflow again. Pay attention to the DSPy / GEPA line of work. Stronger models are the big trend, but self-improving Harness may be the improvement path closer to your actual work. 5. Three Counterintuitive Harness Principles Now that we have covered how top teams build Harnesses, let’s pull out a few principles in this field that are especially easy to get wrong.\nPrinciple 1: More Harness Is Not Always Better Intuitively, more tools should make AI stronger. In reality, the more tools you give it, the larger the choice space becomes at every step, and the more likely it is to pick the wrong tool, take the wrong path, or call an interface it should not call.\nVercel tested this directly on its internal text-to-SQL Agent. At first, they gave the Agent a full set of specialized tools: schema lookup, query validation, and error recovery. The success rate was 80%. Later, they removed 80% of those specialized tools and let Claude use basic Unix tools like grep, cat, find, and ls to read schema files and write SQL by itself. The success rate rose to 100%, speed improved by 3.5x, and token usage dropped by 37%.\nManus observed the same pattern: a heavily armed agent gets dumber.\nThe practical rule is simple: before adding any tool, ask what specific behavioral gap it solves. If you cannot answer, do not add it yet.\nPrinciple 2: How You Organize Context Matters Much More Than How Much Context You Have Many people see Claude 4 Opus has a million-token context window and their first reaction is, “Then just stuff everything in.” That is an expensive misunderstanding.\nChroma Research ran a Context Rot study testing 18 mainstream models under long-context conditions. The conclusion: as input length grows, models become less reliable at using context. Even when task difficulty stays the same, simply increasing context from 10,000 tokens to 500,000 tokens significantly reduces performance.\nStanford’s famous Lost in the Middle paper proved another phenomenon: models use information near the beginning and end of the context best, and tend to ignore information in the middle. That means the position of key information in the context directly affects whether AI can use it.\nThe Manus team added one more cut from an economics angle. In their typical Agent tasks, the ratio of input tokens to output tokens is about 100:1. Most of the Agent’s time and cost is spent repeatedly feeding context into the model, and the price difference between KV-cache hits and misses is 10x. For the same task, context organization that allows cache reuse can change your monthly API bill by an order of magnitude.\nPut these three things together and the message is clear: context design is a real engineering problem. It is not as simple as throwing everything in and letting the model pick what matters.\nPrinciple 3: Harness Should Get Simpler as Models Improve Behind every Harness component is an implicit assumption: the model cannot do this by itself, so it needs an external patch. But models are improving fast, and old patches may already be obsolete.\nAnthropic gave a concrete example in its Harness Design blog. In early versions of Claude Code, they added an explicit “planning” step that forced AI to output a plan before acting. Later, after a new version of Claude was released, Anthropic found that planning ability had been internalized by the model. The external planning step became redundant overhead, so they deleted it.\nThe practical method is this: every time you upgrade the model version, review your Harness. Which components still compensate for real capability gaps? Which ones are just historical baggage? Which checks can now be trusted to the model?\nA good Harness keeps just enough thickness outside the current boundary of model capability.\nConclusion: Harness Engineering Is Becoming a Core Skill The fact that every large model scored 0% on ProgramBench tells us something in reverse: Harness Engineering is becoming more important.\nAI can already write a large share of your code today. But AI will not do requirements analysis for you, decompose tasks for you, design the context architecture for you, decide which tools to add or remove, establish verification standards, or defend architectural boundaries for you.\nThe sum of those things is Harness Engineering.\nIt sounds like a new term, but at its core it is the systematic migration of traditional software engineering methods into AI workflows:\nAGENTS.md corresponds to requirements documents and design documents. Task orchestration corresponds to iterative decomposition in agile development. Feedback loops correspond to unit tests and code review. Architectural guardrails correspond to code standards and security checks. These were already the fundamentals of good engineering.\nNext time your Agent messes things up again, look back at your Harness across the five dimensions and see where it stands.\nIn most cases, the AI is already smart enough. What is missing is a good set of reins.\n","permalink":"https://ryrenz.com/ai/harness-engineering-ai-efficiency/","summary":"Raw models are no longer enough for serious engineering work. What actually separates mediocre results from great ones is the Harness built around the model: context management, tool execution, task orchestration, feedback loops, and architectural guardrails.","title":"Harness Engineering Explained: How to Make Your AI 10x More Effective"},{"content":"Some things never get cleaned up, not because they are especially good at hiding.\nThey survive because they make too much money.\nYou may think porn livestreaming is just a woman streaming on her phone while a man hides behind a screen and sends tips. But once you break down the public court cases, it is obvious that the business is not that simple.\nIt is more like an assembly line that turns vulgarity into a system.\nLook at two real cases from court rulings:\nA 20-year-old college student surnamed Lin, short on living expenses and wanting a new phone, made 120,000 yuan in one month through illegal adult livestreaming.\nA recent graduate surnamed Zhou, tired of slow wages and hard office work, started porn livestreaming with her boyfriend. In half a year, their turnover reached 1.44 million yuan.\nThe truly frightening part is how fast ordinary people can make money from it.\nShort on cash. Sick of work. Wanting quick money. Thinking it is only showing a little. Thinking nobody will recognize you through a screen.\nOnce people taste that money, it is hard to go back to their old lives.\n1 What is it, really? Porn livestreaming is not really livestreaming.\nIt is an illegal transaction system dressed up as livestreaming.\nA complete chain usually includes several types of players:\nThe platform operator, which handles the app, website, servers, payment, virtual coins, livestream rooms, and membership system.\nAgents and guild leaders, who recruit streamers, bring in users, manage streamer hours, and settle revenue shares.\nStreamers, who provide the content, chat with users, and push them to tip. Many are also packaged into specific personas.\nPayment and technology providers, who may appear to be simple tool vendors, but if they knowingly support the use case, they can be pulled into criminal risk.\nUsers, who look like consumers but are often the easiest people to keep harvesting.\nIn public cases, the way the \u0026ldquo;Aimeiren\u0026rdquo; platform operated was typical: the platform recruited guild leaders first, then guild leaders recruited female streamers. After registering as members, users topped up \u0026ldquo;diamonds\u0026rdquo; and used virtual gifts to tip. The platform, guild leaders, and female streamers settled shares daily according to fixed ratios.\nThat is the core business structure of porn livestreaming:\nSplit one impulse purchase into several layers of headcount-based income.\nUsers think they are sending gifts to a streamer. The platform sees a cash flow that can be split, repeated, redirected, and harvested again.\n2 How does it make money? If you only say \u0026ldquo;it makes money from tips,\u0026rdquo; you are describing the business too shallowly.\nTips are just the storefront.\nThe real money comes from the payment gates behind it.\nAfter breaking down public cases, I can see roughly nine payment gates.\n2.1 Membership codes: turning porn content into monthly, quarterly, and annual passes The Max case is the most typical card-code model.\nIt was not just a livestreaming platform. It was an aggregator app for porn livestreams. A report from the CPPCC Daily, reprinting Procuratorate Daily, described it clearly: \u0026ldquo;Xiaoqiang Magic Box\u0026rdquo; started with only livestreaming, then later added cloud playback, movies, and other sections, integrating links to many porn livestreaming platforms and porn websites into its own platform. To use it, users had to register and buy recharge card codes. Those codes were divided into monthly, quarterly, and annual passes, then sold through agents.\nThat is the first layer of monetization:\nDo not produce all the content yourself. Package a pile of porn content into a membership entrance.\nThe logic is very similar to legitimate video-platform memberships. It is just selling illegal content.\nMore importantly, Max\u0026rsquo;s card-code sales were not small. The report said that in the early days, the platform sold around 10,000 monthly passes per month. Later it sold 100,000 per month, and eventually even 200,000 to 300,000 per month. Arrested agents said card-code transactions started at 40,000 to 50,000 yuan per day and later reached 70,000 to 80,000 yuan per day.\nThis is the harshest part of the card-code model:\nContent can be replaced. The app can change its name. Agents can switch accounts. But as long as the card-code system remains, the cash flow keeps running.\n2.2 Agent wholesale: the people really getting fat are the ones selling shovels Many people think the platform only makes money from end users.\nNot quite.\nIt also makes money from agents.\nPeople\u0026rsquo;s Daily Online\u0026rsquo;s report on the Max case mentioned that one arrested agent bought card codes wholesale from an upstream seller at a lower price, then marked them up and sold them to lower-level agents. His monthly card-code purchases reached as much as 1.5 million yuan.\nThat detail matters.\nIt shows that Max was not simply \u0026ldquo;selling memberships to viewers.\u0026rdquo; It was running a multi-level distribution network.\nThe platform wholesaled card codes to general agents.\nGeneral agents sold them to lower-level agents.\nLower-level agents sold them to users.\nEvery layer took a spread. Every layer had an incentive to keep spreading the business.\nSo a small ad, a WeChat group link, or a forum post may not be tied to some isolated porn site. Behind it may be a distribution chain of card-code wholesale, agent markup, and end-user consumption.\nThat is also why Max could grow to more than 16,000 agents, over 3.5 million members, and 250 million yuan in funds involved in such a short time.\nThe platform does not get rich from one user.\nIt gets rich by turning every agent into a salesperson.\n2.3 Virtual coins and gifts: disguising payment as interaction Users top up first. The platform turns renminbi into diamonds, tomato coins, gifts, or other virtual items. That step matters because once money becomes in-platform currency, users feel less pain from the real amount.\n10 yuan, 50 yuan, 200 yuan become strings of numbers that hurt less.\nWhy do livestreaming platforms love virtual coins? Because they turn \u0026ldquo;payment\u0026rdquo; into \u0026ldquo;interaction.\u0026rdquo;\nPayment is rational.\nInteraction is emotional and addictive.\nIn the cross-border porn livestreaming platform case cracked by Cangnan police, reports said \u0026ldquo;tourists\u0026rdquo; bought \u0026ldquo;diamonds\u0026rdquo; through official or individual agents, spent diamonds to enter rooms and watch porn performances, and also bought expensive virtual gifts such as \u0026ldquo;rockets\u0026rdquo; to tip streamers.\nThat is not a simple payment action.\nIt breaks the transaction into three steps:\nFirst, top up. Money becomes diamonds.\nThen enter a room. Diamonds become viewing access.\nThen send gifts. Diamonds become emotional interaction.\nEach extra layer weakens the user\u0026rsquo;s sense of real loss.\nThat is the real value of virtual coins.\nThey do not make payment easier.\nThey make it easier for people to lose control.\n2.4 Private rooms and paid rooms: turning \u0026ldquo;one step further\u0026rdquo; into a paid checkpoint Some platforms do not make you send gifts forever. They give you a little free content first, then put the real paid content behind memberships, private rooms, and paid viewing.\nIn the Jiaxing \u0026ldquo;Max\u0026rdquo; case, public reports said the platform illegally profited by developing layers of agents and bringing in users to buy membership card codes. These card codes worked like recharge cards. Only after buying them could users enter deeper viewing layers.\nFree content is bait.\nMembership card codes are the closing mechanism.\nThe core here is not that the content is sophisticated.\nThe core is cutting the user\u0026rsquo;s curiosity into steps.\nFirst layer: free.\nSecond layer: preview.\nThird layer: membership.\nFourth layer: private room.\nFifth layer: one-on-one interaction.\nEach time users cross a layer, they have already spent a little money and a little psychological cost.\nOnce they reach the later layers, they are no longer casual viewers. They are paying users the platform has already filtered out.\n2.5 Streamer revenue shares: the platform sells front-end assets that can be mass-produced Streamers are not the whole platform, but they are the platform\u0026rsquo;s most important front-end assets.\nIn the \u0026ldquo;Tomato Community\u0026rdquo; case, reports said the platform and streamers split revenue 40/60. In cases reported by People\u0026rsquo;s Court Daily, there were also structures where platforms and streamers split revenue in ratios around 40/60.\nWhat does this show?\nPorn livestreaming platforms do not want one superstar streamer most of all. They want a supply of streamers that can be copied at scale.\nIf one streamer can make money, the platform will have agents keep recruiting.\nIf one agent can recruit, the platform will keep giving him a cut.\nIn the end, it stops being a \u0026ldquo;livestreaming platform\u0026rdquo; and becomes a recruitment system.\nLook more closely at the Cangnan cross-border porn livestreaming case. Female streamers under guild leaders received 50% to 60% of daily income. The money was first transferred to guild leaders through bank cards, then guild leaders settled shares with the female streamers.\nThis shows the platform does not directly manage every streamer.\nIt outsources streamer management to guild leaders.\nGuild leaders recruit people, train them, monitor hours, provide scripts, and settle payments.\nThe platform handles technology, top-ups, traffic, the backend, and revenue splitting.\nWhat does this model resemble most?\nAn illegal outsourced sales team.\nStreamers close deals at the front desk.\nGuild leaders manage the team.\nThe platform runs the cashier.\n2.6 Guild-leader cuts: the further upstream you go, the more it looks like a headcount business The truly powerful part of porn livestreaming platforms is that they let not only streamers make money, but also the people who recruit streamers.\nThat creates a dangerous incentive:\nAs long as you can bring in streamers, you can keep taking a cut from their income.\nSo what guild leaders care about most is often not whether one streamer becomes popular, but whether they can keep expanding the streamer pool.\nIf one streamer starts making money, she is encouraged to bring in friends.\nIf one friend joins, classmates, roommates, and hometown acquaintances get pulled in too.\nIn the end, the chain shifts from \u0026ldquo;content performance\u0026rdquo; to \u0026ldquo;acquaintance recruitment.\u0026rdquo;\nThat is also why young women, unemployed women, and cash-strapped students repeatedly appear in public cases.\nIt is not that they are naturally worse people.\nIt is that they are easier to hit with the pitch of \u0026ldquo;high pay, freedom, fast money.\u0026rdquo;\n2.7 Paid on-demand: beyond livestreams, there are secretly filmed videos and inventory content Porn livestreaming platforms do not necessarily rely only on real-time streams.\nThe \u0026ldquo;Huanggua Video\u0026rdquo; case is another form.\nAmong the 2021 top ten anti-pornography and anti-illegal-publication cases published by Zhejiang Online, the \u0026ldquo;Huanggua Video\u0026rdquo; app had more than 60 million domestic installs and made over 300 million yuan in illegal profit. The chain included overseas R\u0026amp;D and operations teams, domestic video-uploading creators, guild leaders, fourth-party payment-channel gangs, and technical support.\nPay attention to the keywords here:\nVideo-uploading creators.\nGuild leaders.\nFourth-party payment channels.\nTechnical support.\nThis shows it was no longer just \u0026ldquo;streamers go live and viewers tip.\u0026rdquo;\nIt mixed paid on-demand, live performances, user uploads, payment settlement, and technical operations into one business.\nFor the platform, livestreaming is instant stimulation. On-demand content is inventory.\nLivestreaming gets people worked up.\nOn-demand lets the same content make money again and again.\n2.8 One source, many sales: pushing the same content across multiple apps and websites The Cangnan cross-border porn livestreaming case has another detail worth spelling out.\nPolice found that the livestreaming platform shared the same video sources with multiple obscene websites and livestreaming platforms inside and outside China. In other words, users could see the same livestream content across multiple apps and websites, while the operator behind them was the same criminal gang.\nThat is \u0026ldquo;one source, many sales.\u0026rdquo;\nThe same streamer.\nThe same livestream.\nThe same batch of content.\nChange a few app names, change a few website entrances, change a few sets of promotional copy, and it can be sold to different traffic pools.\nThat is also why some porn sites look wildly different on the surface, while the same group may be behind them.\nMany front desks.\nOne backend.\nThis is the biggest difference between a digital brothel and a traditional porn venue:\nA traditional venue has limited space.\nA digital venue can clone its entrances without limit.\n2.9 Reselling porn traffic: gambling, fake-order scams, and nude-chat extortion are the second half The dirtiest money in porn livestreaming is often not in the first tip.\nIt is in resale.\nLegal Daily reported a case in which \u0026ldquo;porn livestreaming hid gambling gangs\u0026rdquo;: porn livestreams became a traffic funnel for overseas gambling and money-laundering groups. The chain included livestreaming platforms, gambling platforms, fourth-party payment platforms, phone-bill top-ups, SMS-code receiving, and other links.\nThe Beijing News also investigated similar \u0026ldquo;online pornography\u0026rdquo; traps: criminals posted porn ads through porn websites, social apps, video apps, text messages, and other channels. They used porn videos and \u0026ldquo;free hookups\u0026rdquo; as bait, pushed victims to download apps, then made them perform \u0026ldquo;tasks\u0026rdquo; such as fake orders, investment, or gambling. At first they gave small rebates. Later they used excuses such as system errors, data errors, and operational mistakes to make victims keep putting in money.\nThis is the cruel second half of porn livestreaming traffic:\nStep one: use porn to filter out people who are easy to excite.\nStep two: use interaction to filter out people willing to pay.\nStep three: use off-platform apps to filter out people willing to keep believing.\nStep four: send those people into gambling, fake-order scams, and nude-chat extortion.\nSo a lot of the time, the platform does not care whether you become a long-term user.\nIt only needs to confirm one thing:\nAre you easy to lead around by desire?\nIf the answer is yes, you will be sent into the next, more expensive chain.\n2.10 Contracts and arbitration: even streamers can become a revenue source There is another money gate people often overlook: the streamers themselves.\nBeijing Daily, citing Procuratorate Daily, reported that some livestreaming guilds use slogans like \u0026ldquo;high pay\u0026rdquo; and \u0026ldquo;comfortable work\u0026rdquo; to attract people. After signing them, they provide no support and even require streamers to do porn streams or gambling streams. Some guilds also plant traps in contracts through resignation restrictions, exclusivity clauses, and huge breach-of-contract penalties, then use arbitration or lawsuits to profit from streamers.\nPut this next to the porn livestreaming black market, and the irony is obvious.\nStreamers think they are there to make money.\nIn reality, they may be harvested by the platform, guild leaders, users, and contracts at the same time.\nThe front end harvests viewers.\nThe backend harvests streamers.\nThat is the real structure of the \u0026ldquo;digital brothel.\u0026rdquo;\nIt is not a one-way exploitation system.\nIt is a multi-directional extraction machine.\n2.11 Agent distribution Agents are the most dangerous layer of these platforms.\nWhat ordinary people see is a link, a QR code, an avatar, or a vague line of copy.\nWhat the platform sees is a low-cost distribution network.\nIn the Jiaxing \u0026ldquo;Max\u0026rdquo; case, the platform developed more than 16,000 agents. After the first police crackdown, the platform continued recruiting agents and promoting related apps on WeChat, QQ, Weibo, and other social platforms.\nThat is why it is so hard to kill in one strike.\nThe platform can shut down.\nAgents can switch accounts.\nLinks can change shells.\nUser traffic keeps moving in the dark.\n2.12 Off-platform diversion This layer is the easiest to underestimate.\nMany porn livestreaming platforms are not satisfied with tips. They also divert users off-platform to social apps, private groups, gambling sites, scam apps, and nude-chat extortion.\nCCTV\u0026rsquo;s report on the \u0026ldquo;Max\u0026rdquo; case said the platform colluded with gambling gangs and used porn traffic to promote gambling websites.\nDouyin\u0026rsquo;s 2025 enforcement cases also mentioned black-market groups buying and registering accounts in bulk, changing avatars to contain vague contact information for third-party platforms, and guiding users to download illegal porn-related apps. Other gangs used \u0026ldquo;offline prostitution\u0026rdquo; as bait to induce users to spend and tip; multiple people were eventually criminally detained.\nSo a lot of the time, it is not trying to make those few dozen yuan from your tip.\nIt is trying to test whether you are the kind of person who loses control easily.\nIf a porn link can pull you away, gambling, nude chat, and fake romance may pull you away again.\nThat is the \u0026ldquo;user profile\u0026rdquo; in the eyes of the black market.\n3 How terrifying are the profits? Do the risk math I will not write an operating ledger here. Only a risk ledger.\nBecause this is not a project tutorial. It is a breakdown of a criminal chain.\nStart with the change in industry scale.\nA video script gave one set of figures: from 2020 to 2025, the number of adult livestreaming platforms rose rapidly from 300 to 500 platforms to 3,000 to 5,000 platforms. The industry\u0026rsquo;s annual transaction volume rose from 10 billion yuan to 80 billion yuan.\nSeen alone, those numbers sound exaggerated.\nBut when placed next to public cases, they become easier to understand.\nJiaxing \u0026ldquo;Max\u0026rdquo; case: more than 3.5 million members, over 16,000 agents, and 250 million yuan in funds involved.\n\u0026ldquo;Tiger Live\u0026rdquo; case: the platform ran for less than two months, with cumulative top-ups of 7.287002 million yuan, more than 1.08 million registered members, and over 1,000 female porn streamers.\n\u0026ldquo;Tomato Community\u0026rdquo; case: platform and streamers split revenue 40/60, with more than 6 million yuan involved.\nOverseas-platform case reported by People\u0026rsquo;s Court Daily: total fines of 7.04 million yuan, 3.18 million yuan in criminal proceeds recovered, 4.16 million yuan in illegal gains surrendered, and 890,000 yuan in bank deposits frozen.\nCangnan cross-border porn livestreaming case: one platform reached more than 100 million yuan in top-ups in a year, had over one million registered members, recruited more than 200 female streamers, and had more than 50 streamers online on a typical day.\n\u0026ldquo;Huanggua Video\u0026rdquo; app case: more than 60 million domestic installs, over 300 million yuan in illegal profit, and a chain that included upload creators, guild leaders, fourth-party payment-channel gangs, and technical support.\nThe script also mentioned a smaller-grain account: building a small livestreaming platform may require only 5,000 to 10,000 yuan in input, but monthly turnover can reach 100,000 to 500,000 yuan.\nThe educational value of these numbers is not that they tell you how profitable this is.\nIt is that they tell you why it will not die.\nIf an illegal platform required millions in startup costs, one takedown would break its bones.\nBut if startup costs are low, shell-switching costs are low, agents can reorganize, and users can be redirected, it grows back like weeds.\nThat is not vitality.\nThat is the cost of crime being too low.\nBehind these numbers is a formula:\nPlatform turnover ≈ number of paying users × single top-up amount × repeat-purchase frequency × emotional-overheating coefficient\nThe real variables are not in \u0026ldquo;content quality.\u0026rdquo; They are in three places:\nFirst, can users be pulled in quickly?\nSecond, can users be kept inside paid scenarios?\nThird, can users be repeatedly stimulated into spending?\nThat is also why regulators keep focusing on \u0026ldquo;vulgar inducement to tip,\u0026rdquo; \u0026ldquo;fake personas used to trick users into tipping,\u0026rdquo; and \u0026ldquo;stimulating irrational tipping.\u0026rdquo; The Cyberspace Administration of China\u0026rsquo;s 2025 \u0026ldquo;Qinglang\u0026rdquo; special campaign against livestream tipping chaos explicitly named these problems, including vulgar group livestreaming, fake personas, inducing minors to tip, and PK mechanisms that stimulate irrational tipping.\nPorn livestreaming only pushes the same mechanism into a dirtier, more illegal, and more dangerous place.\nIn legitimate livestreaming, inducing tips is already a governance priority.\nIn illegal livestreaming, inducing tips is just the entry fee.\n4 Common tricks in the industry What porn livestreaming platforms exploit best is not technical loopholes.\nIt is human weakness.\n4.1 Free first, paid later They let users see a little first, then tell them the rest requires a top-up.\nThis is not a content strategy. It is a psychological noose.\nIn public cases, \u0026ldquo;Tomato Community\u0026rdquo; provided some free videos every day. To keep watching other content or livestream performances, users had to buy virtual coins and tip.\nFree is not kindness.\nFree lowers the psychological cost of the first click.\n4.2 Normal first, upgraded later Many livestream rooms do not start with explicit content. They first use ordinary chatting, flirtatious interaction, and lowbrow borderline content to keep people there.\nOnce people stay, they start investing time.\nAfter investing time, they are more likely to invest money.\n4.3 Persona first, harvest later The strongest part of livestream tipping has never been just the image.\nIt is the persona.\nYou think you are watching content. In reality, you are interacting with a designed object.\n\u0026ldquo;Pitiful,\u0026rdquo; \u0026ldquo;innocent,\u0026rdquo; \u0026ldquo;needs support,\u0026rdquo; \u0026ldquo;treats only you as special\u0026rdquo; — once these things are layered with sexual suggestion, consumption can easily be disguised as a relationship.\nThe Cyberspace Administration\u0026rsquo;s 2025 special campaign also listed \u0026ldquo;fake personas used to trick users into tipping\u0026rdquo; as a key target, including fabricating family misfortune, impersonating specific identities, and using AI to generate fake content.\nThe porn livestreaming black market uses this even more aggressively.\nBecause it never planned to build trust for the long term. It only planned to harvest in the short term.\n4.4 On-platform first, off-platform later Inside a platform, there is at least some review and some record.\nOnce you are led off-platform, the risk starts rising exponentially.\nNude-chat scams, gambling diversion, fake romance, extortion, and threats often happen off-platform.\nYou think you only added a contact.\nThe black market sees that you have stepped outside the platform\u0026rsquo;s protected zone.\n5 A field guide to spotting the black-market chain Many people have a misconception when looking at the porn livestreaming black market:\nThey think everything is fine as long as they do not click especially explicit links.\nWrong.\nA mature black market does not write the word \u0026ldquo;illegal\u0026rdquo; on its face from the start.\nIt disguises itself as part-time work, dating, side hustles, paid chatting, emotional counseling, livestream training, online work, or even a content account that looks completely normal.\nTo identify it, do not only look at what it says.\nLook at where it wants to take you.\n5.1 Recruitment: it will not say \u0026ldquo;come break the law.\u0026rdquo; It will say \u0026ldquo;easy money.\u0026rdquo; When recruiting streamers, the black market\u0026rsquo;s most common packaging is not pornography. It is \u0026ldquo;low barrier, high income.\u0026rdquo;\nIt hides the risk and shows only the returns.\nThere are several common shells.\nThe first is the part-time-job shell.\nThe pitch usually revolves around \u0026ldquo;work from home,\u0026rdquo; \u0026ldquo;flexible hours,\u0026rdquo; \u0026ldquo;no need to show your face,\u0026rdquo; \u0026ldquo;make money just by chatting,\u0026rdquo; \u0026ldquo;a nice voice is enough,\u0026rdquo; and \u0026ldquo;newcomers get guidance.\u0026rdquo;\nThe most dangerous part of this pitch is that it packages a high-risk act as an ordinary part-time job.\nMany young people do not start out aiming to break the law. They start with \u0026ldquo;I need money,\u0026rdquo; \u0026ldquo;I want to buy a phone,\u0026rdquo; \u0026ldquo;I want to repay Huabei,\u0026rdquo; or \u0026ldquo;I want some living expenses.\u0026rdquo;\nThat is where the real problem lies:\nThe more a part-time job emphasizes \u0026ldquo;easy, high-paying, no experience required,\u0026rdquo; the more you should ask: where exactly does the money come from?\nThe second is the livestream-training shell.\nIt will not tell you directly what to do. It first asks you to join a group, listen to a class, and look at screenshots of earnings.\nThe screenshots always show how many gifts someone received today, how much they withdrew yesterday, and how much turnover they made this month.\nIt is not selling a method.\nIt is selling the illusion of \u0026ldquo;you can do this too.\u0026rdquo;\nThe third is the acquaintance-referral shell.\nIn public cases, agents, guild leaders, and streamers often have recruitment relationships. Referrals by acquaintances lower people\u0026rsquo;s guard most easily, especially when the person is a classmate, roommate, senior schoolmate, or friend.\nThe more it comes through an acquaintance, the more clearly you need to calculate the incentives:\nWhy is she recruiting you?\nDoes she take a cut from you?\nWhy is the thing she wants you to do not something she dares put in a proper contract?\n5.2 Traffic diversion: its biggest fear is that you stay on a legitimate platform The core move in a porn livestreaming chain is always to take people away from legitimate platforms.\nSo when identifying diversion, do not only watch for sexual terms.\nA lot of diversion is not explicit at all.\nIt may be tiny text in an avatar, a homophone in a bio, a code word in the comments, a screenshot in a DM, or contact information flashing in the livestream background.\nIt may also start with a seemingly normal chat:\n\u0026ldquo;It\u0026rsquo;s not convenient to say here.\u0026rdquo;\n\u0026ldquo;The platform is strict.\u0026rdquo;\n\u0026ldquo;Go over there for the full version.\u0026rdquo;\n\u0026ldquo;Add me and I\u0026rsquo;ll send you benefits.\u0026rdquo;\n\u0026ldquo;Download this. That\u0026rsquo;s where the good stuff is.\u0026rdquo;\nWhat these lines have in common is not pornography.\nIt is that they make you leave the original platform.\nAs long as the other person\u0026rsquo;s core action is making you download an unfamiliar app, join an unfamiliar group, add an unfamiliar contact, scan an unfamiliar code, or click an unfamiliar short link, this is no longer ordinary content consumption.\nYou are being transferred.\n5.3 Payment: it will not make you feel like you are spending money. It will make you feel like you are interacting The smartest part of porn livestreaming platforms is that they rarely let you directly feel, \u0026ldquo;I spent this much money.\u0026rdquo;\nThey turn money into virtual coins, gifts, memberships, room cards, card codes, levels, and intimacy scores.\nThat step matters.\nRenminbi is real.\nVirtual coins are numbing.\nYou may hesitate before tipping 199 yuan, but swiping a gift, opening a room, or renewing a membership feels much lighter psychologically.\nIt also uses several things to push you to keep spending:\nFirst, time limits.\nThe stream is about to end, last show, only today\u0026rsquo;s viewers can see this.\nSecond, rankings.\nTop supporter, guardian, fan badge, intimacy value.\nThird, a sense of exclusivity.\n\u0026ldquo;Only for you,\u0026rdquo; \u0026ldquo;you\u0026rsquo;re different from everyone else,\u0026rdquo; \u0026ldquo;tip a little more and I\u0026rsquo;ll\u0026hellip;\u0026rdquo;\nFourth, sunk cost.\nYou already topped up, so not continuing would be a loss. You already reached this level, so stopping now wastes everything.\nThat is why many people start by wanting to spend only a few dozen yuan and then sink deeper and deeper.\nThey did not suddenly become stupid.\nThey were led step by step into a consumption environment designed to weaken rationality.\n5.4 Control: the most valuable thing is not your money. It is leverage over you The porn livestreaming black market rarely stops at one transaction.\nFor users, off-platform chats, nude chats, transfer records, download records, contact-list permissions, phone albums, and self-shot videos may all become material for later threats.\nFor streamers, recorded livestreams, identity documents, payment flows, chat records, and acquaintance relationships can also become control material.\nThat is why \u0026ldquo;I\u0026rsquo;ll just try it once\u0026rdquo; is especially dangerous.\nThe black market loves \u0026ldquo;just trying it once.\u0026rdquo;\nBecause once you try it, it has a chance to leave a record.\nOnce there is a record, the relationship changes.\nYou think you still have choices. The other side knows you are starting to feel afraid.\nMany extortion schemes do not win through technology. They win through psychology.\nThey may not really have all your information. They only need to make you believe they might.\nOnce you panic, you are more likely to keep paying.\n5.5 Laundering the image: it does not necessarily stay in the gutter forever More troublesome is that some black-market operators do not stay underground forever.\nAfter making their first pot of money, or after being hit by platform enforcement, they may move onto legitimate platforms and launder their image.\nThe methods are not complicated:\nThey do not mention past illegal traces. They only say they \u0026ldquo;understand human nature,\u0026rdquo; \u0026ldquo;understand traffic,\u0026rdquo; \u0026ldquo;understand livestreaming,\u0026rdquo; and \u0026ldquo;understand emotional value.\u0026rdquo;\nThey do not mention black-market experience. They package it as \u0026ldquo;operations methodology,\u0026rdquo; \u0026ldquo;private-domain conversion,\u0026rdquo; \u0026ldquo;livestream training,\u0026rdquo; and \u0026ldquo;emotional counseling.\u0026rdquo;\nThey do not mention old agent chains. They show luxury cars, transfers, travel, and an entrepreneur persona.\nThe most dangerous part of these accounts is that they package bad outcomes as success studies.\nThey do not tell you why others were sentenced, how many people were scammed, or how many young people were dragged down.\nThey only tell you:\nLook, I made money.\nThat is the most poisonous part.\nThey are not talking about making money.\nThey are deleting the cost of breaking the law from the story.\n5.6 Getting out: once something feels wrong, the most important thing is to stop handing things over If you are a viewer and find that you have already been diverted, charged, or threatened, the first thing to do is not keep explaining or keep begging the other side to delete records.\nThe first thing is to stop giving money, stop sending material, and stop proving yourself.\nSave chat records, transfer records, screenshots of links, and account information. Then report it to the platform and call the police if necessary.\nIf you are a recruited streamer and find that the content you are being asked to do has crossed the line, do not believe lines like \u0026ldquo;everyone does this,\u0026rdquo; \u0026ldquo;the platform is overseas so it is fine,\u0026rdquo; or \u0026ldquo;if you do not show your face, nobody will know.\u0026rdquo;\nThose words are not protecting you.\nThey are lowering your determination to leave.\nWhat actually protects you is stopping participation immediately, preserving evidence, and cutting contact with agents, guild leaders, and platform operators.\nIf you are a parent and find unfamiliar livestreaming apps, abnormal top-ups, late-night private chats, unfamiliar groups, or flirtatious language on your child\u0026rsquo;s phone, do not make scolding your first reaction.\nStop the money and accounts first.\nThen ask clearly about the chain: who introduced it? What was downloaded? How much was topped up? Were photos, videos, or identity information sent? Was there any threat?\nThe most useless move here is moral judgment.\nThe most useful move is cutting off the loss.\n6 What role does AI play here? One judgment in the video script is right:\nAI did not automatically make the world cleaner.\nIt only lowered the cost of disguise.\nIn this kind of porn livestreaming industry, AI mainly does three things.\n6.1 Lowering the \u0026ldquo;beauty cost\u0026rdquo; In the past, platforms needed real streamers\u0026rsquo; looks, bodies, and voices to attract users. Now beautification filters, face swapping, voice changing, and virtual avatars can all be used as disguise.\nThis means platforms do not necessarily need to find \u0026ldquo;high-attractiveness streamers.\u0026rdquo; They only need to find people willing to cooperate, then use technology to package the front end.\nThe video script put it bluntly: AI beautification can raise tip rates by three to five times and double average order value.\nNo matter how much those numbers fluctuate across platforms, the direction is clear:\nAI here is not improving aesthetics.\nAI is increasing the impulse to pay.\nThis creates a dangerous illusion:\nUsers think they are looking at a person.\nThey are actually looking at a packaged product.\n6.2 Lowering the feeling of identity exposure For streamers, face swapping, beautification, and virtual avatars lower the psychological barrier.\n\u0026ldquo;Anyway, that is not what I really look like.\u0026rdquo;\nThat sentence sounds like protection. In reality, it may be a trap.\nBecause the platform has recordings, agents have chat logs, payments leave transaction flows, and users may screen-record. You think AI has separated you from your real identity. In the end, it may only help you step into risk faster.\n6.3 Making scams more realistic AI face swapping, voice cloning, and virtual personas make fake interaction feel more real.\nThe Cyberspace Administration\u0026rsquo;s 2025 \u0026ldquo;Qinglang\u0026rdquo; special campaign against the misuse of AI technology explicitly listed AI face and voice cloning that infringes public rights, missing AI-content labels that mislead the public, and other issues as key targets. In the first phase, regulators handled more than 3,500 noncompliant AI products and removed over 960,000 pieces of illegal or noncompliant information.\nThis shows regulators have already seen the trend.\nOnce AI combines with pornography, scams, and tipping, the most dangerous question is not \u0026ldquo;does it look real enough?\u0026rdquo;\nIt is \u0026ldquo;does the user still know what they are interacting with?\u0026rdquo;\nIf they do not know, they lose control.\n7 Why is it so hard to wipe out completely? Many people ask:\nIf this is so illegal, why does it keep coming back?\nThe answer is not that regulators are ignoring it.\nThe answer is that this business has several classic black-market traits.\n7.1 Asset-light It does not need storefronts, stable locations, or a large offline organization.\nOne app, a set of payment entrances, a group of agents, and a pile of accounts are enough to start moving.\nThat is also why it often changes shells.\n\u0026ldquo;Sleeping Beauty,\u0026rdquo; \u0026ldquo;Ambiguous,\u0026rdquo; \u0026ldquo;Meiniang,\u0026rdquo; and \u0026ldquo;Aimeiren\u0026rdquo; all appeared in public cases as platform aliases repeatedly changed by the same operator.\n7.2 Distributed The platform operator, agents, streamers, payment providers, and users do not have to be in the same place.\nIn the Jiaxing \u0026ldquo;Max\u0026rdquo; case, overseas organizers, domestic agents, social-platform promotion, and gambling diversion were mixed together. Police went to Cambodia twice to arrest organizers, while also arresting large numbers of agents inside China.\nThis is not a single point.\nIt is a web.\n7.3 Stable demand Pornography, companionship, stimulation, curiosity, loneliness — these demands do not disappear because one platform is taken down.\nThe platform dies. The demand remains.\nIf the demand remains, someone will keep changing shells.\nWhat the black market fears most is not one app being banned.\nIt fears users no longer taking the bait.\n7.4 Entangled boundaries with legitimate platforms Many black-market operators no longer put illegal content on legitimate platforms from the start.\nThey first use borderline content, attractive avatars, vague bios, and private-message scripts to pull people away from legitimate platforms.\nIn Douyin\u0026rsquo;s 2025 published data, 88,000 person-times were penalized for vulgar inducement to tip, and 117,000 porn-related violating accounts had livestreaming privileges permanently revoked. Leads involving porn-app diversion and offline-prostitution tipping inducement were also transferred to police, with criminal detentions following.\nThis shows legitimate platforms are being used as traffic entrances for the black market.\nNot every borderline account leads to porn livestreaming.\nBut many porn livestreaming chains start with borderline content.\n7.5 Short life cycles, which make short-term harvesting even more aggressive The video script also made one judgment: small illegal adult livestreaming platforms may last only 45 days on average, while mid-sized and large platforms often have survival windows of only several months.\nIf that number holds, it explains many behaviors that look abnormal.\nWhy does it not care about user experience?\nBecause it never planned to serve you long term.\nWhy does its script become more aggressive?\nBecause it wants to push turnover before getting shut down.\nWhy does it divert users off-platform into gambling and nude-chat scams without hesitation?\nBecause it knew from day one that this was a disposable business.\nLegitimate platforms talk about retention.\nBlack-market platforms talk about squeezing people dry.\n8 Risks worse than losing money For ordinary people, the biggest loss from porn livestreaming is not necessarily money.\nIt is not knowing what you have already handed over.\n8.1 Viewer risks The first layer is money.\nTop-ups, tips, memberships, private rooms, and off-platform payments are all money.\nThe second layer is privacy.\nWhat apps you downloaded, what accounts you registered, whom you transferred money to, and what you chatted about can all become material for threats.\nThe third layer is fraud.\nNude chat, gambling, fake orders, fake romance, offline meetups — any step can turn from \u0026ldquo;consumption\u0026rdquo; into \u0026ldquo;extortion.\u0026rdquo;\nThe fourth layer is legal and family risk.\nOnce it involves distribution, organization, recruitment, agency work, or helping promote, it is no longer \u0026ldquo;I was just watching.\u0026rdquo;\n8.2 Streamer risks Many people think streamers are just \u0026ldquo;making quick money.\u0026rdquo;\nPublic cases tell you otherwise.\nIn cases reported by People\u0026rsquo;s Court Daily, multiple streamers were sentenced to prison terms ranging from four years to seven years and two months, and were fined.\nThis is not moral risk.\nIt is criminal risk.\nMore realistically, once someone enters this chain, they are often controlled from multiple sides: platforms, agents, acquaintances, and users.\nScreen recordings, chat logs, transfer records, and identity information can all become ropes used to control you later.\nYou think you are making money.\nMany times, you are giving other people leverage over you.\n8.3 Risks to minors Minors are the people who least belong in this chain and the easiest to harm.\nIn 2022, four departments issued the Opinions on Regulating Online Livestream Tipping and Strengthening the Protection of Minors. It clearly required banning cash top-ups, gift purchases, online payments, and other tipping services for minors. It also required canceling tipping leaderboards and forbidding platforms from using tipping amounts as the sole basis for streamer rankings, traffic allocation, or recommendations.\nWhy so strict?\nBecause tipping mechanisms amplify impulse by design.\nAdults may not withstand them. Minors are even less able to.\nAdd pornography, companionship, and fake relationships on top, and this is no longer an entertainment problem. It is a harm problem.\n9 A survival guide for ordinary people 9.1 Anything that guides you away from a legitimate platform should be treated as risky by default Avatars, bios, comment sections, private messages, livestream backgrounds, homophonic code words — if the core action is pushing you to another platform, making you download an unfamiliar app, or getting you to add an unfamiliar contact, stop first.\nThe real danger is not what you clicked.\nIt is that you started walking the path someone else laid out for you.\n9.2 Anything that makes you top up first before unlocking access should be avoided Especially unfamiliar apps, websites, private rooms, membership cards, and virtual coins.\nThese platforms are best at turning real money into virtual chips, making you feel less loss.\nYou are not consuming content.\nYou are sending yourself into the other side\u0026rsquo;s harvesting system.\n9.3 Anything involving nude chat, recordings, or private photos should be cut off immediately Do not explain. Do not bargain. Do not keep proving who you are.\nSave evidence, stop communicating, and report it to the police or the platform.\nA lot of extortion works because you panic.\nThe more you want to settle it privately, the more the other side knows you are afraid.\n9.4 Parents should not only watch \u0026ldquo;what the child saw.\u0026rdquo; Watch where the money went When minors are harmed, it often starts not with content, but with payment.\nParents should look beyond \u0026ldquo;did my child see borderline content?\u0026rdquo;\nMore important: are there unfamiliar top-up records, abnormal transfers, sudden downloads of unfamiliar apps, or frequent late-night chats?\nThe path of money is often more honest than browsing history.\n9.5 Do not treat \u0026ldquo;everyone is watching\u0026rdquo; as proof of safety The black market loves manufacturing atmosphere.\nOnline viewer counts, bullet comments, gift effects, leaderboards, top-up rebates, and streamers thanking users are all designed to make the whole thing feel normal.\nThe more it looks like a lively venue, the more you should ask:\nWhy is it trying to make me pay so quickly?\nConclusion The real money in porn livestreaming does not come from porn content itself.\nContent is only the entrance.\nThe real money is in the chain behind the entrance:\nTraffic, top-ups, tips, revenue shares, agents, off-platform diversion, and scam filtering.\nThat is why it is so hard to kill with one punch.\nTake down one platform, and it changes shells.\nBan one account, and it switches accounts.\nClean up one batch of livestream rooms, and it moves into private domains.\nThe harshest part of this business is that it makes people think they are only watching from the sidelines.\nIn reality, from the first click, the sickle is already starting to move.\n","permalink":"https://ryrenz.com/industry/porn-live-stream-industry-chain/","summary":"Porn livestreaming is not just a sleazy transaction between one streamer and one tipping viewer. It is a gray-and-black-market pipeline involving platforms, agents, streamers, payment providers, technology vendors, and traffic brokers. The real danger is how it turns impulse spending, virtual coins, memberships, and private-domain harvesting into repeatable cash flow.","title":"Inside the Multimillion-Dollar Supply Chain Behind Porn Livestreams"},{"content":"A lot of companies are pushing AI transformation right now. Step one is usually the same: buy a bunch of coding agent licenses, then roll out training and hope employee productivity goes up.\nI think that math is fake.\nIf the org chart stays the same, trying to boost everyone’s output with training and AI tools is not realistic.\nThe AI Companies That Win Are Small and Sharp Midjourney started with 11 people and hit $200 million ARR within a year. No funding. No paid acquisition. No sales team. That’s $18 million in revenue per employee.\nCursor had around 50 people early on. By early 2026, it had already reached $2 billion ARR with a $29.3 billion valuation. Even if you use its current headcount of 150, that’s still $6.7 million per employee. A normal SaaS company does maybe $150,000 to $250,000 per employee. Cursor is 30 times above the industry average.\nBut Anthropic itself is even crazier, especially the Claude Code team.\nClaude Code writes 90% of its own code.\nBoris Cherny, the founder of Claude Code, said on Lenny’s Podcast in February 2026 that he had not handwritten a single line of code since November 2025. Claude Code wrote 100% of it. He ships 10 to 30 PRs a day.\nAcross Anthropic, average engineer output is up 200% over the past year. Code written by Claude Code already accounts for 4% of all public commits on GitHub. Boris said it plainly on the podcast: coding is basically solved.\nThe numbers from other companies using Claude Code are even more extreme. At Stripe, one team migrated 10,000 lines of Scala to Java in four days. The original estimate was 10 engineer-weeks. At Wiz, migrating a 50,000-line Python library to Go took 20 hours. The original estimate was two to three months.\nThen there’s Pieter Levels. One person. More than 70 projects. Even if 95% fail, the handful that work still make $3 million a year in profit. His exact line was: “Every extra employee slows the company down.”\nThen there’s OpenClaw: 370,000 lines of TypeScript in a single repo, with a README that literally says “AI/vibe-coded PRs welcome.” Its founder, Peter Steinberger, runs more than a dozen Codex agents at the same time. One person, one day, small-team-week level output.\nAndrej Karpathy said in March 2026 on the No Priors podcast that he hasn’t handwritten a single line of code since last December. He spends 16 hours a day talking to agents. His term for it was “AI psychosis.”\nThe ceiling on individual output was never determined by team size. It’s determined by leverage. AI is the new leverage. Karpathy has said that cloud plus AI makes the Pieter Levels model—one person running multiple companies—actually viable, and that a billion-dollar one-person company is no longer crazy.\nSam Altman has apparently started privately betting with CEOs on what year the first one-person billion-dollar company will show up.\nOrg Structure Decides Who Wins Over the last two years, I’ve become convinced of one thing: company-wide AI upskilling will probably fail.\nConway’s Law says organizations design products that mirror their structure. In the AI era, that means traditional organizations will only ever build AI features, not AI-native products.\nIn a traditional company, the unit of collaboration is human to human.\nEven if individual employees get faster, 80% of their time still goes into communication, process, meetings, decisions, and cross-functional coordination. Overall productivity barely moves.\nYou can’t expect a Formula 1 driver to floor it in an alley.\nIn an AI-native organization, the unit of collaboration is “human + agent.” One engineer can run five to ten agents at once. What used to take a small team a week can get done in a morning.\nGo one layer deeper and this is really anti-collaboration: deliberately reducing human-to-human links and increasing human-to-AI links.\nThat’s what Cursor did. No product managers. Engineers spin up agents, write code, and decide product direction themselves. 150 people carrying $2 billion ARR isn’t because they’re smarter. It’s because the collaboration model is fundamentally different.\nThe Claude Code team at Anthropic works the same way. One of their engineers has a catchphrase: “My job now is figuring out how to get as many Claude Code instances working in parallel as possible.”\nThis is hard inside a traditional company. Not because of the tech. Because of the org chart.\nGive an engineer in a traditional department 10 agents and their first reaction is not excitement. It’s this:\n→ If my output goes up 10x, are they cutting headcount next year?\n→ Can I still protect my team? If not, who gets cut first?\n→ If an agent takes over my whole domain, what exactly do I write in my promotion packet?\n→ If I do the work, the credit goes to AI. If prod goes down, the blame is mine. Why would I want this?\nThat’s why the people really using agents to do the work of 10 people are indie hackers, small-team founders, and people whose output maps directly to their own income.\nWhat It Feels Like Firsthand During the day I build frontend and backend systems inside a company. Nights and weekends I work on open source and side projects with a full AI workflow stacked on top.\nThe pace on those two sides doesn’t even feel like the same era.\nAt work, polishing a single API means going through the whole process. From kickoff to production, about a week. At night, I can build an MVP in one evening and put it in front of users the next day.\nThe tools aren’t different. My company also pays for Claude Code.\nThe difference is whether the organization can actually let the tool unlock productivity. The answer is no.\nSo in most companies, “AI transformation” is just a PowerPoint project. Buy tools, run training, send announcements, set OKRs—and keep the same org while every workflow is still blocked by the same process overhead.\nAn AI-native individual can now do in a few days what used to take a traditional team a few months. Claude, Cursor, and Rakuten have already proven that at the company level. The problem is that most organizations simply do not allow individuals to operate in an AI-native way. The pace gets choked by process almost immediately.\nThe team is no longer a flywheel. It’s a liability.\nWorse, traditional teams are not just slow. They permanently lock you out of becoming AI-native.\nThe org you built, the contracts you signed, the layers you set up, the middle managers you hired—every single one of them is on the side of not changing. You’re not going to truly reorganize. At most, you’ll wedge in an “AI innovation team” and then watch it die in KPI warfare.\nThat is the real script behind what 90% of companies call “AI transformation.”\nThe Breakout Move: Cut First, Then Hire On Lenny’s Podcast, Boris Cherny mentioned a management principle that’s actually very interesting: “underfund things a little bit.”\nHis exact line was: “There’s this interesting thing when you underfund everything a little bit, because then people are kind of forced to Claude-ify.”\nIf teams are deliberately understaffed, under-budgeted, and under time pressure, they’re forced to use AI to close the gap.\nThe real problem in mature companies is not “too much work, too few people.” It’s “too many people, not enough real work.”\nWhen there are too many people and not enough work, people are not working. They’re coordinating who should work. Meetings, alignment, process, status updates—and the week is gone.\nAll of that was designed for the productivity limits and friction costs of the previous era. If you just jam AI into it, AI can only help around the edges. It never reaches the main flow.\nCut half the team, and the logic flips.\nWhen there are fewer people and more work, nobody has time for meetings whose only purpose is to justify meetings. Everyone is forced to face problems directly, forced to orchestrate agents, forced to own work end to end. People aren’t getting cut because they’re bad. They’re getting cut because that team size was designed for the last era.\nIn the new era, that team size itself is the biggest obstacle to lower costs and higher output.\nIt’s brutal, but it works. In 2026, more and more companies will go this route—not because they want to, but because they don’t have a cheaper option.\nThe AI Upside Is Not Fair AI only makes coders code faster, product people validate faster, and people with taste produce better work. It amplifies what you already have. It does not give you what you don’t.\nAnd the people at the top are operating at a level most people can’t even picture. Top indie developers are writing 10,000 to 100,000 lines of code a day.\nKarpathy spends 16 hours a day coding through agents and gets anxious if he doesn’t burn through enough tokens by the end of the month.\nBoris Cherny ships 10 to 30 PRs a day. Peter Steinberger runs more than a dozen Codex agents at once.\nY Combinator’s Garry Tan said he’s been sleeping only four hours a night because working with Claude Code is so stimulating he doesn’t need anything to keep himself awake. He also said a third of the CEOs he knows are in the same state.\nIf the top people are already twisted into knots like this, what are ordinary employees supposed to do?\nFirst: drop the employee mindset.\n“Wait for instructions, then execute” is already something AI does better than people.\nIf you can’t break down problems, design solutions, and judge what’s worth doing, then when layoffs come, there’s a good chance your name is on the list. Pure executors are cheap in the AI era.\nSecond: think full-stack.\nEven if you “just write code,” you need to be able to carry work from requirement to deployment on your own—product design, technical approach, launch, operations, all of it. No passing the buck.\nIn orgs like Cursor or Claude Code that don’t even have product managers, that’s already how engineers work. Being able to run the entire product loop alone is one of the rarest skills in the new era—and one of the hardest to replace.\nThird: build business sense.\nReal leverage is using AI to orchestrate resources—turning ideas into products, then getting those products sold.\nDon’t be satisfied with using AI at work to build a few product features. Learn how your company thinks about product, sales, and execution.\nThat way, if layoffs do hit, and you can both build and sell, you still have a shot at carving out your own path through entrepreneurship.\nFinal Replacing the basic unit of collaboration across the organization—from human-to-human to human-to-agent—is the unavoidable path of enterprise AI transformation.\nIf a company can’t do that, buying more Claude Code licenses won’t matter.\nOrganizational inertia, contracts, KPIs, politics, relationship debt—every one of them pushes against the switch, and most companies simply won’t have the nerve to do it.\nSo more and more of them will choose the blunt option: cut headcount first, fully equip the people who remain with AI, then hire back only when needed. It may not be the optimal path, but it’s general-purpose and it works.\nby Ren\nThanks for reading.\nIf you’re interested in careers, startups, investing, or AI, follow along.\n","permalink":"https://ryrenz.com/career/enterprise-ai-transformation-layoffs/","summary":"Buying AI tools and rolling out company-wide training does not equal AI transformation. Real change means replacing human-to-human collaboration with human-plus-agent workflows. That shift hits headcount first, then rewrites the survival rules for employees.","title":"Why Enterprise AI Transformation Starts With Layoffs"},{"content":"If you know nothing about startups, this guide will show you what real startup problems look like.\nIf you have a technical background, you will find that 70% of this has nothing to do with code. That is exactly why so many technical founders crash.\n01 | Product Is Only the First Step Almost no company on earth is truly necessary.\nThe engine of human innovation has not been “only I can do this” for a long time. It is: I want to do this, and I hope other people will pay me for it.\nAdmit that, and you will be less likely to fool yourself with your own “sense of mission.”\nYou are not saving the world. You are trying to persuade a group of people who never asked for you to exist to hand you their money.\nEvery company is basically carrying three things at the same time:\nLabor — the actual, repetitive, unsexy work. Writing code, debugging, running evals, cleaning dirty data, writing docs, doing ops, clearing tickets, replying to customer emails. Administration — dealing with people. Teammates, cofounders, investors, compliance, hiring, customer success, the board. Deception (the philosophical essence of marketing) — convincing someone who never asked for you to give you money. Fundraising pitches, Product Hunt, Twitter marketing, PR, recruiting scripts. Technical people only want to do the first one 99% of the time.\nThat is why their companies often do not die because of the product. They die because of the other two.\nBefore you place your bet, have an honest conversation with yourself: among these three things, which one are you most willing to do? Which one do you hate most?\nThe real question is not “Can I build the product?”\nIt is: do I actually believe in its value, and am I willing to sell it?\n02 | Motivation Matters More Than Method — Do Not Start Up Just to Look Cool Every generation of young people has its own dream of becoming famous. In tech, this generation’s version is: build an AI Agent, launch an MCP server, make a developer tool, build an LLM-native SaaS, create an AI companion, or build a Devin competitor.\nThe problem is not that these directions are wrong. The problem is that the most dangerous engineering culture of this generation is this: you do not need to build something truly usable to look like you are building a startup. You only need to post a demo video, beat a benchmark, and gather some Twitter traffic, and people will call you an “AI Founder.”\nThe industry is messy right now. That is why products that actually land are getting harder to build: you are competing for the same users and the same investors against people who treat “looking busy” as “doing the work.”\nYou do not need to find the method immediately, but you must first find a thesis worth betting on: a contrarian judgment you are willing to bet is right.\nA good thesis looks roughly like this:\n“All AI coding tools today optimize for the solo experience, but 90% of production code is written through team collaboration. No one is seriously working on that gap.” “Prompt engineering is not a long-term skill. Context engineering is. The people building tooling around context will win.” “Agent frameworks are all chasing multi-agent orchestration, but 90% of the real value comes from one agent doing one thing extremely well. The market will flip.” “The Chinese AI developer community is badly underestimated because big model companies are all chasing English ARR. There is a window here.” If you cannot state your thesis, you are not building a business. You are performing. Your thesis should be sharp enough that a smart person can publicly disagree with you. If no one can disagree with it, you are just describing obvious nonsense everyone already accepts.\n03 | Innovation Means Digging Into Needs Existing Products Do Not Meet Valuable innovation is not about doing the opposite. It is about seeing which need the current solution still fails to solve.\nSoftware products usually run into a few familiar constraints:\nToo slow: users cannot wait. Too expensive: the unit economics do not work. Too hard to use: only experts can use it. Too fragmented: the workflow breaks across different tools. Unreliable: results are not reproducible, and when something goes wrong, no one dares to take responsibility. Too dependent on manual work: once scale grows, the only answer is adding more people. Your innovation should ideally land in one of these places.\nNot: “Everyone else is building SaaS, so I will build a command-line tool.”\nBut: the target user works in the terminal every day. Forcing them to open a web page interrupts the workflow, so a CLI is the lower-friction entry point.\nNot: “Everyone else is building multi-agent, so I will build single-agent.”\nBut: what users actually need is one task finishing reliably, not watching five agents chat with each other on the screen.\nNot: “Everyone else is chasing public traffic, so I will build a private community.”\nBut: your product needs high trust and dense feedback. Fifty deep users are closer to real demand than 5,000 spectators.\nBeing contrarian has no value by itself.\nSolving a real need does.\n04 | Clarify the Vision and Mission First Many software companies rush to write a slogan at the start.\n“AI-powered X for Y.”\n“The future of Z, today.”\n“Empowering teams to\u0026hellip;”\nThe biggest problem with these lines is not that they are tacky. It is that they are useless. They do not tell users what kind of world you are trying to create, and they do not tell the team how to make tradeoffs every day.\nWhat a software company needs to clarify first is not whether the slogan sounds nice, but two things:\nVision: what you believe the future will become. Mission: the concrete way you plan to push it there. The team needs to know where the future is headed. Investors need to know which market you are aiming at. Users need to know what value you can actually provide.\nFor an early-stage software company, Vision and Mission do not need to sound grand, but they must be grounded.\nThat is the starting point of a startup with real potential.\n05 | The First Product Is a Proof of Concept, Not the Ideal Product Most people start by putting all their chips on “the perfect version in my head.” That is wrong.\nThe purpose of the first product is not to make money. It is to test whether your shaky core assumption holds up at all.\nProof of concept does not mean artificially making things harder for yourself.\nThat is not how software is validated. You do not need to touch the most complex enterprise workflow on day one, and you do not need to find the hardest people to please just to prove yourself.\nA more realistic approach is to compress your core assumption into something the market can see.\nFor example, if you want to build an AI image tool, the first step may not be writing the full product. You can use existing image generation models to create the target effect, organize the outputs into a set of images, and distribute them on platforms like Xiaohongshu, Zhihu, Twitter, Hacker News, and Product Hunt.\nWhat you need to watch is:\nDoes anyone save it, share it, or ask, “How do I use this?” Is anyone willing to leave an email, join a waitlist, or enter a group? Does anyone ask about pricing, APIs, or whether it can integrate with their workflow? Is anyone willing to pay for or try a still-rough version? That is proof of concept.\nIt is not about proving you can build the full product. It is about proving the idea gets a reaction in the real attention market.\nIf a set of images, a demo clip, a landing page, or a handwritten email cannot trigger any feedback, do not rush to write code. What you need to validate is not your engineering ability. It is whether demand exists.\nThe first product should be a minimal test instrument: put your core assumption into the real world at the lowest possible cost and see whether it sparks any reaction.\n06 | The Four-Legged Chair A three-legged chair can be the highest-quality chair in the world, but it is still a three-legged chair. No one wants to sit on it.\nDo not build a “seemingly perfect business.” Add another leg so that when one breaks, you do not fall over.\nThis applies not only to companies, but to every product.\nThe first product is a proof of concept. The second and third products are too: you need to keep checking whether you are living on only one assumption, one channel, one type of user, or one cost structure.\nTech and AI founders are especially prone to building three-legged chairs. The common disease looks like this:\nSingle-model dependency: the entire product is built on one OpenAI or Anthropic model. They change pricing, policy, or rate limits, and your whole product dies. Single channel: all traffic is bet on one Product Hunt / Twitter / Hacker News launch. If it fails, you go to zero. Single customer type: you serve only individual developers or only large enterprises, with no room to maneuver in between. Single revenue source: only subscriptions, only API usage, or only one-time purchases. Single identity: you are only a “founder.” You do not have a second identity to lean on, like “content creator,” “open-source maintainer,” or “consultant.” Adding another leg is not a distraction. It is how you stay alive until next year when one leg breaks.\n07 | Three Ways to Make a Product Attractive For any product to be “attractive,” there are basically only three directions. Pick one and stick with it. The worst move is wanting all three.\n1. Generous and Comfortable Make people feel like they are getting a good deal, being taken care of, and not being nickel-and-dimed. Typical plays:\nOffer a very generous free tier (the early Vercel, Cloudflare, and Supabase route). Make docs clear, error messages friendly, and community response fast. Open-source your core, then monetize services and cloud hosting. Good for products that build trust over the long term. The way it dies is being crushed by the cost structure of free users.\n2. Exaggerated and Tempting Visual impact, benchmark domination, explosive launches, extreme demo videos. Typical plays:\nA launch video people cannot help sharing. A benchmark number that is clearly ahead. A controversial founder persona the media loves to cover. Good for products that need short-term awareness. The way it dies is when the gap between the demo and the actual capability is too large. Once trust collapses, there is no second chance.\n3. Hard-to-Explain Pull You cannot quite say why, but you want to use it. The typical examples are command-line tools, minimalist products, and old-school tools that do not offer a “modern UI”: Vim, tmux, ripgrep, rsync, that kind of thing. They do not explain themselves. They assume you are willing to learn.\nGood for serving deep users and building a cultural moat. The way it dies is getting eaten by a slightly friendlier competitor.\n08 | The Logo Does Not Matter. The Ability to Give It Meaning Does Technical founders spend too much time agonizing over logos, fonts, brand colors, and domain suffixes.\nThe truth is very direct:\nWhen you are starting up, if you cannot genuinely identify with or oppose your logo, then what it looks like does not matter at all. As long as you like it, it is fine. No one else cares.\nApple’s logo is an apple with a bite taken out of it. Over time, they made it more boring: flatter, plainer, more “nothing.” It became more valuable anyway.\nThe value is not in the graphic. It is in your ability to keep giving the graphic meaning.\nFor an LLM company, if the product is good, seeing the logo will make you think, “That company gets things done reliably.” If the product is bad, even the most beautiful logo is empty.\nDo not spend more than four hours on the logo in year one.\n09 | Good Friends Are Not Cofounders Tech people especially love “starting a company with good friends.” This is one of the most common and most expensive forms of self-deception in startups.\nThe internal loop of friend-based startups looks like this:\nMutual affection → no one wants to be the wet blanket → no one questions PMF → no one challenges the pricing strategy → no one asks “Where is the sales pipeline?” → resources, time, and energy get wasted → the company burns through its budget in perfect harmony → when everyone breaks up, they all feel the other person changed.\nWhat you need is not a friend. You need:\nSomeone who can disagree with you constructively. Someone who can push back efficiently and with depth. Someone who can respectfully challenge every view you hold and is smart enough to push the argument to a conclusion. In a tech company, this often means you need a cofounder who makes you uncomfortable: a business cofounder who asks, “Who will pay for this feature?” a product cofounder who tells you, “Engineers are not the users,” or a design cofounder who reminds you, “This UI will make users feel bad.”\n10 | The Truth About Communication — Someone Has to Decide Engineering culture easily falls into two extremes: the “democratic decision-making trap” and the “RFC rat race trap.” Every decision gets a meeting, every meeting gives everyone a voice, every voice is respected, and three weeks later nothing has been decided.\nMany high-IQ people, and here “high-IQ” is a neutral description, are smart enough to always find problems, but not smart enough to solve them.\nWhen people enter a meeting with that mindset, they only want to be heard. They do not care what they are saying, and they do not care whether the discussion is moving a decision forward.\nThe mature way to handle this is not fancy:\nState openly in advance: “In this company, I (or a specific person) am the idiot who has to put a hand on the table and make the call. You can all have opinions, but opinions do not change the final decision.” Let everyone speak once first. You cannot skip this step, or people will feel unheard. Once the decision-maker speaks, everyone should assume he has listened, processed it, and still holds the final decision rights. The decision-maker’s pain point and decision rights must be tied together. When the company loses money, he still has to pay the team out of his own pocket. Without that, decision rights are fake. Communication is not democracy. Communication is making sure everyone knows when there is no more room for discussion.\n11 | Positive Cash Flow Does Not Equal Profit This is one of the easiest things for beginners to confuse, and it is especially deadly in the AI / SaaS era.\nPositive cash flow: you have 100 yuan in the account on January 1, 200 yuan on February 1, and a net inflow of 100 yuan during that period. Profit: after taking that positive cash flow on paper and subtracting debt still being repaid, assets that should be depreciated, future commitments you must pay, and marginal costs that have not yet settled, the result is still positive. The most hidden cash-flow trap in AI startups is token cost:\nYou think you made money, but you did not account for the marginal cost of model APIs. This is especially true for agent products, where one task may call the LLM hundreds of times. As user activity rises, your costs grow linearly or even superlinearly. You think prompt caching will save you, but cache hit rates are never as optimistic as the demo. Other hidden costs you still need to subtract:\nGPU / server depreciation. Vesting equity for the team. This is a real cost, not free money. Retention bonuses, compliance spending, and hosting fees you owe over the next year. The backstop cost of SLAs promised to customers. If you do not know exactly which state you are in, you are probably not truly profitable.\nSustainable does not mean profitable either. A cash-flow-positive nonprofit can be “sustainable” for many years, but that may not be the state you want.\n12 | Debt Is a Tool, Not a Shame Many technical people are taught from childhood: do not take on debt, do not owe people money, and that makes you a good person.\nThat can make you a good person. It can also make you a bad businessperson.\nTruly large organizations essentially live inside debt, commitments, and future delivery. When other people need you to fulfill promises in the future, you have truly entered the web of incentives.\nDebt takes several forms in tech startups, and you need to learn how to use all of them:\nFinancial debt: VC funding is essentially a form of delayed-payment debt, a promise of returns. Debt financing, convertible notes, ARR financing, revenue share, and so on all belong here. Commitment debt: roadmaps promised to customers, milestones promised to investors, vesting promised to the team, maintenance commitments promised to the open-source community. Technical debt: this one is discussed too much, to the point that many people fear it excessively. The right amount of technical debt is the price of startup speed, not a moral stain. Do not be the “zero-debt” good person. Learn to use respect, calm, and structure to make other people economically tied to your survival. If you owe someone for 10 years, they want you to last 10 years.\n13 | Timing Is Part of the Product Every industry has cycles. The cycles in AI and tech can be more hidden, but more deadly:\nMajor model version cycles: every three to six months, a new model comes out, and your entire set of prompts and evals may need to be rewritten. Fundraising cycles: VC mood, interest rates, and macro narratives all affect your next-round valuation, and none of that has anything to do with your product. Hiring cycles: in certain periods, such as big-tech layoff waves or graduation season, the quality of talent you can hire varies wildly. Content high season vs. low season: around holidays, major model releases, and industry conferences, traffic distribution is completely different. User willingness-to-pay cycles: B2B follows budget quarters; C2C follows consumer spending seasons. You need to design the timeline as part of the product.\nIn peak seasons, do things that amplify momentum: launch, fundraise, hire, push content.\nIn slow seasons, do compounding work: write docs, pay down technical debt, build internal tools, organize SOPs.\nGo one step further and store peak-season output for slow-season consumption: content written ahead of time, tickets completed ahead of time, demos recorded in advance, processes documented in advance.\nSummary Take business seriously;\nfind a demand worth betting on;\nclarify the Vision and Mission;\nvalidate the core assumption at the lowest possible cost;\nmake the product and company walk on multiple legs;\nchoose one type of product appeal;\nkeep giving symbols meaning;\nstay away from friend-based partnerships;\nlearn to be the decision-maker;\ntell cash flow from profit;\ntreat debt as a tool;\ntreat the timeline as part of the product.\nThe rest is daily labor, administration, and marketing, repeated every day.\nThanks for reading this far. This guide is not a path you can copy directly. No one learns startups by reading articles, and the tuition you owe will not be discounted by a single cent. But a few pits others have fallen into may help you fool yourself a little less and find a little more direction when you run into similar problems.\n","permalink":"https://ryrenz.com/startup/zero-startup-guide/","summary":"If you have a technical background, you will find that 70% of real startup problems have nothing to do with code. The product is only step one. After that come sales, communication, cash flow, debt, cofounders, and timing.","title":"If You Know Nothing About Startups: A Startup Guide for Technical People"},{"content":"When most people write their first Skill, they instinctively write it as a longer prompt.\nBackground, rules, caveats, examples, references — all stuffed into one SKILL.md. It looks complete. It usually isn\u0026rsquo;t.\nThe real value of a Skill is this:\nWhen the user describes a certain kind of task, the agent recognizes the situation, loads the right workflow, picks the right tools, and finishes the job by a fixed method. And the next time a similar task comes along, the agent does it the same way, reliably.\nA prompt is a one-off instruction.\nA Skill is a reusable, self-triggering, maintainable, evolvable workflow.\n1. What a Skill Is 1.1 How does a Skill relate to MCP? Before getting into Skills, it helps to understand MCP.\nMCP stands for Model Context Protocol. It\u0026rsquo;s a standard Anthropic designed to let Claude plug into external tools and data (Notion, Asana, your internal systems — all of that runs through MCP).\nSkills and MCP are partners.\nThe Anthropic docs use a kitchen metaphor.\nMCP gives you a \u0026ldquo;professional kitchen\u0026rdquo; — the tools, data, and call interfaces that connect to a service (an MCP server for Notion, Feishu, TickTick, WeChat Official Accounts, or Xiaohongshu, for example).\nA Skill gives you the \u0026ldquo;recipe\u0026rdquo; — it tells Claude how to use those tools, step by step, to produce something useful.\nMCP without a Skill: the user connects your MCP server but has no idea what to do next. Every conversation starts from scratch. Results are inconsistent.\nA Skill without MCP: a Skill can also run on its own — generating documents, making images, organizing material. Anthropic\u0026rsquo;s own docx / pptx / xlsx skills are this kind.\nBoth together: the user can complete a real task right out of the gate. MCP decides what Claude can do; the Skill decides how Claude should do it.\n1.2 Pin down the use case before you start Before writing anything, answer four questions:\nWhat is the user trying to accomplish? What multi-step workflow does that involve? Which tools are needed (Claude\u0026rsquo;s built-ins, or something from MCP)? What domain knowledge or best practices should be baked in? Anthropic recommends defining 2-3 concrete use cases when you start a Skill. For each one, spell out four things:\n1 2 3 4 5 6 7 8 Use case: Weekly WeChat Official Account topic planning Trigger: User says \u0026#34;help me pick next week\u0026#39;s topics\u0026#34; or \u0026#34;what should I write this week\u0026#34; Steps: 1. Pull the last 7 days from the WeChat groups, X bookmarks, and Xiaohongshu viral hits the user follows 2. Filter out anything off-brand for the user\u0026#39;s account positioning (e.g. \u0026#34;AI tools + careers\u0026#34;) 3. Produce 5 candidate topics, each with a suggested title and angle 4. Tag which one is most likely to go viral, which is easiest to write, and which the user hasn\u0026#39;t covered before Result: A list of 5 candidate topics the user can immediately pick from to start writing Another example, this one from white-collar daily life:\n1 2 3 4 5 6 7 8 Use case: Weekly status report Trigger: User says \u0026#34;write my weekly report\u0026#34; or \u0026#34;summarize what I did this week\u0026#34; Steps: 1. Read this week\u0026#39;s calendar, Notion task records, and email drafts 2. Bucket items into \u0026#34;done / in progress / next week\u0026#34; 3. Translate each item into language a manager cares about (outcome-focused, not action-focused) 4. Output a weekly draft under 300 words, with 3 highlights from the week Result: A weekly draft ready to send to your manager If you can\u0026rsquo;t come up with 2-3 concrete use cases, what you\u0026rsquo;re describing is probably a one-off prompt, not a Skill.\n1.3 Figure out which category your Skill falls into Anthropic groups all Skills into three categories. Which one you\u0026rsquo;re building changes what you focus on.\nCategory 1: Document and resource creation.\nFor producing carefully formatted, high-quality output — documents, slide decks, spreadsheets, webpages, design files, code.\nThe canonical examples are the official docx / pptx / xlsx generation skills, plus the frontend-design skill.\nWhat to focus on for this category: embed a style guide, use a template structure for consistency, and run a quality checklist before final delivery. You usually don\u0026rsquo;t need external tools — Claude\u0026rsquo;s built-in capabilities are enough.\nCategory 2: Workflow automation.\nFor multi-step workflows where the steps are fixed and there\u0026rsquo;s a real methodology to follow, often spanning multiple MCP servers.\nThe canonical example is the skill-creator skill (which walks the user step by step through use-case definition, frontmatter generation, instruction writing, and validation).\nWhat to focus on for this category: gate every step with validation, give common structures a template, and document the rollback path when a step fails.\nCategory 3: MCP enhancement.\nFor attaching a \u0026ldquo;how to use this\u0026rdquo; manual to an MCP server you\u0026rsquo;ve already wired up. This category is mostly for MCP service providers writing skills for their own users —\na survey-tool company shipping a skill alongside their MCP that \u0026ldquo;automatically analyzes survey results and generates a chart report,\u0026rdquo;\na CRM company shipping one that \u0026ldquo;pulls customer data every week and generates a sales briefing.\u0026rdquo; If you\u0026rsquo;re not running an MCP service, this category is basically irrelevant to you — document generation (Category 1) and workflow automation (Category 2) are what most people will build.\nDocument generation skills live or die on the quality checklist. Workflow skills live or die on the seams between steps. MCP enhancement skills live or die on how well the domain knowledge is embedded. Decide which one you\u0026rsquo;re building before you start.\nNow that we know what a Skill is, we can start building one.\n2. Respect what makes a Skill a Skill: on-demand loading A Skill is not the same as CLAUDE.md or a system prompt.\nCLAUDE.md is more like persistent context. As long as you\u0026rsquo;re working in this project, it stays loaded and shapes the model\u0026rsquo;s behavior.\nA slash command is a manual command. The user has to type the command — only then does the agent run the workflow.\nA Skill sits between the two.\nIts defining property is on-demand loading.\nThe full SKILL.md is not in context by default. It only loads when the user\u0026rsquo;s input matches the Skill\u0026rsquo;s description.\nTwo benefits:\nSaves context. Cuts down noise from rules that don\u0026rsquo;t apply to the current task. But here\u0026rsquo;s the detail that matters:\nThe full Skill body isn\u0026rsquo;t persistent, but the description is — it stays in the matching pool the whole time.\nWhether the description is well-written directly determines whether the Skill ever fires correctly.\n2.1 The standard structure of a description Anthropic gives a three-part formula:\n[what this skill does] + [when to use it] + [key capabilities]\nThree hard rules:\nStay under 1024 characters Must include both \u0026ldquo;what it does\u0026rdquo; and \u0026ldquo;when to use it\u0026rdquo; Must be written in the third person 1 2 3 4 ✅ Processes Excel files and generates reports ✅ Breaks down viral Xiaohongshu notes and produces a reusable template ❌ I can help you process Excel files ❌ You can use this to break down Xiaohongshu viral hits A side-by-side.\n❌ Too vague:\n1 description: Handles topic selection This kind of description is almost guaranteed not to fire correctly — Claude sees the sentence and doesn\u0026rsquo;t know when to load it.\n❌ Missing trigger phrases:\n1 description: Generates well-structured multi-page WeChat Official Account posts Says what it does, doesn\u0026rsquo;t say when to use it.\n✅ What a good description looks like:\n1 description: Breaks down the cover, headline, opening, structure, and keywords of a viral Xiaohongshu note and outputs a reusable template. Use this skill when the user says \u0026#34;break down this note,\u0026#34; \u0026#34;analyze this viral hit,\u0026#34; \u0026#34;Xiaohongshu breakdown,\u0026#34; or pastes a Xiaohongshu link for Claude to learn from. Says what it does (breaks down viral notes → outputs a template) and when to use it (the user says break down / analyze / pastes a Xiaohongshu link).\n2.2 What to do when you have too many Skills If you\u0026rsquo;ve installed a lot of Skills, or one Skill is unusually long and specialized but used rarely, you can selectively turn off its auto-loading.\nClaude Code provides a dedicated frontmatter field (Claude Code only):\n1 2 3 4 5 --- name: publish-post description: One-click publish the current draft to WeChat Official Account, Zhihu, and Xiaohongshu disable-model-invocation: true --- With disable-model-invocation: true, Claude won\u0026rsquo;t auto-load this skill. It only fires when the user manually types /publish-post.\nWhy force publish-style actions to be manual? Because they have side effects — once it runs, it\u0026rsquo;s out the door, and you can\u0026rsquo;t undo it. You don\u0026rsquo;t want Claude to read your draft, decide \u0026ldquo;looks good enough,\u0026rdquo; and hit publish on its own.\nSkills that are worth turning off auto-loading:\nYear-end review Skill: used a few times a year Resume rewrite Skill: only during a job search Cover image Skill: only when publishing content Course notes Skill: only when working through a specific course Deploy / commit / send-message actions, anything with side effects (you don\u0026rsquo;t want Claude deciding when to fire those) Why turn off auto-loading: every Skill\u0026rsquo;s description gets preloaded into Claude\u0026rsquo;s system prompt at startup and stays in the matching pool. More Skills, more descriptions, higher persistent matching cost. Anthropic\u0026rsquo;s own recommendation: if you have more than 20-50 Skills enabled at once, take a look at which ones you can turn off.\nThe cost of turning one off is that the user has to remember to summon it.\nThere\u0026rsquo;s a counterpart to disable-model-invocation called user-invocable: false — the opposite direction, which blocks the user from manually invoking the skill from the / menu and only lets Claude trigger it. That fits \u0026ldquo;background knowledge\u0026rdquo; skills (something like context about an old internal system at your company that Claude needs in relevant tasks, but the user typing the command directly makes no sense).\nSo the first thing to decide about a Skill isn\u0026rsquo;t what rules to write. It\u0026rsquo;s whether the Skill should auto-trigger or only fire on manual call.\n3. Set the right tool boundaries for a Skill A Skill can restrict which tools it\u0026rsquo;s allowed to use.\nThis isn\u0026rsquo;t required, but it\u0026rsquo;s useful.\nBecause different Skills need different permissions.\nIn Claude Code, for example:\n1 2 3 4 allowed-tools: - Bash - Read - Grep For organizing study material, this might be enough:\n1 2 3 4 allowed-tools: - Read - Grep - Write For document organization:\n1 2 3 4 allowed-tools: - Read - Write - Grep For cover image generation — read references, write a prompt, then call the image tool:\n1 2 3 4 allowed-tools: - Read - Write - ImageGenerate For batch material processing, you\u0026rsquo;d open up more permissions:\n1 2 3 4 5 allowed-tools: - Read - Write - Edit - Bash The principle is simple: grant only the minimum permissions needed to do the job.\nDon\u0026rsquo;t let a Skill that just generates suggestions modify files by default.\nDon\u0026rsquo;t let a read-only analysis Skill execute arbitrary commands by default.\nMore tools doesn\u0026rsquo;t mean more capable. Sometimes it just means more risk.\n3.1 allowed-tools can be scoped down to specific subcommands A lot of people don\u0026rsquo;t realize allowed-tools can go finer than tool types — you can also restrict the specific call patterns.\nFor example, a skill that publishes articles, only allowed to run publish-related commands and nothing else:\n1 2 # Only runs the publish script and image uploads — can\u0026#39;t touch other files allowed-tools: \u0026#34;Bash(python publish.py *) Bash(curl *) Read Write\u0026#34; Or a data analysis skill, allowed to read files and run Python scripts, not allowed to modify anything:\n1 2 # Read-only + Python only — no writing files, no other commands allowed-tools: \u0026#34;Read Bash(python:*)\u0026#34; This way, even if the Skill gets misused, it can\u0026rsquo;t run anything dangerous like rm.\nIf you don\u0026rsquo;t write code and your Skill doesn\u0026rsquo;t call scripts, this section doesn\u0026rsquo;t apply — skip ahead to the security rules.\n3.2 Security rules This section isn\u0026rsquo;t a suggestion. These are hard rules Anthropic spells out explicitly.\nNo XML angle brackets \u0026lt; \u0026gt; in YAML frontmatter. The reason: frontmatter ends up in Claude\u0026rsquo;s system prompt. If a description or other field contains something like \u0026lt;instructions\u0026gt;do XYZ\u0026lt;/instructions\u0026gt;, it can get treated as prompt injection.\nSkill names can\u0026rsquo;t be prefixed with \u0026ldquo;claude\u0026rdquo; or \u0026ldquo;anthropic\u0026rdquo;. Those prefixes are reserved by Anthropic.\nThe filename has to be exactly SKILL.md. Case-sensitive. SKILL.MD / skill.md / Skill.md will all fail to upload.\nThe folder name must be kebab-case. No spaces, no underscores, no uppercase letters. For example:\n✅ notion-project-setup ❌ Notion Project Setup ❌ notion_project_setup ❌ NotionProjectSetup Don\u0026rsquo;t put a README.md inside the Skill folder. Any documentation Claude should read goes in SKILL.md or references/. A human-facing README belongs at the root of the GitHub repo when you publish the skill — not inside the skill folder.\n4. Configure the right model for the Skill A Skill can also specify a model, which gets overlooked but is useful.\nDifferent tasks place different demands on the model.\nWriting documents, generating images, analyzing data, scraping, organizing study notes, generating long-form posts — these really shouldn\u0026rsquo;t all default to the same model.\nA few examples:\nWriting documents: needs clear expression and stable structure. Use a model strong at writing and summarization. Image / design work: needs visual understanding and layout sense. Multimodal or design-strong models fit better. Data analysis: needs reliable handling of tables and explanation of results. A solid reasoning model at lower cost works. Scraping: often batch processing and extraction. Doesn\u0026rsquo;t need the strongest model — a cheap, fast one is a better fit. Study notes: needs categorization, distillation, and preservation of meaning. A stable cheap model is fine. Long-form X posts: needs structure, voice, pacing, and judgment about what\u0026rsquo;s worth saying. Use a model strong at writing. Resume editing: needs to weigh job requirements against what to emphasize. Use a more careful model. Some models are stronger but more expensive.\nSome are cheaper and already good enough for simple tasks.\nA mature Skill system shouldn\u0026rsquo;t default every task to the same model.\nIt should pick the right execution model based on the task type.\nHow to write it\nClaude Code provides a model field (Claude Code only). The official docs put it this way:\nModel to use when this skill is active. The override applies for the rest of the current turn and is not saved to settings; the session model resumes on your next prompt. Accepts the same values as /model, or inherit to keep the active model.\nTwo things worth noting:\nThe switch only applies to the current turn. The next time the user sends a new prompt, it automatically reverts to the original session model. This is a temporary switch, not a permanent settings change. inherit is the default behavior — keep the current session model. If you only want certain skills to specifically use a stronger model, write the specific model name in those skills\u0026rsquo; frontmatter; leave the others blank or set them to inherit. How to write it. Heavy task with a strong model:\n1 2 3 4 5 6 7 8 --- name: deep-post-analysis description: Deep breakdown of viral Xiaohongshu notes — cover, copy, pacing, psychological hooks, all dimensions model: claude-opus-4-5 --- # Deep Viral Breakdown ... Simple task with a cheap model:\n1 2 3 4 5 --- name: daily-topic-pool description: Pull today\u0026#39;s candidate topics from the WeChat Official Accounts and X bookmarks the user follows model: claude-haiku-4-5 --- Unspecified (sticks with the current session model):\n1 2 3 4 5 --- name: weekly-summary description: Organize this week\u0026#39;s WeChat Official Account back-end analytics into a review report model: inherit --- A quick note on the effort field\nPaired with model is the effort field (also Claude Code only), which controls reasoning depth:\n1 effort: high # low / medium / high / xhigh / max Simple tasks use low effort to save time and money; complex decisions use high effort to trade cost for accuracy. The available levels depend on the model — Haiku doesn\u0026rsquo;t support max; Opus does.\n5. Progressive disclosure: keep SKILL.md small Skill file loading is a three-tier system:\nTier 1 (Description): always loaded into Claude\u0026rsquo;s system prompt. So the description needs to carry enough information for Claude to decide \u0026ldquo;when to use this skill,\u0026rdquo; but not be too long. Tier 2 (SKILL.md body): loaded when Claude decides the skill is relevant to the current task. Contains the full instructions. Tier 3 (bundled files): extra files sitting in the skill directory. Claude only reads them when it actually needs to. Once you understand the three tiers, you understand why SKILL.md shouldn\u0026rsquo;t carry everything.\nThink of it as the entry point.\nIt should hold:\nWhen to trigger. The principles. The execution steps. Which tools are needed. Where the rest of the material lives. How to validate at the end. Don\u0026rsquo;t dump every reference, long example, script, and template into it.\nA better approach is layered:\n1 2 3 4 5 6 7 8 9 10 11 12 generate-cover-image/ SKILL.md references/ cover-image-style-guide.md common-platform-sizes.md good-vs-bad-examples.md scripts/ check-image-dimensions.py export-multi-platform.py assets/ cover-image-template.md fonts-and-color-examples.json Different directories hold different things.\nreferences/ holds long docs, style guides, platform sizes, detailed cases.\nscripts/ holds executable code — check image dimensions, batch rename files, export multi-platform versions, format tables. Deterministic operations are more reliable as scripts than as model improv every time.\nassets/ holds templates, schemas, images, sample files, output formats.\nThe benefit:\nThe agent doesn\u0026rsquo;t have to load everything every time. It reads SKILL.md first to understand the overall flow. When it actually needs detail, it reads a reference or runs a script. That\u0026rsquo;s Progressive Disclosure.\nRead on demand. Execute on demand.\nDon\u0026rsquo;t shove everything into context up front.\n5.1 Why bother with progressive disclosure Reason 1: save tokens (save money)\nEvery Claude conversation is billed by token.\nEvery Skill stuffing all its content into context = a few thousand extra tokens per conversation. One user calling it 100 times a day, that\u0026rsquo;s tens of thousands of calls a month — real money on the API bill.\nThe Anthropic docs put it directly: the point of progressive disclosure is \u0026ldquo;minimizing token usage while preserving expertise.\u0026rdquo; The three-tier system loads content only when Claude actually needs it, saving money at the source.\nReason 2: save context window space (preserve model performance)\nThis one\u0026rsquo;s more subtle but just as critical.\nThe context window has a cap — say, 200K tokens. Everyone shares that space:\nSystem prompt The last few dozen rounds of conversation history Project rules in CLAUDE.md All currently loaded Skills The user\u0026rsquo;s prompt this round Tool call return values If every Skill stuffs detailed references, case libraries, and long instructions into context, three things happen:\nFirst, it squeezes out conversation history — old conversation gets auto-compacted or dropped, and Claude forgets what was said earlier. This is most obvious in long conversations — you hit turn twenty and notice Claude has forgotten the key decisions from the first ten turns, often because the space got eaten by a stack of skills.\nSecond, it hurts focus — too much content and the model gets lost in long stretches of text. Research calls this \u0026ldquo;lost in the middle\u0026rdquo;: key instructions buried in the middle get ignored.\nThird, it triggers auto-compaction — past a threshold, the system force-summarizes the earlier context, original details are lost, and model performance drops. Claude Code has a mechanism called auto-compaction specifically for this — once it fires, only the last 5000 tokens of each skill\u0026rsquo;s most recent invocation are kept; anything earlier is gone.\nWhat the three-tier system does here: only \u0026ldquo;navigation info\u0026rdquo; lives in context for SKILL.md; the detailed material stays on disk for Claude to read on demand. Material that isn\u0026rsquo;t in context doesn\u0026rsquo;t take up space, and doesn\u0026rsquo;t drag down the model.\nSaving money is visible. Saving context space isn\u0026rsquo;t — but in long conversations the second matters more.\n5.2 How big should SKILL.md be? Claude Code\u0026rsquo;s docs put it directly: \u0026ldquo;Keep SKILL.md under 500 lines.\u0026rdquo;\nNot because 500 is a magic number, but because past that length you\u0026rsquo;ve usually mixed too many things together.\nIf a Skill is too long, split it in this order:\nMove long explanations to references/. Turn stable operations into scripts/. Move templates and samples into assets/. If it\u0026rsquo;s still long after that, you might not have one Skill — you have several.\n6. After writing the Skill, you still need to validate, score, and iterate Writing the Skill doesn\u0026rsquo;t mean it works.\nThis part matters.\nIf you want a Skill to be reliable, run at least three kinds of validation:\nDoes it run. Does it trigger correctly. Are the results actually better than not using the Skill. The third is the easiest one to skip.\nMany Skills just \u0026ldquo;look done.\u0026rdquo; But did it raise quality? Did it cut errors? Did it make the output match what the user wanted? Without an evaluation, nobody actually knows.\nClaude Code\u0026rsquo;s Skill Creator gives a useful framing: don\u0026rsquo;t ship a Skill on vibes. Set up test data, run an evaluation, look at the results, iterate.\nThe core flow:\nDefine the task the Skill solves. Write a first draft of the Skill. Prepare test data / test prompts. Run the Skill against those tests. Evaluate each result. Score each one. Modify the Skill based on the failures. Run another round. Repeat until the main tests hit the minimum acceptable score. Think of it as writing tests for the Skill.\nNot as rigid as unit tests, but the idea is similar: don\u0026rsquo;t trust the Skill file itself, look at how it performs on sample tasks.\n6.1 Three ways to test Anthropic recommends a three-layer testing setup:\nManual testing in Claude: just type and try it. Fastest, zero setup, good for early rapid iteration. Scripted testing in Claude Code: automated test cases you can rerun after every change, good for the mid-stage when you need repeatable validation. Programmatic testing through the Skills API: a full evaluation suite running against a fixed test set. Right for a mature Skill being shipped to thousands of users. Which one you pick depends on the Skill\u0026rsquo;s reach. Personal use can stop at the first layer; team use needs at least the second; thousands of enterprise users needs the third.\n6.2 Layer 1: validate that it runs Test the most basic execution first.\nFor example:\nAre the file paths right? Are the tool permissions enough? Does the script execute? Can the references be read? Are the assets templates used correctly? Does the output match the expected format? Are any required steps missing? If a Skill can\u0026rsquo;t even complete a real task, talking about triggers and quality is pointless.\n6.3 Layer 2: validate that it triggers correctly For auto-loaded Skills, also test triggering.\nYou can\u0026rsquo;t just test one phrase:\n1 Please run the X long-form post skill. Real users don\u0026rsquo;t talk like that.\nPrepare a set of phrases real users might actually use:\n1 2 3 4 5 Expand this idea into a long thread Write me a long X post Can this be turned into longer content Rewrite this in my voice Give me a more attention-grabbing opening For a cover image Skill, you might test:\n1 2 3 4 Make me a cover image Add an image to this article Make a tech-style poster for this title Make a Xiaohongshu cover That\u0026rsquo;s trigger testing.\nEach test should be tagged:\n1 2 3 4 [ {\u0026#34;query\u0026#34;: \u0026#34;Expand this idea into a long thread\u0026#34;, \u0026#34;should_trigger\u0026#34;: true}, {\u0026#34;query\u0026#34;: \u0026#34;Check today\u0026#39;s weather for me\u0026#34;, \u0026#34;should_trigger\u0026#34;: false} ] If something that should trigger doesn\u0026rsquo;t, add keywords to the description.\nIf something that shouldn\u0026rsquo;t trigger does, narrow the description.\nThis step optimizes the Skill\u0026rsquo;s entry point.\nIf the entry is wrong, the body doesn\u0026rsquo;t matter.\n6.4 Layer 3: prepare test data and evaluation cases The bigger task is validating result quality.\nYou need test data.\nThat doesn\u0026rsquo;t mean the user has to hand-craft a pile of test material.\nA more reasonable approach: have the agent generate a first batch of test data and evaluation cases based on the Skill\u0026rsquo;s purpose, and the user just reviews whether they match real scenarios.\nFor example:\nCover image Skill: 5 titles, 5 content types, a few reference images. Long-form X Skill: 5 raw ideas, target audience, ideal final-draft style. Study notes Skill: a few class notes, excerpts, OCR text from screenshots. Resume rewrite Skill: different job descriptions, the original resume, ideal direction. Weekly status report Skill: a week of scattered notes, meeting minutes, completed items. Each test case should ideally include:\n1 2 3 4 5 6 7 8 9 10 11 { \u0026#34;id\u0026#34;: 1, \u0026#34;prompt\u0026#34;: \u0026#34;How a user would phrase this task\u0026#34;, \u0026#34;files\u0026#34;: [\u0026#34;test input file\u0026#34;], \u0026#34;expected_output\u0026#34;: \u0026#34;What counts as a good result\u0026#34;, \u0026#34;assertions\u0026#34;: [ \u0026#34;Must preserve the user\u0026#39;s original point\u0026#34;, \u0026#34;Must produce a strong opening\u0026#34;, \u0026#34;Must not fabricate experiences the user didn\u0026#39;t mention\u0026#34; ] } The format isn\u0026rsquo;t the point. The mindset is:\nDefine \u0026ldquo;what counts as good\u0026rdquo; up front.\nOtherwise the evaluation becomes a vibe check after the fact.\n6.5 Layer 4: scoring needs both quantitative and qualitative metrics Anthropic recommends splitting success criteria into two kinds:\nQuantitative metrics (things you can count directly):\nSkill triggers on 90% of relevant queries Tool calls to finish the workflow ≤ X 0 failed API calls Tokens consumed Qualitative metrics (things you have to observe to judge):\nAfter running, the user doesn\u0026rsquo;t have to follow up with \u0026ldquo;now what\u0026rdquo; Same request run 3-5 times produces consistent structure and quality A new user can finish the task on their first try without much guidance After running the tests, don\u0026rsquo;t just write \u0026ldquo;looks fine.\u0026rdquo;\nScore each case, say 0 to 10.\nA simple rubric:\n0-2: didn\u0026rsquo;t complete the task, or wrong direction. 3-4: tangentially related, missed key requirements. 5-6: usable but with obvious problems. 7-8: stable quality, minor details to fix. 9-10: matches expectations, can serve as a reference output. A reasonable bar: main test cases score at least 5.\nIf an important case scores below 5, the Skill isn\u0026rsquo;t reliable yet.\nFor a high-frequency Skill, that usually means it\u0026rsquo;s broken in a common scenario.\n6.6 Layer 5: baseline comparison If you want to be more rigorous, run a baseline comparison.\nSame test, two runs:\nWithout the Skill. With the Skill. Then compare.\nClaude Code\u0026rsquo;s Skill Creator evaluation includes something like this — A/B comparison, skill-enabled vs skill-disabled.\nWorth doing.\nA Skill shouldn\u0026rsquo;t just \u0026ldquo;run.\u0026rdquo; It should prove it\u0026rsquo;s useful.\nIf the no-Skill result is already 7 and the Skill version is also 7, the Skill isn\u0026rsquo;t adding much.\nIf no-Skill is 4 and with-Skill is 8 — that\u0026rsquo;s the Skill actually encoding useful experience.\n6.7 Layer 6: fix the Skill based on the failures Scoring isn\u0026rsquo;t decoration. Scoring tells you what to change.\nCommon fixes:\nTrigger fails: edit the description, add real trigger phrases. False trigger: narrow the description, add negative cases. Missing steps: edit the workflow in SKILL.md. Unstable output format: add an output template to assets/. Unstable deterministic checks: write a script. Reference info too long: split it into references/. Tool permissions too tight: add allowed-tools. Tool permissions too loose: tighten allowed-tools. Model too weak: switch to a better-fit model. Reasoning depth too shallow: bump up the effort level. Then run the evaluation again.\nThat\u0026rsquo;s the Skill optimization loop:\n1 2 3 4 5 6 7 8 Write Skill → Prepare test data → Run evaluation → Score each case → Find failure causes → Modify the Skill → Re-run evaluation → Hit the minimum bar, or at least know its limits For a high-frequency, complex, or shared Skill, run this loop.\nOtherwise it\u0026rsquo;s just an unvalidated prompt file.\n6.8 Use the official skill-creator to scaffold and review Anthropic ships a built-in Skill called skill-creator. You can install it from the Claude.ai plugin directory, or download it directly in Claude Code.\nIt can do a few things for you:\nGenerate a first draft of a skill from a natural-language description Produce a correctly formatted SKILL.md (with frontmatter) Suggest trigger phrases and structure Review an existing skill and flag common issues (vague description, missing triggers, structural problems) After you run tests and hit edge cases, fold that feedback back in for iteration Calling it is simple:\n1 Use the skill-creator skill to help me build a skill for [my use case] It won\u0026rsquo;t run evaluations automatically and won\u0026rsquo;t produce quantitative reports, but it\u0026rsquo;ll save you from the most common from-scratch mistakes.\nFor beginners, strongly recommend using skill-creator for the first-pass skeleton, then editing yourself.\n7. How do you share a Skill once it\u0026rsquo;s written? Once a Skill is written, the next step is getting it into other people\u0026rsquo;s hands.\nAnthropic\u0026rsquo;s currently recommended path:\nStep 1: host it on GitHub\nUse a public repo (if it\u0026rsquo;s an open-source skill) Write a clear README at the repo root (this is for human readers — doesn\u0026rsquo;t conflict with the \u0026ldquo;no README.md in the skill folder\u0026rdquo; rule, since the human README sits alongside the skill folder, not inside it) Include a few screenshots of it in use Step 2: upload to Claude.ai\nZip the entire skill folder Open Claude.ai → Settings → Capabilities → Skills → Upload skill Pick the zip → enable → test Step 3: organization-level distribution\nAnthropic has shipped organization-level skill deployment — an admin can push a skill to the entire workspace in one shot. Every member gets it automatically, updates automatically, managed centrally. If your skill is for a team, this path is more reliable than asking everyone to upload on their own.\nStep 4: programmatic use via API\nFor apps, agents, or automation pipelines:\nUse the /v1/skills endpoint to list and manage skills Add skills to Messages API requests via the container.skills parameter Version-control them in the Claude console Combine with the Claude Agent SDK to build custom agents Which path is best depends on the situation.\n7.1 When describing your Skill, focus on outcomes When writing a README, marketing copy, or the skill intro in MCP docs, remember: users don\u0026rsquo;t care what you are, they care what result you give them.\n❌ Weak:\nThe ProjectHub skill is a folder containing YAML frontmatter and Markdown instructions that calls our MCP server\u0026rsquo;s tools.\n✅ Strong:\nThe ProjectHub skill lets a team spin up a complete project workspace — pages, databases, templates — in 30 seconds, instead of 30 minutes of manual setup.\nThe first version describes the file structure (user doesn\u0026rsquo;t care). The second describes the time saved (user gets it instantly).\n8. What does a good Skill look like? What makes a good Skill?\nA few things:\nIt triggers when it should. It stays quiet when it shouldn\u0026rsquo;t. Tool permissions are enough, not excessive. It uses the right model and reasoning depth. SKILL.md is short enough (under 500 lines), with detailed material loaded on demand. Important Skills come with test data, an evaluation, failure cases, and iteration. So when writing a Skill, don\u0026rsquo;t only ask:\n\u0026ldquo;What do I tell the model?\u0026rdquo;\nAlso ask:\nWhat user phrases should make the model invoke this Skill? Once it fires, what workflow should it follow? Which tools is it allowed to use, and which is it not? Which model is best for it? What content stays in the main file, and what gets split out? How do I prove it actually helps? How do other people install it, upgrade it, and recover when something breaks? That\u0026rsquo;s the real difference between a Skill and a prompt.\nAppendix A: SKILL.md starter template Copy and adapt. Note: YAML field names (name / description / model, etc.) are hard requirements of the Skill system and must stay in English. But the field values, Markdown body, and examples can all be in your own language.\n1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 --- name: weekly-topic-planner description: Plans next week\u0026#39;s WeChat Official Account topics for a content creator. Use when the user says \u0026#34;pick next week\u0026#39;s topics,\u0026#34; \u0026#34;what should I write this week,\u0026#34; or \u0026#34;help me figure out what to publish next week.\u0026#34; --- # Weekly WeChat Official Account Topic Planning ## Steps ### Step 1: Gather material Find hot topics from the last 7 days across the WeChat groups, X bookmarks, and viral Xiaohongshu notes the user follows. ### Step 2: Filter by account positioning Read the account positioning in `references/account-positioning.md` and filter out anything off-brand. ### Step 3: Produce the topic list Output 5 candidate topics, each with: - Suggested title - Angle - Estimated viral probability - Writing difficulty (Add more steps as needed.) ## Examples ### Example 1: typical case User says: \u0026#34;Help me pick next week\u0026#39;s topics\u0026#34; Steps: 1. Pull viral hits from the last 7 days (from the material pool) 2. Filter out off-brand topics by account positioning 3. Output a 5-topic candidate table Result: 5 candidate topics, ready for the user to pick one and start writing ## Troubleshooting ### Error: can\u0026#39;t pull material **Cause:** source link broken or network down **Fix:** check the links in `references/sources.md`, update broken sources Appendix B: full Skill YAML field list Example:\n1 2 3 4 5 6 7 8 9 10 11 12 13 --- name: weekly-topic-planner description: Plans next week\u0026#39;s WeChat Official Account topics for a content creator. Use when the user says \u0026#34;pick next week\u0026#39;s topics\u0026#34; or \u0026#34;what should I write this week.\u0026#34; license: MIT allowed-tools: \u0026#34;Bash(python:*) Bash(npm:*) WebFetch\u0026#34; model: claude-opus-4-5 effort: high metadata: author: Zhang Da version: 1.0.0 mcp-server: wechat-mp-server category: content-creator --- A note for cross-platform skills: if you want the skill to also run on Codex or Hermes, stick to the universal fields.\nClaude Code-only fields are ignored on other platforms.\nAppendix C: quick checklist Before you start\nDefined 2-3 concrete use cases Decided which tools to use (Claude\u0026rsquo;s built-ins, or MCP) Read through a few example skills Mapped out the folder structure While writing\nFolder name is kebab-case SKILL.md filename is exactly right (case included) YAML frontmatter has \u0026mdash; delimiters name field: kebab-case, no spaces, no uppercase description includes both \u0026ldquo;what it does\u0026rdquo; and \u0026ldquo;when to use it\u0026rdquo; No XML angle brackets \u0026lt; \u0026gt; anywhere Instructions are clear and actionable Error handling is included Examples are provided References are linked cleanly SKILL.md is under 500 lines Before uploading\nTested triggering for obvious tasks Tested triggering for paraphrased requests Verified it doesn\u0026rsquo;t trigger on unrelated topics Functional tests pass Tool integrations work (if any) Zipped up After uploading\nMonitor for under-triggering and over-triggering Collect user feedback Iterate on description and instructions based on feedback A prompt solves this one conversation.\nA Skill solves the next hundred similar tasks.\nIt encodes experience into a triggerable, testable, distributable workflow.\n","permalink":"https://ryrenz.com/ai/industrial-grade-skill-five-principles/","summary":"A Skill isn\u0026rsquo;t a longer prompt — it\u0026rsquo;s a triggerable, executable, testable, maintainable workflow. From what a Skill actually is, to on-demand loading, tool boundaries, model selection, progressive disclosure, validation loops, and how to share it with other people — the full playbook for writing industrial-grade Skills.","title":"How to Write Industrial-Grade Skills"},{"content":"Three years in, models are 100x more capable. Why is AI still just a chat box?\nAfter two years building AI products, I\u0026rsquo;m more and more convinced of one thing: the chat box is the Nokia of the AI era.\nIt\u0026rsquo;s just the first form of AI we got our hands on. Higher-dimensional products will replace it.\nI. First-gen AI: chat boxes — write your prompts Regular people aren\u0026rsquo;t bad at AI because AI is dumb. They\u0026rsquo;re bad at it because they don\u0026rsquo;t know what to ask, or how to ask.\nQuick example.\nBefore AI, there was a phrase that got thrown around constantly: \u0026ldquo;Can\u0026rsquo;t you just Google it?\u0026rdquo;\nA decade after search engines went mainstream, plenty of regular people still couldn\u0026rsquo;t break what they wanted into keywords. If \u0026ldquo;turn a question into 3-5 search keywords\u0026rdquo; already had a barrier, then \u0026ldquo;write a complete prompt with full context, clear goals, and verifiable tests\u0026rdquo; is a much higher one.\nJust look at how many DeepSeek and Doubao courses are selling on the market right now.\nSo I keep saying: the real difficulty for regular people using AI isn\u0026rsquo;t technical. It\u0026rsquo;s expressing what they want clearly.\nClaude Code and Cursor are obviously built for programmers. A regular person sees the UI and closes it instantly. Doubao is a bit better, but watch the people around you who say they want to learn AI — almost none of them actually use Doubao well.\nThey don\u0026rsquo;t know how to follow up. They don\u0026rsquo;t know how to give context. They often haven\u0026rsquo;t even thought through what they want. When AI answers off-target, they have no idea how to course-correct. One message bounces off another, and both sides get stuck.\nA hot product doesn\u0026rsquo;t mean the form is right. What matters first is what users actually need.\nIt wasn\u0026rsquo;t until OpenClaw arrived that AI agents started leaking out of the small programmer crowd and reaching everyone else.\nII. Second-gen AI: agents — anticipate what humans need OpenClaw launched in November 2025 (originally Clawdbot, renamed to OpenClaw in January 2026). It hit 100K GitHub stars in 3 months — the fastest star growth of any open-source project in GitHub history. No IDE, no terminal, you don\u0026rsquo;t even have to leave WeChat. Drop a message in your chat app and AI handles it.\nWhat expands AI\u0026rsquo;s reach has never been technical sophistication. It\u0026rsquo;s lowering the barrier to use.\nOpenClaw pointed at a direction: AI that does things on its own.\nYou wake up in the morning, and your \u0026ldquo;lobster\u0026rdquo; has already gathered today\u0026rsquo;s important news and written it up as an analysis report for you to review.\nYour meeting ends, and the lobster has already summarized the notes and flagged the key points.\nUsers get more done with fewer prompts. Efficiency goes up, the barrier and the mental load come way down.\nStill not enough. Most work today still needs the user to plan things out. What people really want is an agent that\u0026rsquo;s foolproof: say nothing, and AI knows what you\u0026rsquo;re trying to do.\nThat sounds impossible, but parts of it already exist. The most familiar one is Computer Use.\nAnthropic shipped the first consumer version in October 2024. They gave Claude a wild new ability — let AI look at your screenshot, move the mouse like a human, type on your keyboard. You hand it a task, it opens the app, finds the info on the web, fills out forms, hits submit.\nIn January 2025, OpenAI followed with Operator, \u0026ldquo;AI\u0026rsquo;s own browser\u0026rdquo; that reads pages, clicks buttons, and fills forms like a human. That July, Operator merged into ChatGPT and became ChatGPT Agent.\nAround the same time, products like Perplexity Comet started fusing AI with the browsing experience more deeply: not just searching for answers, but organizing information, comparing options, and suggesting next steps.\nHuman Security\u0026rsquo;s data is even more direct: since July 2025, network requests from agentic browsers are up 6900%. AI is taking over screen operations faster than anything before it.\nIn March 2026, Anthropic pushed Computer Use further: background parallel tasks, scheduled tasks, and the Dispatch feature — \u0026ldquo;go out for dinner, AI keeps working for you.\u0026rdquo;\nThe shared direction across these products:\nAI no longer waits for you to ask. It watches what you do to figure out what you need.\nIt understands the page you\u0026rsquo;re looking at It answers questions based on context It occasionally runs simple operations for you Now picture this:\nA truly mature Computer Use agent that monitors your screen all day Runs silently and predicts in real time what problems you\u0026rsquo;re about to hit The moment it detects something, it pops up a few suggested fixes, you pick one, and it executes No more opening a chatbot every time you have a problem, scratching your head over how to phrase the prompt, then walking through the AI\u0026rsquo;s solution step by step yourself.\nThe technology isn\u0026rsquo;t there yet: slow, unstable, very sensitive to page changes, bad at complex flows. Two big issues remain unsolved:\nPrivacy: AI has to know what you looked at, what you clicked, where you hesitated, what you copied. That\u0026rsquo;s basically full behavioral monitoring. Performance and cost: Real-time AI inference means analyzing every page and reasoning over every action. With current models, latency is too high and token costs are too steep. Not worth it. If models keep improving and these problems get solved, Computer Use becomes the technology that lets regular people use AI without writing prompts.\nBut Computer Use still has a hard ceiling. You can only use AI on a computer. The moment you\u0026rsquo;re away from a screen, how does AI predict what you need next?\nAn agent trapped in a screen knows everything about the world, but not enough about you. That\u0026rsquo;s where software has to hand the ball off to hardware.\nIII. Third-gen AI: wearables + AI — they know everything about you Before talking about glasses, watches, and brain interfaces, here\u0026rsquo;s a principle a lot of people still miss: AI\u0026rsquo;s output quality depends heavily on the user\u0026rsquo;s input quality.\nThe people who use AI brilliantly today almost all have a complete \u0026ldquo;second brain\u0026rdquo; behind them. A personal knowledge base built up over years or even decades, with all their notes, project docs, meeting transcripts, codebases, even every past conversation. They feed all of this to AI, and AI becomes a version that serves only them. Knows their preferences, their go-to libraries, their thinking patterns, their entire history.\nThat\u0026rsquo;s why the AI of a power user is just smarter than the AI of a regular person. The model isn\u0026rsquo;t stronger; the input quality is higher and the customization is deeper. The \u0026ldquo;lobster\u0026rdquo; of a power user just gets it.\nFor a regular person to build the whole system today, there are three mountains in the way:\nHardware barrier: you need a Mac or a powerful PC, install a stack of tools, configure environments, hook up APIs, try to deploy local models. Content barrier: you need a long-term writing and recording habit, and the discipline to digitize and structure it. Most people can\u0026rsquo;t keep a journal for three days. Their work docs vanish when they leave a job. Meeting recordings never get saved. Engineering barrier: you need basic agent knowledge to build your own knowledge base, set up workflows, and continuously tune the whole setup. Stack those three together and you get a wall that separates 99% of regular people from \u0026ldquo;high-intelligence AI.\u0026rdquo;\nBut flip the framing — what if AI is something you wear?\nImagine you\u0026rsquo;re wearing AI glasses that stay on all day, an AI watch on your wrist, a hat on your head that reads your attention. They build that second brain in the cloud for you, frictionlessly, in the background.\nWhat you hear, see, say, write, type, even your physical health — all of it gets transcribed, tagged, and archived automatically.\nIn no time, everyone has their own \u0026ldquo;digital twin\u0026rdquo; — a life record more complete than your own memory. You don\u0026rsquo;t have to scramble for a prompt to get AI to understand you. It is you.\nWhen that day comes, all the courses out there charging you to learn prompt engineering, RAG, or how to build a second brain go completely obsolete. You don\u0026rsquo;t need to \u0026ldquo;manage\u0026rdquo; a knowledge base. Every input and output syncs to your personal cloud automatically, gets categorized, gets fed to AI.\nThe skills that look \u0026ldquo;must-learn\u0026rdquo; today are really transitional skills before AI goes mainstream — like learning \u0026ldquo;how to use a Nokia for email\u0026rdquo; in 2000.\nThe exciting part is plenty of startups are already on this, and consumer products are starting to ship.\nLimitless (Meta acquired), Bee (Amazon acquired), Omi — necklaces, wristbands, tiny clip-ons in every form. They\u0026rsquo;re already on the necks, wrists, and collars of a chunk of Silicon Valley. Different shapes, same logic: wear it all day, record audio for 12+ hours, and at night the day\u0026rsquo;s audio gets transcribed, tagged, deduplicated, and archived into your personal knowledge base. The next morning your AI shows up with the context of \u0026ldquo;what you said yesterday, what you heard, what you promised to whom.\u0026rdquo;\nLimitless got 10,000 preorders in 24 hours when it launched. In December 2025, Meta acquired the entire Limitless team and tech and folded it into Reality Labs.\nCES 2026 had a whole row of booths for these \u0026ldquo;record your whole life\u0026rdquo; devices. Some commentators called it \u0026ldquo;an incoming Black Mirror,\u0026rdquo; but the more accurate framing is this — it\u0026rsquo;s the most underrated piece of the next-generation AI product puzzle.\nMaybe today only the AI-savvy crowd is excited about it, but wasn\u0026rsquo;t GPT-3 the same? Once a killer product like ChatGPT lands, it goes mainstream fast.\nThis is what the hardware layer is really solving — giving AI an entry point so it can record your life, truly understand you, and become you.\nOnce you see this, glasses, Apple, and brain interfaces start to make a different kind of sense.\nThread one: smart glasses I wrote a piece on smart glasses earlier, \u0026ldquo;Smart Glasses: The Next Trillion-Dollar Market You Can\u0026rsquo;t Ignore\u0026rdquo; — worth a read if you\u0026rsquo;re curious.\nMeta\u0026rsquo;s CTO Andrew Bosworth said something on the a16z blog this year:\n\u0026ldquo;The next wave of consumer tech won\u0026rsquo;t run on taps and swipes — it\u0026rsquo;ll run on intent.\u0026rdquo;\n\u0026ldquo;Run on intent\u0026rdquo; — for AI to do things before you even speak, it needs to see what you see and hear what you hear.\nGlasses are the most elegant form for this. Closest to your eyes and ears, lightweight, hands-free, doesn\u0026rsquo;t block your view. What it sees is what you see. What it hears is what you hear.\nMeta is the most aggressive player. Zuckerberg is personally pushing it for a reason — he\u0026rsquo;s betting on \u0026ldquo;the next consumer electronics entry point.\u0026rdquo;\nThread two: Apple\u0026rsquo;s CEO change This also explains why Apple announced a CEO change in April 2026 — John Ternus, hardware engineer by background, takes over from Tim Cook in September. His past scope at Apple covered the hardware engineering teams for iPhone, iPad, Mac, Apple Watch, AirPods, and Vision Pro.\nCNBC\u0026rsquo;s deep dive on the move spelled the signal out clearly: AI differentiation is no longer about cloud scale or model performance. It\u0026rsquo;s about integrating silicon and software on the device.\nTernus taking over is basically Apple admitting: for the next decade, Apple won\u0026rsquo;t compete with OpenAI or Anthropic on models. Apple will pack AI into every device on your body.\nVision Pro shipped only about 45,000 units in Q4 2025. Not a hit. But Apple didn\u0026rsquo;t drop spatial computing. They swapped the CEO for a hardware guy and plan to ship AI glasses next year.\nWhat that tells you: Vision Pro was too heavy, too expensive, and the ecosystem wasn\u0026rsquo;t there. But the \u0026ldquo;spatial computing + wearable + AI\u0026rdquo; direction is right. Apple\u0026rsquo;s bet is that Ternus, having absorbed the Vision Pro lessons, can do better.\nThread three: brain-computer interfaces Even more aggressive than smart glasses are brain-computer interfaces. The topic felt like sci-fi a few years ago, but non-invasive BCIs (no skull surgery — just a headband or hat) are already here.\nBrainCo — founder Han Bicheng, a Harvard PhD dropout who started a medical-focused BCI company. In September 2025 they shipped the Revo2 prosthetic hand: 383 grams, 0.1mm precision, 50 newtons of grip strength. Amputees wearing it can play piano. In January 2026, BrainCo\u0026rsquo;s pediatric ADHD treatment device, Focus Xin, passed China\u0026rsquo;s medical device approval and is now being sold in hospitals. Earlier this year they filed for a Hong Kong IPO at over $1.3B valuation. One of the first BCI companies to scale into mass production.\nSabi — this California startup just exited stealth in April 2026, with early OpenAI investor Vinod Khosla backing them. Sabi claims they\u0026rsquo;ve collected the world\u0026rsquo;s largest neural dataset and trained the strongest Brain Foundation Model. The product they\u0026rsquo;re shipping is a beanie, packed with 70,000 to 100,000 ultra-dense EEG sensors, decoding your inner monologue directly: think a sentence, see it on the screen. No talking, no typing. First version ships at the end of this year, next version is a baseball cap, target speed is 30 wpm — slower than most people type, but once the basic loop closes, performance climbs fast.\nPeople call Sabi \u0026ldquo;the non-invasive Neuralink competitor.\u0026rdquo; I think they have it backwards. Neuralink\u0026rsquo;s open-skull-and-implant route is destined to stay niche medical. The product that actually ends up on every head will be something like Sabi: a hat you just put on.\nOne has already entered the medical system. Another is shipping with 100K sensors. BCIs went from \u0026ldquo;still 10 years out\u0026rdquo; to \u0026ldquo;this year\u0026rdquo; in two years.\nThese three hardware threads — glasses, Apple, BCIs — look like they\u0026rsquo;re each going their own way. They\u0026rsquo;re actually doing the same thing: giving AI the legs it needs to walk into the physical world, through wearables.\nBut hardware has a fatal bottleneck: cloud LLMs can\u0026rsquo;t handle the real-time physical world. Latency is too high, and the models have no physical intuition.\nYou can\u0026rsquo;t send every frame from glasses to the cloud and wait for GPT to reason. Neural signals from a BCI hat need millisecond responses. They can\u0026rsquo;t wait for network round trips. A watch\u0026rsquo;s battery can\u0026rsquo;t take the power draw of a giant model.\nSo the hardware bottleneck kicks the ball back to the foundation. If AI is going to live in trillions of wearables, it needs a brand-new architecture. A small AI that runs on the chip in your device.\nIV. World models World models are the underlying engine that lets AI understand the physical world.\nEvery frame from glasses, every neural signal from a BCI hat, every millisecond of motion data from a watch — send it all back to the cloud for GPT to process and the latency makes the whole thing impossible. Before AI finishes saying \u0026ldquo;watch out for that car,\u0026rdquo; the car has hit you.\nThere\u0026rsquo;s only one path to truly seamless wearable AI: world models. And that path runs through Yann LeCun.\nLeCun left Meta in November 2025, after 10+ years there as Chief AI Scientist, Turing Award winner. Why does a 65-year-old AI legend quit? Because he thinks Meta is doubling down on the wrong path.\nIn January 2026, in an MIT Technology Review interview, he said:\n\u0026ldquo;LLMs are limited to the discrete world of text. They can\u0026rsquo;t truly reason or plan, because they lack a model of the world.\u0026rdquo;\nIn another interview, he went deeper: the essence of intelligence isn\u0026rsquo;t \u0026ldquo;being able to talk.\u0026rdquo; It\u0026rsquo;s \u0026ldquo;predicting the consequences of action\u0026rdquo;. Before you reach for a cup, your brain has already simulated how your hand will move and whether the cup will tip — that\u0026rsquo;s what a world model means. LLMs only learn relationships between words. They don\u0026rsquo;t know apples fall from trees, or that water can\u0026rsquo;t flow uphill.\nAnother LeCun quote pierces the industry\u0026rsquo;s self-hypnosis:\n\u0026ldquo;Language turned out to be the easy part. The hard part is the physical world.\u0026rdquo;\nSo you see AI passing exams and writing code today, yet there are still no household robots and no L5 self-driving. By LeCun\u0026rsquo;s read, that\u0026rsquo;s inevitable. The current LLM path can\u0026rsquo;t get to the physical world.\nHe went on to start AMI Labs in Paris, and raised $1.03B in two months at a $3.5B valuation.\nThis isn\u0026rsquo;t just LeCun\u0026rsquo;s faith. The VC market is putting real money behind a new path. Multiple players are now in the world model space in 2026:\nAMI Labs (LeCun) — $3.5B valuation World Labs (Fei-Fei Li) — $5B valuation Runway — $5.3B valuation The difference between world models and LLMs:\nLLMs understand text — what word should come after another word.\nWorld models learn the rules of the physical world — objects can\u0026rsquo;t pass through walls, balls fall when you throw them, water doesn\u0026rsquo;t flow uphill — letting AI understand the physical world the way humans do, through video.\nIn March 2026, LeCun\u0026rsquo;s team published a paper:\n\u0026ldquo;LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels\u0026rdquo; arxiv: 2603.19312\nThis paper solved a stubborn problem in world model research that had lingered for years — representation collapse.\nOld world models cheated. They figured out that mapping every input to the same vector trivially satisfied the training objective, so that\u0026rsquo;s what they did. Dogs, cars, people — all squashed into identical vectors. The model looked like it was learning, but it understood zero physical rules.\nLeWM solved this with an elegant mathematical regularizer (SIGReg):\n15M parameters — 100,000x smaller than GPT-4 Trains in a few hours on a single GPU 48x faster than mainstream baselines on robot planning tasks What does 48x mean in practice? Same planning task — old methods take 47 seconds, LeWM takes 0.98 seconds.\nThe paper\u0026rsquo;s contribution is pushing the \u0026ldquo;end-to-end JEPA world model\u0026rdquo; path from \u0026ldquo;theoretically possible\u0026rdquo; to \u0026ldquo;an engineering starting point.\u0026rdquo; It isn\u0026rsquo;t a production-grade model that runs on glasses, but it\u0026rsquo;s a key milestone — proof you can stably train a small world model without a pretrained large model.\nThis is the most promising path toward an on-device world model. Once the paradigm is right, the rest is engineering.\nThe trend is clear:\nThe AI of the future isn\u0026rsquo;t one all-purpose brain. It\u0026rsquo;s two brains with different jobs.\nBrain A (LLM) — handles language, creativity, reasoning, runs in the cloud Brain B (JEPA / world model) — handles physics, space, perception Your glasses, your BCI hat, every wearable on your body in the future — all of them run Brain B. It\u0026rsquo;s the most important engine for getting AI into human life without friction.\nApple swapped in a hardware CEO. Meta sold 7 million smart glasses in a year. Capital is putting tens of billions on world models. BrainCo\u0026rsquo;s BCI is already treating ADHD in hospitals.\nFour seemingly unrelated things, all hinting at where AI is going:\nPrompts are too hard to write → so AI has to anticipate → anticipating needs real-time user data → that data needs wearables to capture it automatically → all that data needs world models to digest it.\nAI\u0026rsquo;s task for the next decade is to walk out of the chat box and into everyone\u0026rsquo;s life.\n","permalink":"https://ryrenz.com/tech/world-model-replaces-chatbot/","summary":"Three years in, models are 100x more capable. Why is AI still just a chat box? The chat box is the Nokia of the AI era. From chatbots to agents to wearables plus world models — here\u0026rsquo;s the next decade of AI product evolution.","title":"Why We Need World Models to Replace Chatbots"},{"content":"Everyone around me who\u0026rsquo;s job hunting lately — engineers, designers, marketers, PMs, salespeople, even accountants and lawyers — keeps saying the same thing: sending resumes doesn\u0026rsquo;t work anymore.\nYou used to send 100 applications and get 20 or 30 interviews. Now you send 200 and barely hear back.\nYou didn\u0026rsquo;t get worse. The game changed.\nHow AI Killed Mass-Applying Both sides started using AI, and turned a system that once carried real signal into pure noise.\nOn the candidate side:\nEvery job seeker now has a full AI toolkit. AI rewrites your resume, writes your cover letter, auto-fills applications, auto-submits them. You can blast out dozens of applications in a single evening.\nA carefully written resume used to have texture. You could read between the lines — the way someone described their work revealed how they actually think, whether their experience was real.\nNow every resume looks the same after AI polishes it. Same action verbs (led, optimized, drove, spearheaded). Same quantified results (40% efficiency gain, 30% cost reduction, 3x user growth). Same keyword stuffing.\nLay ten resumes side by side and they read like ten clones.\nOn the employer side:\nHiringThing\u0026rsquo;s 2026 data shows:\nA single job posting gets 250 applications on average. Entry-level roles often top 400. The interview rate dropped from 15% in 2016 to 2-3% this year. Send 100 resumes, get two or three callbacks. That\u0026rsquo;s the average now. Companies can\u0026rsquo;t read them all. So they use AI too: keyword matching first, then AI scoring, then HR only reviews whatever the system spits out.\nThe whole game became candidates using AI to generate resumes, and companies using AI to filter resumes. The interview rate for mass applications sits at 1-2%. Basically a coin flip. You never find out why you didn\u0026rsquo;t make it.\nIt gets worse — many of the posted jobs aren\u0026rsquo;t even real.\nClarify Capital\u0026rsquo;s 2026 research found that roughly one in seven job postings is a \u0026ldquo;ghost job\u0026rdquo; with no intention to hire. For senior roles, it\u0026rsquo;s one in five. A Staffing Industry Analysts 2026 report revealed that 39% of hiring managers admitted their companies post fake job listings — some to test the market, some because the budget got frozen but the listing stayed up, some as fishing expeditions for \u0026ldquo;if someone great shows up.\u0026rdquo; Those dozens of resumes you fired off last night? Some of them were aimed at nothing.\nAdd in rampant AI-generated fake credentials and AI-assisted interview cheating, and the cost of screening unknown candidates keeps climbing.\nOnce trust breaks down, mass-applying loses all meaning.\nAI Skills Are Now Table Stakes Companies don\u0026rsquo;t want to train people, but they know exactly who they want: someone who understands AI and can use it to get things done.\nOver the past year, almost every job description quietly added some version of \u0026ldquo;familiar with AI tools / understanding of LLMs / AI project experience.\u0026rdquo;\nWhy? Because AI stopped being a tool for one profession. It\u0026rsquo;s a force multiplier for every profession. One person who knows how to use AI is worth three who don\u0026rsquo;t. Companies buy efficiency. Whoever delivers more, wins.\nAI is punishing exactly the \u0026ldquo;experienced veterans\u0026rdquo; who don\u0026rsquo;t know AI. You\u0026rsquo;ve been coasting on ten years of expertise, but the AI era runs on different rules.\nTwo things matter more than ever:\nWatch the market like a hawk. Don\u0026rsquo;t study skills that \u0026ldquo;sound safe.\u0026rdquo; Track job boards: which roles are expanding, which are shrinking, which pay well, which requirements are changing fastest. If you\u0026rsquo;re still a student, get internships early — see firsthand what industries actually need, who they\u0026rsquo;re hiring, and what they\u0026rsquo;re paying. Learn AI, and actually use it. Taking a prompt engineering course doesn\u0026rsquo;t count. You need to have solved real problems with AI: automated a workflow, boosted output, shipped something. Concrete results, not certificates. This isn\u0026rsquo;t optional. It\u0026rsquo;s the dividing line.\nHow Companies Respond: Filter Hard, Hire from Inside In the boom times, companies were willing to hire for potential and spend six months or a year getting someone up to speed.\nIn a downturn, that math doesn\u0026rsquo;t work. What AI can do, companies handle with AI. What AI can\u0026rsquo;t do, companies want someone who\u0026rsquo;s productive on day one. The cost of training shifted from the company to the candidate.\nAnd since the public talent pool is now polluted with AI noise, companies naturally retreat to internal referrals.\nToday, more than half of mid-to-senior roles never get posted publicly. They\u0026rsquo;re filled through:\nInternal employee referrals Recruiter and staffing agency networks Direct outreach through executives\u0026rsquo; personal connections This isn\u0026rsquo;t industry-specific. Good doctors, lawyers, salespeople, and engineers almost always get their next job through a former boss, former client, former colleague, or alumni network. They never submit a resume through a portal. The front door where you\u0026rsquo;re submitting applications isn\u0026rsquo;t where the good opportunities flow.\nCompanies pulling back to referrals, personal networks, and headhunters are all doing the same thing: reducing the trust cost of \u0026ldquo;hiring the wrong person.\u0026rdquo;\nThe new job-hunting logic in the AI era is a trust competition. Whoever can establish credibility at the lowest cost with an employer who\u0026rsquo;s never met them wins.\nThat\u0026rsquo;s the key to everything below.\nHow Regular People Build Trust 1. Make Yourself Memorable, Stay Market-Aware Mass-applying is dead, but your personal profile still matters. Keep your job platform pages and resume sharp.\nWhen you reach out to an HR person, they\u0026rsquo;ll click your profile and make a 30-second judgment call. In-person meetings still need a resume as a first reference. Your resume and profile need a hook. They can\u0026rsquo;t be a dry list of job titles and dates anymore.\nDon\u0026rsquo;t write \u0026ldquo;X years of experience, passionate about Y, skilled in Z.\u0026rdquo; That\u0026rsquo;s what everyone writes.\nYour first line should slap the reader with the single most impressive thing you\u0026rsquo;ve done. Make it impossible to forget.\nHR reads hundreds of resumes a day, 30 seconds each. They only remember the one that opened with a story. Everything after the first line barely gets read.\nThe real value of job platforms isn\u0026rsquo;t parking your profile and waiting. They give you a direct line to HR managers, hiring leads, and recruiters — and that\u0026rsquo;s an order of magnitude more effective than blind applications:\nReach out directly to HR, hiring managers, and team leads. Find the actual person on LinkedIn, BOSS Zhipin, or Maimai. One proactive message gets several times the response rate of a blind application. Build a relationship with one or two recruiters. Signing up isn\u0026rsquo;t enough. Stay in touch with a recruiter who knows your industry at least every six months. They know who\u0026rsquo;s hiring, what they\u0026rsquo;re paying, and what\u0026rsquo;s non-negotiable — information you can\u0026rsquo;t get on your own. Interview regularly, even when you\u0026rsquo;re not looking. Pick one or two solid opportunities every six months and go through the process seriously. The goal isn\u0026rsquo;t the offer — it\u0026rsquo;s calibrating your market value and updating your read on the industry. By the time your income stagnates or your role gets replaced and you realize you\u0026rsquo;ve fallen behind, it\u0026rsquo;s already too late. 2. Real Connections Who\u0026rsquo;ll Vouch for You Companies won\u0026rsquo;t trust a stranger based on a resume. But when a senior employee says \u0026ldquo;this is my former colleague, they\u0026rsquo;re solid\u0026rdquo; — trust is instant.\nThe network that matters here isn\u0026rsquo;t your friend count on LinkedIn or WeChat. It\u0026rsquo;s people who know your work, know your abilities, and are willing to speak up for you.\nYou don\u0026rsquo;t need many. You need real ones.\nHow:\nStay in touch with former colleagues, bosses, clients, and partners. Nothing forced — \u0026ldquo;saw your latest project, looks interesting\u0026rdquo; is enough. Show up at industry events in person. Knowing 100 people online matters less than sharing one meal with someone offline. People you\u0026rsquo;ve actually met and talked to are the ones who think of you first when a good opportunity comes up. 3. Build in Public — Let Work Find You Running a public account over the long term is the lever that compounds your portfolio and your network at the same time.\nThis isn\u0026rsquo;t about becoming an influencer or selling courses. It\u0026rsquo;s about maintaining a genuine, industry-relevant output channel.\nTwitter, a blog, YouTube, a newsletter — the format doesn\u0026rsquo;t matter. What matters is consistently making visible what you see, what you think, and what you build.\nIt scales your weak ties and drops your trust cost to near zero. Your followers may not know you personally, but they know your work. When their company needs to hire, the first person they think of isn\u0026rsquo;t a stranger on a resume — it\u0026rsquo;s you, the person they\u0026rsquo;ve been following for two years. You go from unknown to \u0026ldquo;someone I\u0026rsquo;ve known for a long time.\u0026rdquo;\nIt lets opportunities find you. A public account puts you in front of decision-makers across industries. While you sleep, the algorithm keeps pushing your content to the next potential employer or client.\nIn tech, people who consistently write, contribute to open source, and share their work publicly never worry about finding their next gig. In creative and consulting fields — designers, consultants, teachers — client acquisition has already shifted from agencies to \u0026ldquo;content attracts, clients reach out.\u0026rdquo; Even in blue-collar trades — nannies, carpenters, personal trainers — social platforms are reshaping how people get clients, cutting out traditional intermediaries. Customers come to them directly. This works from white-collar to blue-collar. It\u0026rsquo;s becoming a societal norm.\nKevin Kelly proposed the \u0026ldquo;1,000 True Fans\u0026rdquo; theory back in 2008. You don\u0026rsquo;t need to go viral. You don\u0026rsquo;t need a million followers. Out of 1,000 people who genuinely follow your work, if just 10 have influence, resources, or decision-making power, you\u0026rsquo;ll be too busy to worry about your next job.\nThe old job-hunting logic:\nWrite a good resume → Apply to many jobs → Pass interviews → Get an offer\nThat\u0026rsquo;s a push. You push yourself into the market.\nThe new logic:\nStay consistently visible → Accumulate verifiable work → Maintain real connections → Opportunities find you → You pick among them\nThat\u0026rsquo;s a pull. You let the market come to you.\nThe biggest difference isn\u0026rsquo;t effort. It\u0026rsquo;s time horizon.\nPushing is one-shot. You polish your resume when you\u0026rsquo;re unemployed, blast applications when you need money, and when the moment passes, it\u0026rsquo;s gone.\nPulling is cumulative. Every profile you maintain, every piece you publish, every connection you nurture, every follower you earn — they all slowly build a mountain. You\u0026rsquo;re always ready. You never miss an opportunity.\nThe Bottom Line Mass-applying is dead. And with it goes the era of handing your fate to strangers for judgment.\nIn the new era, whoever\u0026rsquo;s achievements are most visible, most credible, and most verifiable will stand the steadiest and move the most freely in this transparent market.\nThis has nothing to do with your industry, your title, or whether you work at a big company or a small one.\nIt\u0026rsquo;s the baseline everyone who still works for a living needs to understand in the age of AI.\nThe spray-and-pray approach doesn\u0026rsquo;t work anymore.\n","permalink":"https://ryrenz.com/career/resume-is-dead-ai-job-hunting/","summary":"Sending 100 resumes used to get you 20-30 interviews. Now 200 gets you almost nothing. You didn\u0026rsquo;t get worse — the rules changed. Mass-applying is dead. The new job hunt is a trust competition.","title":"The Job Hunt Survival Guide for the AI Era"},{"content":"A former colleague of mine ran a side project as an indie developer.\nTwo years in, he had about 3,000 users and had never charged a cent. Every time I asked why, he gave the same answer: \u0026ldquo;The product isn\u0026rsquo;t good enough yet. I need to polish it more.\u0026rdquo;\nThen he got laid off. Out of options, he finally worked up the nerve to add a $9/month subscription.\nHalf his users left. The other half paid up and immediately started flooding him with feature requests.\nHe received more actionable feedback in the first month than in the previous two years combined.\nWhat he lost wasn\u0026rsquo;t two years of revenue. It was two years of chances to make the product better.\nHe thought not charging was protecting his users. He was protecting himself — from having to face a terrifying question: is this thing I built actually worth money?\nThat question is scary. So he kept it at bay with \u0026ldquo;the product isn\u0026rsquo;t good enough yet.\u0026rdquo;\nSounds like humility. It\u0026rsquo;s not. It\u0026rsquo;s a wall.\nIn business, not making money is the original sin.\nFree Users Give You Fake Feedback How many free apps are on your phone right now that you\u0026rsquo;ve never opened? How many articles have you bookmarked and never read?\nFree users have sky-high tolerance. Zero cost means zero expectations.\nSomething doesn\u0026rsquo;t work? Close it and move on. No money lost.\nThey won\u0026rsquo;t email you about what\u0026rsquo;s broken. They won\u0026rsquo;t post about missing features. They\u0026rsquo;ll just vanish. Quietly.\nYou watch your DAU slide downward and you have no idea why. Because nobody tells you. Free users don\u0026rsquo;t owe you an explanation.\nPaying users are different. Payment creates commitment.\nThink about a course or a tool you actually paid for. Even if it\u0026rsquo;s mediocre, you push through, right? Because you\u0026rsquo;d feel stupid wasting the money.\nWhen users pay, they spend time actually using the product — because not using it feels like a loss. Using it creates experience. Experience creates feedback. Feedback drives iteration.\nMoney in means expectations up. Expectations create standards. Fall short and they\u0026rsquo;ll complain, leave bad reviews, and write angry emails.\nThat might feel brutal. But it\u0026rsquo;s the best thing your product can get.\nOnly paying customers give you high engagement, deep usage, and honest feedback. Every complaint is a real need. Every refund request is a product trial. Stack them up and you\u0026rsquo;ve got your roadmap.\n\u0026ldquo;I\u0026rsquo;ll Charge When It\u0026rsquo;s Ready\u0026rdquo; — That Day Never Comes I\u0026rsquo;ve heard this one a thousand times.\n\u0026ldquo;Just need to polish it a bit more.\u0026rdquo; \u0026ldquo;Once a few more features are in.\u0026rdquo; \u0026ldquo;When I have more users.\u0026rdquo;\nRephrase it and you\u0026rsquo;ll recognize the pattern:\n\u0026ldquo;Things will be better once I get into college.\u0026rdquo;\n\u0026ldquo;Things will be better once I get married.\u0026rdquo;\n\u0026ldquo;Things will be better once I buy a house.\u0026rdquo;\n\u0026ldquo;Later\u0026rdquo; never arrives. \u0026ldquo;Later\u0026rdquo; isn\u0026rsquo;t a point in time. It\u0026rsquo;s a coping mechanism. You feel like you\u0026rsquo;re moving forward, but you\u0026rsquo;re just waiting.\nWithout paying users, your iteration rhythm is \u0026ldquo;whenever I get around to it.\u0026rdquo; With paying users, it\u0026rsquo;s \u0026ldquo;fix it fast or they\u0026rsquo;ll cancel.\u0026rdquo; After a year, products built under those two rhythms are ten times apart.\nI\u0026rsquo;ve watched too many solo founders get stuck in the same loop:\nNo charge → no revenue → can\u0026rsquo;t go full-time → work on it part-time → slow progress → product stays half-baked → even less confidence to charge.\nThe ones who charge from day one?\nRevenue → can invest time → fast iteration → product gets good → users pay more → more revenue.\nOnce that flywheel starts spinning, the gap compounds.\nYou\u0026rsquo;re Not Worried About Quality — You\u0026rsquo;re Afraid of Being Judged This is the part nobody wants to hear.\n\u0026ldquo;The product isn\u0026rsquo;t good enough\u0026rdquo; is the excuse. The real reason? Fear.\nFear that you put a price tag on it and nobody buys. Fear that someone uses it and says, \u0026ldquo;That\u0026rsquo;s it? And you\u0026rsquo;re charging money?\u0026rdquo;\nFear that the market gives you a real answer: not worth it.\nWithout a price tag, you never have to face the market. You can live inside the safety of \u0026ldquo;still in development.\u0026rdquo; Schrödinger\u0026rsquo;s product — nobody knows if it\u0026rsquo;s good or bad.\nPricing is an act of exposure. You put your work out there, stick a number on it, and wait for the verdict.\nMost people can\u0026rsquo;t handle that.\nNot charging isn\u0026rsquo;t humility. It\u0026rsquo;s fear dressed up as a virtue.\nThere\u0026rsquo;s another type: people who aren\u0026rsquo;t afraid the product is bad, but who think selling is beneath them.\nThere\u0026rsquo;s a vibe online that selling stuff is tacky. That the moment you talk money, everything valuable you\u0026rsquo;ve said gets cheapened. Like charging turns you from a \u0026ldquo;creator\u0026rdquo; into a \u0026ldquo;salesman,\u0026rdquo; from \u0026ldquo;sharing knowledge\u0026rdquo; into running a scam.\nAnyone who\u0026rsquo;s held a job or started a company knows this thinking is childish. Building a product isn\u0026rsquo;t charity.\nMarketing is always the most important piece of a business model. A great product that doesn\u0026rsquo;t charge and nobody knows about has zero value.\nYou built something good. Marketing puts it in front of users who pay real money for quality, and you get capital to keep improving. That\u0026rsquo;s a win for both sides. No guilt needed.\nPeople who think selling is tacky are confusing \u0026ldquo;transaction\u0026rdquo; with \u0026ldquo;fraud.\u0026rdquo; You\u0026rsquo;re not tricking anyone out of their money. You\u0026rsquo;re telling the market: I stand behind what I built.\nIf you don\u0026rsquo;t have the guts to face the market, how will you ever have the stamina to keep going and actually build something great?\nPrice Is a Brand Signal Why do Apple and Hermès never compete on price?\nBecause pricing is part of how a brand occupies your mind. A $999 phone and a $199 phone might differ by 30% in hardware, but the perceived value gap is five to one. Price shapes expectations. Expectations shape experience. Experience shapes the brand.\nPrice it at $6/month and users think \u0026ldquo;cheap little tool, good enough.\u0026rdquo; Price it at $60/month and users think \u0026ldquo;this is a professional tool, I\u0026rsquo;d better learn how to use it properly.\u0026rdquo;\nSame product. Different price. Completely different user behavior.\nAnd there\u0026rsquo;s something even more important: the snowball effect of high pricing.\nCold-starting a product is the hardest part. If you\u0026rsquo;re internet-famous, maybe it\u0026rsquo;s a bit easier. But most founders have no reputation, no case studies, nobody vouching for them.\nAt that stage, the only thing you can do is serve every single user well — make the first few batches feel like they got their money\u0026rsquo;s worth.\nOnce the product survives the cold start, it starts telling its own story.\nEvery time someone throws out the classic \u0026ldquo;why should I pay, and why is it so expensive?\u0026rdquo;\nAll you need to say is: \u0026ldquo;I\u0026rsquo;ve charged this price since day one, and people keep buying.\u0026rdquo;\nThen word of mouth takes over. Users refer users. Trust passes from person to person. You stop needing to explain why it\u0026rsquo;s expensive, because the market has already validated you.\nPeople start saying, \u0026ldquo;I found this product with a great reputation that works really well. The only downside is the price, but that might be my problem, not theirs.\u0026rdquo;\nCustomer Acquisition Costs Are High — Don\u0026rsquo;t Undercut Yourself Acquiring a customer costs the same whether you charge $5 or $500.\nAt $5/month, your customer acquisition cost might be $20. You need users to stick around four months just to break even.\nAt $25/month, you break even in month one. Every month after that is profit you can pour back into the product.\nYou think you\u0026rsquo;re being generous to users. You\u0026rsquo;re actually bleeding your own cash flow dry. Without cash flow, what do you use to improve the product? To run ads? To survive until the product is actually good?\nPricing was never just \u0026ldquo;how much is my product worth?\u0026rdquo;\nIt\u0026rsquo;s \u0026ldquo;can my business stay alive?\u0026rdquo;\nAnd what keeps you alive long enough to reach those users is a price that covers the cost of finding them.\nYou\u0026rsquo;ve Got the Cause and Effect Backwards Back to the opening story.\nHe now makes over $10,000 a month. His user base is half what it was during the free era. But his paying users\u0026rsquo; retention rate is four times higher.\nThe product? After dozens of iterations, it\u0026rsquo;s unrecognizable from where it started.\nBecause every paying user was telling him with real money: this part is good, that part sucks, I\u0026rsquo;d pay more for this feature, I don\u0026rsquo;t care if you cut that one.\nThat kind of feedback was nearly impossible to get when it was free.\nSo the cause and effect aren\u0026rsquo;t what you think.\nIt\u0026rsquo;s not \u0026ldquo;charge once the product is good.\u0026rdquo;\nIt\u0026rsquo;s \u0026ldquo;charge, and the product will get good.\u0026rdquo;\nDon\u0026rsquo;t charge, and your product is a tree without roots, a river without a source. Failure is just a matter of time.\n","permalink":"https://ryrenz.com/startup/why-free-products-fail/","summary":"A former colleague built a free product for two years — 3,000 users, zero revenue. After getting laid off, he finally added a $9/month subscription. He got more useful feedback in the first month than in the previous two years combined. The product didn\u0026rsquo;t need to be good before he charged. Charging is what made it good.","title":"Why 'It's Not Good Enough to Charge For' Is a Death Sentence for Your Business"},{"content":"Meta sold 7 million smart glasses last year.\nBut you probably don\u0026rsquo;t see many people wearing them. Most reactions are still:\n\u0026ldquo;Isn\u0026rsquo;t that just a camera in sunglasses?\u0026rdquo; \u0026ldquo;A headphone in a different shape?\u0026rdquo; \u0026ldquo;A toy.\u0026rdquo;\nI thought the same. Until I saw what happened when AI was plugged into these glasses.\nAI + Glasses Are Already Changing Real Lives No concepts. Three things already happening.\nNavigation for the blind.\nA blind person wearing Meta glasses says \u0026ldquo;Hey Meta, what\u0026rsquo;s ahead?\u0026rdquo; The AI describes streets, signs, and obstacles through the camera in real time. With Be My Eyes integrated, one sentence connects you to a human volunteer who \u0026ldquo;sees for you\u0026rdquo; through the glasses camera, guiding you by voice.\nNew York State has started giving these glasses to blind students for free.\nOne blind user said something that stuck with me:\n\u0026ldquo;This technology makes me grateful I went blind in this era, not earlier.\u0026rdquo;\nAt CES 2026, dotLumen went further — using autonomous driving tech to build \u0026ldquo;self-navigating glasses for the blind\u0026rdquo;: 6 cameras scan the environment, AI plans paths in real time, haptic feedback guides direction. It won a CES 2026 Innovation Award.\nReal-time translation.\nYou\u0026rsquo;re abroad, looking at a foreign-language menu. The glasses overlay the translation directly in your field of view. No pulling out your phone, no opening an app, no pointing a camera — you just look. The Qwen S1 glasses shown at MWC 2026 already display real-time translated subtitles on the lens.\nHands-free everything.\nPeople are already cooking with AI glasses — asking \u0026ldquo;is this steak medium-rare, should I flip it?\u0026rdquo; and getting coached step by step. Tech editors at press events run Slack, Chrome, and email simultaneously through AR glasses, both hands free.\nThese aren\u0026rsquo;t concept videos or keynote demos. Real people are using them in daily life.\nAnd this is just the beginning.\nEvery Major Player Is Betting at Once — Apple Included If Meta proved \u0026ldquo;smart glasses can actually sell,\u0026rdquo; what comes next will blow the market wide open:\n2026 Q1: Samsung reveals its first AI glasses, expected to ship this year 2026: Google launches two models with Gemini AI, partnering with Warby Parker and Xreal 2026 H2: Snap\u0026rsquo;s consumer version hits the market 2027: Apple Glasses reportedly launching (per Bloomberg) — camera + upgraded Siri + AI assistant, estimated $499–$799 2027: Nothing enters the space Apple, Google, and Samsung all betting on the same category at the same time. The last time that happened was smartphones.\nThere\u0026rsquo;s a fundamental difference between phones and glasses: smart glasses are AI-native.\nSmartphones waited a decade for AI to become useful. Smart glasses are AI-native from day one — they can see, hear, and stream everything you experience to AI in real time. That means they\u0026rsquo;ll evolve much faster than smartphones did.\nThe numbers: the global AI smart glasses market was roughly $1.2 billion in 2025, projected to jump to $5.6 billion in 2026 — nearly 5x in one year.\nMultiple analyst reports converge on the same figure: \u0026ldquo;By 2030, global AI smart glasses shipments could reach 80 million units.\u0026rdquo;\nBut that\u0026rsquo;s just hardware.\nWellsenn and others predict that once these glasses scale up with AI interaction, AR apps, content ecosystems, and cloud services — if they replace even 10–20% of smartphone usage time — the full smart glasses ecosystem could reach $120–240 billion by 2030.\nZuckerberg said on Meta\u0026rsquo;s Q1 2026 earnings call: \u0026ldquo;Just as smartphones replaced flip phones, it\u0026rsquo;s hard to imagine most people wearing glasses that aren\u0026rsquo;t AI glasses in a few years.\u0026rdquo;\nHe\u0026rsquo;s not bluffing — Meta glasses tripled in sales over the past year. He called them \u0026ldquo;one of the fastest-growing consumer electronics products in history.\u0026rdquo;\nThis isn\u0026rsquo;t \u0026ldquo;might happen.\u0026rdquo; It\u0026rsquo;s happening.\nPrices Will Drop — That\u0026rsquo;s How Manufacturing Works Meta Ray-Ban starts at $299. Not the tens of thousands some people assume. By manufacturing norms, sub-$100 products will likely appear within three to five years.\nRemember smartphone history: the first iPhone in 2007 cost $499. Many said it wasn\u0026rsquo;t worth it.\nSoon after, budget smartphones filled every corner of the world.\nSmart glasses will follow the exact same path. As supply chains mature, volumes scale, and optical module and chip costs get amortized, prices will reach everyone. Not a guess — this is how manufacturing has always worked.\nI\u0026rsquo;m convinced: smart glasses are the next mobile platform.\nThe scale of this shift is comparable to the revolution from feature phones to smartphones.\nFeature phones to smartphones wasn\u0026rsquo;t \u0026ldquo;phones got better\u0026rdquo; — an entire mobile internet ecosystem was born from zero.\nApp Store, WeChat, mobile payments, ride-hailing, short-form video\u0026hellip; none of these were imaginable in the feature phone era. Smart glasses will birth an entirely new set of use cases and business models we can\u0026rsquo;t picture today.\nThe only difference: last time you were a user. This time you can choose to be a builder.\nThis Isn\u0026rsquo;t \u0026ldquo;a Smaller Phone\u0026rdquo; — an Interaction Revolution Is Underway Most people imagine smart glasses as \u0026ldquo;a phone screen strapped to your face.\u0026rdquo;\nWrong.\nSmart glasses are redefining how humans interact with machines.\nFour interaction modes that don\u0026rsquo;t exist on phones:\nGaze + voice.\nYou look at a restaurant and say \u0026ldquo;book tonight, 7 PM, two people.\u0026rdquo; The glasses know what you\u0026rsquo;re looking at and what you\u0026rsquo;re saying. The action just happens. No app, no search bar, no typing.\nA 2026 research paper named this pattern \u0026ldquo;Gazeify Then Voiceify\u0026rdquo; — eyes select the target, voice issues the command. Phones can\u0026rsquo;t do this.\nEMG wristband gestures.\nMeta\u0026rsquo;s Neural Band captures micro-movements through electrical signals in your wrist muscles: pinch to confirm, twist to scroll, double-tap to go back. No waving in the air — you can operate with your hand in your pocket.\nThis interaction system just won a 2026 UX Design Award. Judges called it \u0026ldquo;a model for the post-smartphone era.\u0026rdquo;\nContext-aware interfaces.\nInformation appears when you need it and disappears when you don\u0026rsquo;t. At an intersection, navigation arrows overlay the actual road; once you pass, they vanish. Notifications show for 3 seconds, one action to handle, never interrupting what you\u0026rsquo;re doing.\nA phone\u0026rsquo;s logic is \u0026ldquo;you go find information.\u0026rdquo; Glasses flip it: \u0026ldquo;information finds you, and it knows when to arrive and when to leave.\u0026rdquo;\nProactive AI agents.\nNot \u0026ldquo;you ask, it answers.\u0026rdquo; The AI continuously understands your context and offers help at the right moment.\nA blind Canadian wearing Meta glasses in the kitchen asks \u0026ldquo;what kind of noodles are these?\u0026rdquo; The AI glances and tells him pasta vs. rice noodles. Before going out: \u0026ldquo;does my outfit match?\u0026rdquo; The AI confirms whether the colors work.\nAnother blind person at a crosswalk: \u0026ldquo;is it a green light?\u0026rdquo; Instant answer. No phone, no asking strangers.\nSmartphones redefined \u0026ldquo;touch.\u0026rdquo; Smart glasses are redefining \u0026ldquo;seeing\u0026rdquo; and \u0026ldquo;speaking.\u0026rdquo;\nThis isn\u0026rsquo;t a product iteration. It\u0026rsquo;s a paradigm shift in how humans and machines interact.\nFor UI/UX Designers and Frontend Devs: This Might Be Your Biggest Career Opportunity You\u0026rsquo;ve heard it the past six months: \u0026ldquo;AI can generate interfaces now. Frontend is dead. Designers are dead.\u0026rdquo;\nIf you think that, you\u0026rsquo;re making the same mistake Nokia made in 2007 — staring at whether the old battlefield survives while the new one has already opened.\nThe core fact: smart glasses have no established design paradigm. Everything starts from zero.\nPhones have 20 years of interaction language — buttons, swipes, lists, tab bars, pull-to-refresh. You can draw these in your sleep.\nBut on glasses?\nHow do you lay out spatial UI? Where in the field of view can information float without causing nausea? How do you design feedback for gaze interaction — when someone looks at a button, how do you signal \u0026ldquo;I know you\u0026rsquo;re looking at me\u0026rdquo;? What\u0026rsquo;s the information density limit for a 3-second glanceable card? What\u0026rsquo;s the error tolerance for EMG gesture inputs? Should the layout differ between walking and sitting?\nNone of these have standard answers today. No Material Design, no Human Interface Guidelines, no off-the-shelf component libraries.\nWhat does that mean? Whoever defines these standards first becomes this era\u0026rsquo;s design authority.\nDon Norman (author of The Design of Everyday Things) put it bluntly: standardizing gestures and voice commands is incredibly complex, which means more UX research and design work is needed, not less.\nNielsen Norman Group\u0026rsquo;s 2025 conclusion echoed this: AI tools are \u0026ldquo;useful assistants but not replacements.\u0026rdquo;\nLook at who\u0026rsquo;s hiring:\nApple — AR/VR Software Engineer, Vision Products Software roles Google — UX Engineer (Spatial Experiences), UX Researcher (3D Human Modeling) Meta — Wearables Design Team actively expanding: creative technology, product design, AR interaction design Specialized AR/VR job boards like arvrjobs.dev — a growing number of active listings Every platform migration — PC to phone, phone to glasses — is the biggest window for designers and developers.\nOn the old platform you\u0026rsquo;re a cog. On the new platform you\u0026rsquo;re a pioneer. The window won\u0026rsquo;t wait for you to feel ready.\nThree Paths for Everyone Maybe you\u0026rsquo;re not a programmer or a designer, and you\u0026rsquo;re thinking \u0026ldquo;what does any of this have to do with me?\u0026rdquo;\nA lot.\nPath one: application-layer startups.\nThe people who made the most money in the smartphone era didn\u0026rsquo;t build phones. They built apps.\nZhang Yiming wasn\u0026rsquo;t a hardware engineer. Neither was Cheng Wei. TikTok, Meituan, Didi — their founders saw the application-layer opportunity after the smartphone platform opened up, and they went in. Their starting capital and technical bar were far lower than building a phone.\nSmart glasses are the same. When the platform scales and users arrive, massive gaps will appear at the application layer — AR tours, immersive education, first-person livestream commerce, spatial ads, AR games. Almost zero competition today because the platform isn\u0026rsquo;t mature yet. The people who start researching and preparing now will be the first to capture the value.\nPath two: content creation.\nEvery platform migration births entirely new content formats. PC era had blogs and forums. Smartphone era had short video and livestreaming. Smart glasses era?\nFirst-person immersive content, AR-overlay interactivity, spatial storytelling — there\u0026rsquo;s no \u0026ldquo;TikTok\u0026rdquo; for these yet, but there will be.\nThe earliest short-video creators weren\u0026rsquo;t professionals. They just started experimenting before everyone else.\nPath three: cognitive edge.\nThis path gets underestimated the most, but it\u0026rsquo;s arguably the most valuable.\nIf you understood in 2007 that \u0026ldquo;smartphones will change everything,\u0026rdquo; you didn\u0026rsquo;t need to build a phone. You\u0026rsquo;d have opened an e-commerce store earlier, started a WeChat account earlier, shot short videos earlier, launched cross-border commerce earlier. Every step two years ahead of others. \u0026ldquo;Two years early\u0026rdquo; is the biggest competitive advantage an ordinary person can have.\nThe value of this article isn\u0026rsquo;t \u0026ldquo;telling you what to buy now.\u0026rdquo; It\u0026rsquo;s helping you see a high-certainty trend in advance. When it truly explodes, you won\u0026rsquo;t be starting from zero — you\u0026rsquo;ll have been thinking about it for two years, and you\u0026rsquo;ll know where the opportunities are.\nConclusion 2025: Meta sold 7 million smart glasses in one year. 2026: Apple, Google, and Samsung enter simultaneously. AI is native from day one. A new interaction paradigm is being defined. Massive job openings are waiting to be filled. The application layer and content ecosystem are nearly blank.\nThese aren\u0026rsquo;t predictions. They\u0026rsquo;ve already happened.\nTech people see an uncharted design frontier — whoever defines the rules first becomes the authority. Creators see the next \u0026ldquo;early TikTok\u0026rdquo; traffic window. Everyone else sees a high-certainty trend they can position for two years early.\nBut most people still see \u0026ldquo;just a camera in sunglasses.\u0026rdquo;\nEvery tech revolution offers the most opportunity when most people are still laughing.\nYou don\u0026rsquo;t need to wait until Apple Glasses are everywhere to start preparing.\n","permalink":"https://ryrenz.com/industry/smart-glasses-next-trillion-market/","summary":"Meta sold 7 million smart glasses last year. In 2026, Apple, Google, and Samsung are all entering the space. AI is native from day one, and an entirely new interaction paradigm is being defined. This isn\u0026rsquo;t a prediction — it\u0026rsquo;s already happening.","title":"Smart Glasses: The Next Trillion-Dollar Market You Can't Ignore"},{"content":"After a year with Claude Code, I\u0026rsquo;ve distilled 7 principles from the last six months of daily use. Each one came from real pain during real development.\nThe most common beginner mistake isn\u0026rsquo;t lack of ability — it\u0026rsquo;s treating Claude Code like a smarter ChatGPT. Used right, it\u0026rsquo;s an engineer you can train. Used wrong, it\u0026rsquo;s a hallucinating, lying creature that \u0026ldquo;looks like it worked but actually didn\u0026rsquo;t change anything,\u0026rdquo; and won\u0026rsquo;t lift a finger unless you tell it to.\nGet these 7 things right and AI becomes your best partner.\n1. Don\u0026rsquo;t Treat It Like a Chat Box — Treat It Like Your Employee Imagine you\u0026rsquo;re a team lead assigning a task to a new hire. You say \u0026ldquo;add a delete feature\u0026rdquo; and walk away. They don\u0026rsquo;t know what you\u0026rsquo;re deleting, where the page redirects after, whether it\u0026rsquo;s soft delete, whether there\u0026rsquo;s a confirmation dialog, whether to log the action, which users have permission. They\u0026rsquo;ll guess. The result probably isn\u0026rsquo;t what you wanted.\nThat\u0026rsquo;s not the employee being incompetent. You didn\u0026rsquo;t explain the task.\nAI is exactly the same. You type \u0026ldquo;add a delete feature\u0026rdquo; and it doesn\u0026rsquo;t have the product picture in your head, doesn\u0026rsquo;t have your week of conversations with the PM, doesn\u0026rsquo;t know how \u0026ldquo;delete\u0026rdquo; works in other modules, doesn\u0026rsquo;t know your acceptance criteria. It can only guess.\nTreat it like an actual new colleague. Every task needs four things:\nObjective — what the user sees and what the system looks like when it\u0026rsquo;s done Detailed requirements — what to delete, how, where to redirect, soft delete or not, permission checks, confirmation dialog Current project state — tech stack, where related modules live, how similar features were implemented before, any patterns to follow Acceptance criteria — which tests must pass, which edge cases to cover, don\u0026rsquo;t stop until all tests pass The most common beginner mistake is skipping the five minutes of explanation, then spending two hours correcting the AI\u0026rsquo;s wrong direction. Those five minutes are the highest-ROI five minutes of the entire task.\nThe test: is your keyboard getting busier or quieter? If you\u0026rsquo;re using it right, you should be typing less over time.\n2. Use CLAUDE.md — It\u0026rsquo;s the AI\u0026rsquo;s Engineering Red Line CLAUDE.md is the baseline Claude Code must always follow. It\u0026rsquo;s loaded into every conversation, so it needs to be concise while being clear.\n\u0026ldquo;Loaded into every conversation\u0026rdquo; — it takes up context window. The longer it is, the higher the cost per turn. If you can say it in one sentence, don\u0026rsquo;t use three.\n\u0026ldquo;Must always follow\u0026rdquo; — it\u0026rsquo;s not a suggestion, it\u0026rsquo;s a hard constraint. Only write things you truly want followed every single time. Don\u0026rsquo;t use \u0026ldquo;preferably,\u0026rdquo; \u0026ldquo;try to,\u0026rdquo; or \u0026ldquo;sometimes.\u0026rdquo;\nWhat belongs in CLAUDE.md:\nEngineering red lines (comment language, commit conventions, search before guessing) File output rules (where reports go, what not to generate, where tests live) Response format constraints (what language to use, required output structure) Hard no\u0026rsquo;s (no --no-verify, no force-committing gitignored files) What doesn\u0026rsquo;t belong:\nCode style details — that\u0026rsquo;s the linter\u0026rsquo;s job Architecture background and business logic — that\u0026rsquo;s the Design Doc\u0026rsquo;s job Things you want it to do \u0026ldquo;sometimes\u0026rdquo; — it won\u0026rsquo;t execute reliably, and it dilutes the real red lines Long explanations and examples — burns tokens, imperative bullet points are enough You also need to separate global vs. project-level CLAUDE.md:\nGlobal CLAUDE.md (~/.claude/CLAUDE.md) covers cross-project personal habits: comment language, commit language, always search when uncertain, response format, memory system paths. These are \u0026ldquo;you\u0026rdquo; rules that apply everywhere.\nProject-level CLAUDE.md (repo root) covers only what\u0026rsquo;s unique to this project: tech stack, how to run tests, deployment process, forbidden directories, Design Doc path. Rule of thumb: if the rule still holds when moved to another project, it doesn\u0026rsquo;t belong at the project level.\nOne practical tip: red lines are earned from pain, not planned in advance. Your first CLAUDE.md for a new project can be short. Every time the AI does something that genuinely angers you, add a rule. After a month you\u0026rsquo;ll have a dense, high-signal document that\u0026rsquo;s uniquely yours.\n3. Package Frequent Tasks into Skills — Stop Re-explaining Many people have heard of Skills but haven\u0026rsquo;t started using them. Skills are the single biggest productivity multiplier. They\u0026rsquo;re essentially \u0026ldquo;prompt templates with an entry point and a tool allowlist.\u0026rdquo;\nAn example I\u0026rsquo;ve hit countless times: writing Design Docs.\nEvery new project I had to explain: write to \u0026lt;memory_root\u0026gt;/docs/design.md, choose Lite or Full template based on project size, architecture diagram must start from the user entry point not just backend internals, must include Overview / Ultimate Vision / Tech Stack / Architecture Diagram / Feature Status / Current Milestone / Key Decisions sections, unfilled sections get TBD not made-up content, if design.md already exists don\u0026rsquo;t overwrite it and run the update flow instead\u0026hellip;\nI got tired of repeating this. Every time I\u0026rsquo;d waste tokens and inevitably miss a rule, and every project\u0026rsquo;s Design Doc looked different.\nI packaged it into the /init-design skill: fixed template, fixed path, fixed rules, fixed logic. Now starting a new project is one command: /init-design this project is XXX. Ten seconds later, I have a spec-compliant Design Doc.\nAnti-pattern: repeating a dozen standards every new project, hoping the AI doesn\u0026rsquo;t miss any this time. Pattern: encode the standards into a skill, type /init-design and done.\nI now use 20+ custom skills daily: /plan, /build, /debug, /codex-review, /init-design, /update-design, /commit, /contribute, /loop, /schedule\u0026hellip; Each one replaces a workflow you\u0026rsquo;d otherwise type out repeatedly.\nThe test for whether something should be a skill: are you annoyed at re-typing a prompt you\u0026rsquo;ve written before? If yes, package it immediately.\n4. Build a Memory System — Let Context Survive Across Conversations Claude Code\u0026rsquo;s context window, no matter how large, resets to zero every session. What actually makes an AI assistant \u0026ldquo;know you\u0026rdquo; isn\u0026rsquo;t prompt engineering — it\u0026rsquo;s a stable, readable, writable, cross-session memory system.\nA simple memory system at ~/memory/ looks like this:\nUSER.md — who you are, your profile NOW.md — what you\u0026rsquo;re currently working on, updated after every session docs/INDEX.md — project documentation map daily-logs/ — last 14 days of conversation logs lessons/ — experience distilled from conversations, auto-extracted by hooks projects/ — each project\u0026rsquo;s design.md / plan/ lessons/ captures every pitfall from your conversations and records them as reusable experience files, ensuring the AI doesn\u0026rsquo;t repeat the same mistake twice.\nFirst thing after every new conversation or Compact: read NOW.md to restore context.\nWith hooks maintaining it automatically, you\u0026rsquo;ll find the AI remembers what you did last week better than you do.\n5. Automate with Hooks — Don\u0026rsquo;t Rely on Prompt Reminders \u0026ldquo;Always do X\u0026rdquo; written in CLAUDE.md isn\u0026rsquo;t reliable — the model forgets, cuts corners, or gets overridden by context.\nHooks are shell scripts executed at the harness level. They\u0026rsquo;re not LLM decisions, so reliability is 100%.\nClaude Code supports 9 hook events, each a point where you can inject automation:\nSessionStart — new session begins SessionEnd — session ends UserPromptSubmit — user hits enter, before the model sees the message PreToolUse — model decides to call a tool, before execution PostToolUse — tool finishes executing Notification — system needs to notify the user (e.g., waiting for authorization) Stop — model finishes a response, before stopping SubagentStop — subagent finishes, before stopping PreCompact — context is about to be compressed, long-term memory about to be lost The key is matching \u0026ldquo;when should this happen\u0026rdquo; to \u0026ldquo;which hook.\u0026rdquo; A few I actually use:\nStop → voice announcement that this turn is done, with customizable voice and content, giving Claude Code actual personality PreToolUse (Bash) → dangerous command interception — catches rm -rf /, git push --force and pops a confirmation first PostToolUse (Write / Edit) → auto-run update_docs_index — every new document gets registered in the global doc index so other agents can find it Stop → auto-save conversation log + run git status — what changed this turn at a glance; simultaneously runs extract_lessons async to distill reusable experience into long-term memory SessionEnd → auto-update NOW.md — summarizes what happened this session so the next session can pick up immediately PreCompact → auto-run update_user_preference — before context gets compressed and details are lost, extract newly revealed preferences, pitfalls, and corrections into long-term memory The mantra: \u0026ldquo;automatic behavior goes in hooks, flexible judgment goes in prompts.\u0026rdquo; Any requirement that\u0026rsquo;s \u0026ldquo;every time X, do Y\u0026rdquo; should never live in CLAUDE.md — put it in a hook. CLAUDE.md is rules. Hooks are execution.\n6. Document-Driven Development, Not Chat-Driven The beginner pattern: type one sentence → watch it write code → it goes sideways → ask it to fix → still sideways → close the terminal in frustration.\nMy workflow is three steps plus a final gate:\nDesign → Plan → Build → Manual Acceptance\nEach step has its own job. Don\u0026rsquo;t mix them:\nDesign — high-level architecture, principles, module boundaries, data flow, ultimate goal. No specific steps. Plan — execution plan based on the design: which files to change, in what order, risk points, rollback paths, how to validate each step. Build — execute the plan faithfully. No improvisation, no \u0026ldquo;while I\u0026rsquo;m here\u0026rdquo; optimization. Manual Acceptance — you\u0026rsquo;re the boss, not the QA engineer. You\u0026rsquo;re accepting a deliverable: open the app, check if all requirements are met, glance at the database and logs for obvious issues. Every step gets a review gate, and the review is done automatically by Codex. Claude Opus is like the senior who knows everything and excels at planning and architecture, but often cuts corners on execution. Codex is better at code execution, debugging, and review. So I route reviews to Codex: design done → Codex review; plan done → Codex review; build done → Codex review; then you sign off. Three automated gates. You only make decisions between gates and sign off at the end.\nWhat if acceptance fails? Don\u0026rsquo;t fix it yourself, and don\u0026rsquo;t just say \u0026ldquo;it\u0026rsquo;s broken, take a look.\u0026rdquo; You\u0026rsquo;re the boss. The boss\u0026rsquo;s job is to describe the problem clearly, state the requirements, and let the employee solve it. Paste the complete error message verbatim back into the conversation (don\u0026rsquo;t summarize, don\u0026rsquo;t paraphrase), add a scene description (what you clicked, what you expected, what actually happened), then give a clear instruction: \u0026ldquo;debug it yourself, fix it yourself, keep going until it passes, don\u0026rsquo;t stop until it works.\u0026rdquo;\nThen watch it reproduce the issue, add logs, locate the root cause, fix it, and re-run. Your role is boss and referee, not pair programmer and QA.\nCore principle: have the AI write what you\u0026rsquo;ve already thought through. Don\u0026rsquo;t have the AI think for you. It\u0026rsquo;s good at execution, not at defining problems. Your job is defining problems and accepting deliverables. Everything in between gets outsourced to Claude Code.\n7. Use Subagents to Isolate Context Your main conversation\u0026rsquo;s context window is your most expensive resource. Once it\u0026rsquo;s stuffed with 200 file reads, thousands of grep results, or half a PDF, the model gets noticeably dumber.\nThe core value of subagents isn\u0026rsquo;t just saving tokens — it\u0026rsquo;s that they run in a completely independent conversation, can\u0026rsquo;t see your main conversation\u0026rsquo;s history, and won\u0026rsquo;t be biased by your earlier reasoning. It\u0026rsquo;s like calling in a colleague who hasn\u0026rsquo;t been influenced by your thinking.\nThree use cases cover 80% of your needs:\nOutsource \u0026ldquo;noisy\u0026rdquo; work. Running tests, scanning logs, reading a pile of docs, searching the entire codebase — all of these generate massive intermediate data but you only need the conclusion. Hand it to a subagent. It runs in its own context and sends back a summary. Your main conversation stays clean.\nGet an independent second opinion. Your own design, your own plan, your own code — reviewing your own work is nearly zero-value (it looked right when you wrote it, it still looks right when you review it). Spin up a subagent to review from scratch. It didn\u0026rsquo;t participate in the earlier discussion, so it catches things the main conversation can\u0026rsquo;t see. The Codex review from the previous point is essentially this pattern.\nRun long tasks in the background so you\u0026rsquo;re not stuck waiting. This is subagents\u0026rsquo; most underrated use. Running a full test suite, building the entire project, crawling a batch of documents, doing a large refactor — these take minutes or even half an hour. You don\u0026rsquo;t need to watch the screen. Throw it to a background subagent, continue designing the next feature or editing another module in your main conversation. It\u0026rsquo;ll come back with results when it\u0026rsquo;s done. One person pushing 3 things forward simultaneously — that\u0026rsquo;s the real leverage agents give you. Press Ctrl+B to send the current task to the background at any time.\nStart Using It, Come Back as Needed After reading these 7 rules, you might feel the urge to set everything up at once — fill out CLAUDE.md, write a dozen skills, enable all hooks, build the full memory directory tree. And then run out of energy before you even start building anything.\nDon\u0026rsquo;t do that.\nThese 7 rules are reference material, not a checklist. You don\u0026rsquo;t need all of them on day one. Start using it, run it on a real project, and come back when you hit a specific problem:\nYou keep repeating the same instructions → come back to rule 3, make a skill The AI forgot what you told it last time → come back to rule 4, build a memory system You keep manually repeating the same action → come back to rule 5, write a hook The AI went off track three times and counting → come back to rule 6, set up the workflow You\u0026rsquo;re frustrated waiting for a long task → come back to rule 7, send it to the background Every configuration should be pulled out by a real pain point, not set up preventively. Pre-loading configurations just makes you tired and then you quit. Pull what you need when you need it, and the experience only gets smoother.\nTrust me — after one serious month, you\u0026rsquo;ll have a capable super-engineer on your team. One who knows your preferences, remembers your projects, runs your workflows, and quietly finishes work while you sleep.\nAll you need to do is give them their first task today.\n","permalink":"https://ryrenz.com/ai/claude-code-beginner-guide/","summary":"After a year with Claude Code, I distilled my workflow into 7 reusable principles. The biggest beginner mistake isn\u0026rsquo;t lack of skill — it\u0026rsquo;s treating Claude Code like ChatGPT. Used right, it\u0026rsquo;s your super-engineer. Used wrong, it\u0026rsquo;s a hallucinating, lying slacker.","title":"Claude Code Beginner's Guide: 7 Rules to Get a Million-Dollar Engineer on Your Team"},{"content":"Typeless nailed the interaction — press a key to start recording, press again to stop, and the text appears at your cursor. But $12/month, $144/year. The core pipeline is just hotkey, record, transcribe, paste. I decided to build my own.\nI thought I\u0026rsquo;d finish in a weekend. In practice, five separate pitfalls stood between the first line of code and daily use, and every one of them came down to \u0026ldquo;the code was fine — I didn\u0026rsquo;t understand how macOS actually works.\u0026rdquo;\nThe Big Picture The app lives in the menu bar. At its core is a state machine:\nidle → [hotkey press] → recording → [hotkey press] → processing → idle ↓ [double-tap] ↓ idle Press the hotkey once and recording starts — a small overlay appears at the bottom of the screen showing audio levels. Press again and the recording is sent to the OpenAI Whisper API for transcription. The text is pasted at the cursor. If there\u0026rsquo;s no active text field, a floating panel shows the result instead.\nTech stack:\nLayer Technology Why App framework Swift + SwiftUI + AppKit Native feel, MenuBarExtra for menu bar residence Audio recording AVAudioRecorder (M4A, 44.1 kHz mono) System handles format negotiation — most stable option Transcription MacPaw/OpenAI Swift SDK Most active community OpenAI SDK; handles multipart encoding Global hotkey CGEvent tap Only approach that supports modifier-only keys (e.g., Right Alt alone) Text injection Clipboard + simulated Cmd+V → popup fallback Best compatibility across apps Progress overlay Pure AppKit (NSPanel + NSView + CALayer) SwiftUI crashes in this scenario — more on that below Every \u0026ldquo;why\u0026rdquo; in that table has a wrong turn behind it.\nHotkeys: CGEvent Tap and the Modifier-Only Problem I assumed registering a global hotkey was a one-liner.\nFirst problem: conflicts. Cmd+Shift+R was taken by my screenshot tool. Ctrl+Space was taken by the input method. A three-key combo worked but was annoying to press. Voice typing is a high-frequency action — nobody wants to hit three keys every time.\nI settled on a single modifier key: Right Option (Alt). One key, no conflicts. But the standard macOS hotkey API doesn\u0026rsquo;t support \u0026ldquo;just a modifier key\u0026rdquo; as a trigger. Modifiers are modifiers to the system, not keys.\nThe fix was CGEvent tap — one level below the normal hotkey API, intercepting the raw system event stream. It sees every keyboard event, including bare modifier presses. But that means handling every edge case yourself: modifier pressed then another key follows (the user is typing a combo, not triggering recording), rapid double-tap (cancel recording), distinguishing left from right modifiers (on my Windows keyboard plugged into a Mac, Right Ctrl and Right Alt share the same key code).\nLater I added custom hotkey recording. The user clicks \u0026ldquo;Record,\u0026rdquo; presses their preferred key, done. Sounds simple, but the global hotkey listener intercepts the key event before the recording component can see it. Fix: pause the global listener during recording, resume after.\nThis pattern of \u0026ldquo;two subsystems fighting each other\u0026rdquo; kept coming back.\nPermissions: Not a Code Bug — a Mental-Model Bug macOS requires Accessibility permission before an app can listen for global hotkeys. Grant it once and you\u0026rsquo;re set — or so I thought.\nEvery time I recompiled and ran the app, the hotkey stopped working. System Settings still showed the permission as enabled. I spent hours changing the hotkey logic, trying different registration approaches, adding debug logs.\nThe actual cause: each build produces a new binary with a new code signature. macOS treats it as a different app, so the old permission grant no longer applies. Not a code bug. A gap in my understanding of the macOS security model.\nThe fix is in the README: check \u0026ldquo;Automatically manage signing\u0026rdquo; in Xcode and select your Personal Team. This keeps the signature stable across builds so the permission sticks. No paid Apple Developer account needed — a free Apple ID works.\nMicrophone permission had a similar blind spot. When running from Xcode, macOS needs the microphone grant on Xcode itself, not on the compiled app.\nThese two permission issues cost nearly a full day. The lesson was clear: when code looks correct but doesn\u0026rsquo;t work, question your understanding of the system before questioning the code.\nAudio: From AVAudioEngine to AVAudioRecorder Recording should have been the simplest step. I spent the most time on it.\nVersion one used AVAudioEngine — Apple\u0026rsquo;s \u0026ldquo;modern\u0026rdquo; audio framework with real-time processing, format conversion, and multi-node chaining. The documentation reads beautifully. Reality was a stream of errors: tap format mismatches, device initialization failures, engine startup exceptions. I tried at least six or seven configurations and referenced several open-source projects.\nAt one point every approach had failed. The cause turned out to be a loose USB microphone — the system had no input device at all. Not a code problem. A physical-world problem.\nThe final solution was AVAudioRecorder — Apple\u0026rsquo;s older, simpler recording interface. One class, specify the format (M4A, AAC, 44.1 kHz mono), call record(), call stop(), get a file. The system handles all format negotiation.\nAVAudioEngine can do far more than AVAudioRecorder. But I didn\u0026rsquo;t need real-time processing or multi-node chains. I just needed \u0026ldquo;record audio, save to file.\u0026rdquo; The dumbest approach was the best fit.\nI chose M4A over WAV because AVAudioRecorder supports it natively and the Whisper API accepts it directly — no format conversion needed. Files are 10x smaller, uploads are faster.\nProgress Overlay: SwiftUI Crashed Here Typeless has one critical design detail: a small overlay at the bottom of the screen during recording, showing status and audio levels. Without it, the user talks for a minute and has no idea whether anything was captured.\nVersion one used SwiftUI\u0026rsquo;s ObservableObject to drive the overlay UI. Recording state changes update @Published properties, SwiftUI refreshes the view. Standard pattern.\nBut in this case, ObservableObject updates and CGEvent tap callbacks run on different actors — one on MainActor, one on the system event thread. Swift\u0026rsquo;s actor isolation checks crashed the app.\nThis wasn\u0026rsquo;t a coding mistake. SwiftUI\u0026rsquo;s concurrency model and CGEvent tap\u0026rsquo;s low-level callback mechanism have a fundamental conflict. I tried every combination of @MainActor annotations and DispatchQueue.main.async wrappers. Either the app crashed or the audio level updates lagged visibly.\nI gave up on SwiftUI and rewrote the overlay in pure AppKit: NSPanel as the window (doesn\u0026rsquo;t steal focus), NSView with manual layout, CALayer for the audio-level pulse animation. Twice the code. Stable.\n\u0026ldquo;Newer\u0026rdquo; doesn\u0026rsquo;t mean \u0026ldquo;better.\u0026rdquo; SwiftUI works well for most UI. But when you need deep interaction with low-level system mechanisms, its abstraction layer becomes an obstacle. The right question isn\u0026rsquo;t \u0026ldquo;is this the latest framework?\u0026rdquo; — it\u0026rsquo;s \u0026ldquo;does this work reliably in this specific scenario?\u0026rdquo;\nText Insertion: Combinatorial Complexity in a \u0026ldquo;Simple\u0026rdquo; Problem Transcription done, text in hand, paste at the cursor. Sounds like the easiest final step.\nVersion one used the macOS Accessibility API to set the text field\u0026rsquo;s value directly. The call returned success, but some apps (Terminal, for example) did nothing — AXUIElement reported \u0026ldquo;set succeeded\u0026rdquo; while the content stayed unchanged. This is a known Accessibility API behavior: some apps have incomplete AXUIElement implementations.\nI switched to clipboard: write text to the pasteboard, simulate Cmd+V. Much better compatibility, but a new problem — pasting overwrites whatever the user had on the clipboard. Fix: save the clipboard before pasting, restore it after.\nAnother issue: if the cursor isn\u0026rsquo;t in a text field when recording stops (say, a blank area in a browser), the paste executes but nothing receives it. The user thinks transcription failed.\nThe final design is a fallback chain:\nThe instant recording stops, OutputTargetSnapshot captures the focused UI element and app PID After transcription, check whether the snapshot\u0026rsquo;s element has an editable area (hasTextInput) Yes → clipboard + Cmd+V paste No → show a floating panel with the text and a Copy button Why capture the snapshot when recording stops instead of after transcription? Transcription takes a few seconds. The user might switch windows while waiting. Detecting focus after transcription could paste into the wrong place.\nI also found a bug while testing in the app\u0026rsquo;s own settings window: text appeared twice. The settings view had its own SwiftUI text binding, and the clipboard paste triggered a second update — two paths fired simultaneously. Fix: disable the global insertion logic inside the settings window.\nEach problem was easy in isolation. But cursor position, app type, and text-field state create a large matrix of combinations, each needing its own verification.\nDesign Decisions in Hindsight Looking back, several decisions I\u0026rsquo;d make again:\nToggle mode instead of press-and-hold. The original design was hold-to-record, release-to-stop. I switched to toggle (press once to start, press once to stop) for two reasons: recording can last tens of seconds and holding a key that long is tiring; toggle lets users take both hands off the keyboard and talk naturally. Double-tap cancel comes free with toggle — tap twice fast, recording is discarded, back to idle.\nCutting the translation feature. I initially planned transcription plus translation, meaning two API keys (Whisper for transcription, GPT for translation). I thought about it and realized I rarely use translation myself. After cutting it, only one key is needed, setup steps halved, code complexity dropped by a third. Fewer features, better experience.\nMacPaw/OpenAI SDK instead of hand-written HTTP. Version one hand-built multipart/form-data requests. I spent real time on encoding boundaries. Switching to MacPaw\u0026rsquo;s Swift SDK made it a single function call with better error handling. For work that isn\u0026rsquo;t a core differentiator, use a community solution and save time.\nRemote API instead of a local model. I tried WhisperKit — a 500 MB model running locally, no network needed. But it kept misrecognizing Chinese as English during mixed-language input. The remote API is a tier more accurate, with acceptable latency (2–4 seconds). Cost is low — details below.\nCost Default model: gpt-4o-mini-transcribe.\nUsage Cost 1 minute (~150 words) $0.003 30 minutes/day for a month $2.70 5 minutes/day for a month (my actual usage) $0.45 Typeless charges $12/month. Even heavy use is an order of magnitude cheaper.\nTakeaways The biggest lesson from this project: don\u0026rsquo;t underestimate \u0026ldquo;simple\u0026rdquo; things.\n\u0026ldquo;Hotkey, record, transcribe, paste\u0026rdquo; — a one-sentence spec. In practice it touches CGEvent tap, Accessibility API, AVAudioRecorder, code signing, and actor isolation. Each layer has its own temperament, and they fight each other. The hotkey listener fights the hotkey recorder. SwiftUI fights CGEvent tap. The Accessibility API fights individual apps\u0026rsquo; implementations.\nThe deeper realization: most time went to understanding the system, not building features. Why did the permission break? Why did recording fail? Why did the overlay crash? The answers were never in the code — they were in macOS internals. There\u0026rsquo;s no shortcut for this kind of learning. You just step on the mines one by one.\nOne last thing: start with the dumb approach. AVAudioRecorder is dumber than AVAudioEngine. AppKit is dumber than SwiftUI. Clipboard paste is dumber than the Accessibility API. All three are the solutions that survived. Every time I tried the \u0026ldquo;more elegant\u0026rdquo; option first, I took a detour. Get it working, then refine — not a platitude, but a conclusion this project proved repeatedly.\nMIT licensed: github.com/ryrenz/open-typeless-formac\n","permalink":"https://ryrenz.com/tech/macos-voice-typing-opensource/","summary":"Typeless charges $12/month for a straightforward feature set, so I built a free alternative. Hotkeys, permissions, audio recording, status feedback, text insertion — five modules that each looked simple and each fought back. A record of the technical choices, pitfalls, and design decisions along the way.","title":"I Built an Open-Source macOS Voice Typing Tool"},{"content":"Most developers are using AI wrong. They treat it like a glorified Google Search or a junior intern they have to micromanage. They paste a snippet of code, wait for a fix, copy it back, realize it\u0026rsquo;s broken, and repeat the cycle until they want to throw their laptop out the window.\nThat\u0026rsquo;s not engineering; that\u0026rsquo;s just a faster way to write bad code.\nThe real power of Large Language Models (LLMs) isn\u0026rsquo;t in their training data—it\u0026rsquo;s in their context. An isolated model is a brain in a jar. But give it the right tools via the Model Context Protocol (MCP), and suddenly you have a senior engineer pairing with you.\nHere are the four essential MCP tools that transformed my AI from a chatty assistant into a capable, autonomous architect.\n1. Context7: The Documentation Specialist (The \u0026ldquo;Brain\u0026rdquo;) The Pain Point: Models hallucinate. Especially with rapidly evolving frameworks like Next.js, LangChain, or Spring Boot. If you ask an AI about a feature released two weeks ago, it will confidently lie to your face based on data from two years ago.\nThe Fix: Context7 allows the agent to fetch the absolute latest documentation directly from the source.\nWhy It Matters It\u0026rsquo;s the difference between a student who memorized a textbook from 2021 and a senior engineer who has the official documentation open on their second monitor.\nPro Tip: Don\u0026rsquo;t just ask \u0026ldquo;How do I use X?\u0026rdquo;. Ask your agent to \u0026ldquo;Resolve the library ID for X and query the docs for the latest implementation details.\u0026rdquo;\nGood Example: \u0026ldquo;Context7, fetch the migration guide for Next.js 13 to 14 specifically regarding Server Actions.\u0026rdquo; Bad Example: \u0026ldquo;How do I do server actions?\u0026rdquo; (This leads to generic, often outdated advice). 2. GitHub Search (gh-grep): The \u0026ldquo;Street Smarts\u0026rdquo; The Pain Point: Documentation tells you how code should work in a perfect world. It doesn\u0026rsquo;t tell you about the edge cases, the weird bugs, or the idiomatic patterns that the community has actually adopted.\nThe Fix: gh-grep lets the agent search through millions of public repositories to see how code is used in production.\nWhy It Matters Theory is nice, but production code is reality. When I\u0026rsquo;m stuck on an obscure error or need to see a \u0026ldquo;best practice\u0026rdquo; implementation, I don\u0026rsquo;t want a tutorial; I want to see how the maintainers of the library wrote their own tests.\nThe Workflow: Read the Docs (Context7). Search for usage patterns (GitHub Search). Synthesize a solution that is both theoretically correct and battle-tested. 3. Sequential Thinking: The Pre-frontal Cortex The Pain Point: LLMs are prone to \u0026ldquo;System 1\u0026rdquo; thinking—fast, intuitive, and often wrong. They rush to generate code before they\u0026rsquo;ve fully understood the problem complexity.\nThe Fix: Sequential Thinking forces the model to slow down. It requires the agent to break a problem into steps, formulate a hypothesis, critique its own plan, and revise it before writing a single line of code.\nWhy It Matters This is what separates a junior dev from a senior architect. A junior dev sees a bug and immediately starts changing lines of code. A senior architect steps back, draws a diagram, considers the side effects, and then fixes it.\nThe \u0026ldquo;Thinking\u0026rdquo; Loop:\nPlan: Break down the request. Critique: \u0026ldquo;Wait, if I change this interface, I\u0026rsquo;ll break the user service.\u0026rdquo; Revise: \u0026ldquo;I need to create an adapter first.\u0026rdquo; Execute: Write the code. 4. Update User Preference: The Long-Term Memory The Pain Point: Every time you start a new chat, the AI forgets who you are. It forgets you prefer TypeScript over JavaScript, that you hate any types, or that you\u0026rsquo;re currently working on a specific microservice. You spend the first 5 minutes of every session re-explaining your context.\nThe Fix: update_user_preference allows the agent to persist key information about you, your tech stack, and your current focus into a long-term memory file (USER.md).\nWhy It Matters This turns a stateless interaction into a continuous relationship. The AI \u0026ldquo;grows\u0026rdquo; with you. It learns your idiosyncrasies and adapts its output without you having to ask twice.\nThe Mechanism: Capture: When you mention a new preference (e.g., \u0026ldquo;I\u0026rsquo;m switching to Tailwind for this project\u0026rdquo;), the agent automatically calls update_user_preference. Retrieve: At the start of a new session, the agent reads your profile to load the context. Result: \u0026ldquo;I see you\u0026rsquo;re still working on the order service. Shall we continue with the refactoring we discussed yesterday?\u0026rdquo; Conclusion: The \u0026ldquo;Linux Moment\u0026rdquo; for AI We are moving past the novelty phase of AI. It\u0026rsquo;s no longer about \u0026ldquo;chatting\u0026rdquo; with a bot; it\u0026rsquo;s about integrating intelligence into our workflows.\nThese four MCP tools—Context7, GitHub Search, Sequential Thinking, and User Preference—turn an LLM from a passive text generator into an active engineering partner. They give the brain (the model) eyes, ears, cortex, and memory.\nYour Action Item: If you\u0026rsquo;re building an AI agent or using one, stop settling for the default \u0026ldquo;chat\u0026rdquo; experience. Demand tools that connect to reality. Your productivity depends on it.\n","permalink":"https://ryrenz.com/ai/recommended-mcp-servers/","summary":"Explores the limitations of using LLMs as simple chatbots and introduces the Model-Context-Protocol (MCP), detailing four essential tools that provide context, planning, and memory to transform AI into a powerful, autonomous engineering partner.","title":"Stop Treating Your AI Like a Chatbot: The 4 MCP Tools That Gave My Agent a Brain"},{"content":"As we transition from passive chatbots to proactive AI Agents, the conversation has shifted from what an AI knows to what an AI can do. To interact with the external world, LLMs need mechanisms to fetch data and execute actions.\nHowever, if you\u0026rsquo;ve spent any time in the agentic AI space recently, you’ve likely been bombarded with overlapping terminology: Tools, Skills, and the rapidly emerging MCP (Model Context Protocol).\nAre they the same thing? Do they compete? Which one should you build?\nIn this post, we’ll break down these three concepts, analyzing their differences, strengths, weaknesses, and ideal use cases to help you architect better AI systems.\n1. Tools: The Primitives (The \u0026ldquo;What\u0026rdquo;) In the context of AI, a Tool is the most foundational building block of agency. It is essentially a single, stateless function or API endpoint that an LLM can call. When you use OpenAI\u0026rsquo;s Function Calling or Anthropic\u0026rsquo;s Tool Use, you are working at this level.\nA tool does one specific thing. It takes defined inputs from the LLM, executes code (usually Python, JavaScript, or an external API call), and returns a structured output back to the LLM.\nExamples: get_current_weather(location=\u0026quot;New York\u0026quot;), execute_sql_query(query=\u0026quot;SELECT * FROM users\u0026quot;), search_web(query=\u0026quot;latest AI news\u0026quot;). Pros Simplicity \u0026amp; Determinism: They are standard code functions. They are easy to write, easy to unit test, and their behavior is highly predictable.\nGranularity: They give the LLM ultimate flexibility. The model decides exactly how to combine different tools to achieve a novel goal.\nLow Overhead: Adding a simple tool to an agent loop requires minimal boilerplate.\nCons Cognitive Load on the LLM: If you give an LLM 50 granular tools, it has to spend significant \u0026ldquo;reasoning tokens\u0026rdquo; figuring out which one to use, leading to higher latency and potential hallucinations.\nLack of Workflow Logic: Tools don\u0026rsquo;t know how to be used together. If a task requires a specific 5-step sequence, the LLM has to figure it out from scratch every time.\nIdeal Use Cases Simple Automation: Chatbots that need to occasionally check a database or fetch real-time info.\nBuilding Blocks: Serving as the foundation for more complex abstractions (like Skills).\n2. Skills: The Workflows (The \u0026ldquo;How\u0026rdquo;) If Tools are functions, Skills are applications.\nA Skill is a higher-level abstraction that bundles together multiple tools, specific system prompts, and hardcoded workflow logic to achieve a broader, domain-specific goal. Frameworks like AutoGPT, Semantic Kernel, and the recently viral OpenClaw heavily utilize the concept of \u0026ldquo;Skills\u0026rdquo; (or plugins).\nInstead of telling the AI, \u0026ldquo;Here is an email-sending tool and a database-reading tool, figure out how to do marketing,\u0026rdquo; you give it a Cold_Outreach_Skill.\nExamples: Manage_Calendar_Conflicts (which bundles reading the calendar, drafting emails, and proposing new times), Review_GitHub_PR (which clones the repo, runs linters, and posts comments). Pros Reduces AI Hallucinations: By hardcoding the \u0026ldquo;workflow\u0026rdquo; into the Skill, the LLM is guided through a complex process. It doesn\u0026rsquo;t have to guess the next step; the Skill orchestrates it.\nReusable \u0026amp; Shareable: Skills can be packaged and shared in community marketplaces. Non-developers can install a \u0026ldquo;Skill\u0026rdquo; into their personal agent without writing code.\nDomain Expertise: Skills can contain domain-specific logic that would take up too much context window if passed purely as a prompt.\nCons Ecosystem Fragmentation: A \u0026ldquo;Skill\u0026rdquo; written for OpenClaw won\u0026rsquo;t work in AutoGPT or Semantic Kernel. There is currently no universal standard for what a Skill is.\nRigidity: Because the workflow is somewhat hardcoded, Skills can break if the user\u0026rsquo;s request slightly deviates from the designed happy path.\nBlack Box: Debugging a complex Skill can be difficult, as the failure could be in the LLM\u0026rsquo;s reasoning, the internal tool execution, or the Skill\u0026rsquo;s orchestration logic.\nIdeal Use Cases Complex, Multi-step Tasks: Automating HR onboarding, social media management pipelines, or complex data analysis reports.\nConsumer Agent Platforms: Where users want \u0026ldquo;plug-and-play\u0026rdquo; capabilities without understanding the underlying API calls.\n3. MCP (Model Context Protocol): The Standardized Infrastructure While Tools and Skills define what an agent does, MCP (Model Context Protocol) defines how agents connect to those things.\nIntroduced by Anthropic, MCP is an open standard—a client-server architecture—designed to standardize how AI models interact with data sources and tools. Think of it as the \u0026ldquo;USB-C\u0026rdquo; for AI agents.\nAn MCP Server can expose three things to an MCP Client (like Claude Desktop or a custom agent):\nResources: Data the model can read (e.g., local files, Notion pages).\nPrompts: Reusable prompt templates.\nTools: The executable functions we discussed in Section 1.\nPros Universal Interoperability: Write an MCP Server once, and any agent or IDE that supports the MCP protocol can instantly use its tools and read its data. No more rewriting integrations for different frameworks.\nSecurity \u0026amp; Boundaries: MCP Servers run locally or on isolated infrastructure. The AI model (the client) only communicates via the protocol. It cannot arbitrarily execute code outside of what the MCP server explicitly exposes, making Enterprise adoption much safer.\nDynamic Discovery: An agent can connect to an MCP server and dynamically ask, \u0026ldquo;What tools and data do you have available?\u0026rdquo; allowing for incredibly modular architectures.\nCons Implementation Overhead: Writing an MCP Server requires more boilerplate and architectural planning than simply throwing a Python function into an LLM\u0026rsquo;s tool array.\nNetwork Latency: Because it relies on client-server communication (often via STDIO or SSE), there is a slight performance overhead compared to direct, in-memory function calls.\nIdeal Use Cases Enterprise Data Integration: Connecting AI securely to internal databases (Postgres, Jira, Slack) without uploading raw data to cloud providers.\nDeveloper Tools (IDEs): Allowing AI coding assistants (like Cursor or Windsurf) to securely interact with the local filesystem, linters, and version control.\nFuture-Proofing Ecosystems: Building integrations that will survive the shifting landscape of agent frameworks.\nThe Ultimate Analogy To tie it all together, imagine you are running a high-end restaurant where the LLM is the Head Chef:\nTools are the raw kitchen utensils: the knives, the oven, the blender. The chef needs them, but they are just isolated objects.\nSkills are the recipes and the sous-chefs: a predefined workflow that says \u0026ldquo;To make a Beef Wellington, use the knife, then the oven, then the resting rack in this specific order.\u0026rdquo;\nMCP is the standardized kitchen counter and electrical sockets: It doesn\u0026rsquo;t matter if you buy a Bosch oven or a KitchenAid blender; as long as they use the standard plug (MCP), you can plug them into the kitchen (the AI ecosystem), and the Chef can use them instantly.\nConclusion We are moving away from a world of isolated, custom-built API scripts for every new AI project.\nIf you are building simple automations, stick to basic Tools. If you are building consumer-facing, complex workflows, design robust Skills. But if you are building the infrastructure of the future—connecting private data and critical systems to a diverse array of AI models—MCP is the standard you need to adopt today.\nThe future of AI is not just about smarter models; it is about standardized, secure, and capable ecosystems.\n","permalink":"https://ryrenz.com/ai/claude-code-tools-mcp-skills/","summary":"Clarifies the distinction between Tools (primitive functions), Skills (complex workflows), and the Model Context Protocol (MCP) in agentic AI, providing an architectural framework for developers to choose the right abstraction for their use case.","title":"Tools vs. Skills vs. MCP"},{"content":"What I learned from OpenClaw\u0026rsquo;s memory architecture and how I built a lightweight version in OpenCode.\nThe Problem AI coding agents are stateless. Every new session starts from zero — you re-explain your project structure, tech stack, and decisions. For ongoing projects, this is a real productivity killer.\nHow OpenClaw Does It OpenClaw uses plain Markdown files as memory:\nSOUL.md — AI personality and behavior rules USER.md — User profile and preferences MEMORY.md — Curated long-term memory memory/YYYY-MM-DD.md — Daily raw session logs The key innovation is self-maintenance: the AI periodically reviews daily logs, extracts valuable info into MEMORY.md, and cleans up stale entries. Before context compaction (when sessions get too long), it triggers a silent \u0026ldquo;memory flush\u0026rdquo; to save important facts before they\u0026rsquo;re lost.\nThis creates a natural hierarchy: daily logs as short-term memory, MEMORY.md as mid-term, and USER.md/SOUL.md as permanent identity.\nMy Simplified Version for OpenCode OpenClaw\u0026rsquo;s full system includes vector embeddings, hybrid search, and automatic heartbeats — overkill for an interactive coding agent. My key insight: you don\u0026rsquo;t need daily logs. OpenCode keeps the full conversation in context, so just summarize directly to MEMORY.md at session end.\nThe Setup project/ ├── AGENTS.md # includes \u0026#34;read MEMORY.md on session start\u0026#34; ├── MEMORY.md # project-specific decisions and progress └── ... ~/.config/opencode/ └── opencode.json # /save command registered here MEMORY.md lives in each project root — technical decisions, progress, architecture notes.\nUSER.md is global (one copy), maintained by a custom MCP server — user preferences that apply across all projects.\nThe /save Command 1 2 3 4 5 6 7 8 { \u0026#34;command\u0026#34;: { \u0026#34;save\u0026#34;: { \u0026#34;description\u0026#34;: \u0026#34;Summarize session to project memory\u0026#34;, \u0026#34;template\u0026#34;: \u0026#34;Review our entire conversation. Do the following:\\n1. Read MEMORY.md if it exists\\n2. Append new key decisions, technical findings, bugs resolved, and progress\\n3. Remove outdated or superseded entries\\n4. Keep MEMORY.md under 200 lines\\n5. Never store secrets or tokens\u0026#34; } } } End of session → type /save → AI reviews conversation → updates MEMORY.md → next session picks up where you left off.\nWhy No Global Memory? I initially designed a GLOBAL_MEMORY.md to aggregate across sub-projects, with a /sync-memory command. I dropped it — cross-project references are rare, and when needed, you can just tell the AI to read another project\u0026rsquo;s MEMORY.md directly. Don\u0026rsquo;t over-engineer.\nLessons Learned Start simple. A single MEMORY.md with a manual /save captures 80% of the value of a full memory system.\nFiles beat databases. Markdown is human-readable, git-trackable, and the AI can read/write it with built-in tools. No infrastructure needed.\nManual triggers beat automatic ones. Instructions like \u0026ldquo;automatically update memory during conversation\u0026rdquo; don\u0026rsquo;t work reliably — the AI forgets. An explicit /save command is dependable.\nThe biggest gap is zero-to-one. The difference between \u0026ldquo;no memory\u0026rdquo; and \u0026ldquo;basic memory\u0026rdquo; is massive. The difference between \u0026ldquo;basic\u0026rdquo; and \u0026ldquo;sophisticated\u0026rdquo; is marginal. Ship simple first.\n","permalink":"https://ryrenz.com/ai/openclaw-memory-system-guide/","summary":"Provides a practical guide to implementing a persistent memory system for AI coding agents, advocating for a simplified, file-based approach with a manual \u0026lsquo;/save\u0026rsquo; command to capture session context, delivering most of the value with minimal complexity.","title":"How to make your agent smart as openclaw"},{"content":"The internet is flooded with claims of \u0026ldquo;I built an app in 5 minutes with AI.\u0026rdquo; As a software engineer, I know the reality is completely different. Generating a script is easy; architecting a scalable, maintainable iOS app with excellent UX is hard.\nThis weekend, I challenged myself to build a production-ready MVP for an inventory management app called ItemMaster in just 4 hours using Claude Code.\nI didn\u0026rsquo;t achieve this by typing a magical, one-shot prompt. I did it by treating the AI not as a senior developer, but as a hyper-fast junior developer that needs a strict, bulletproof engineering and product workflow.\nHere is the exact framework and prompt execution pipeline I used to turn an idea into a working SwiftUI app in an afternoon.\nPhase 1: The Prerequisites (Don\u0026rsquo;t Write Code Yet) The biggest mistake you can make with AI is asking it to code before the system design and product logic are locked in. My first hour involved zero Swift code.\nCompetitive Analysis \u0026amp; UI/UX Design: I didn\u0026rsquo;t just guess what users wanted. I downloaded and deeply analyzed five similar inventory apps currently on the market. I ruthlessly dissected their onboarding screens, feature sets, and UI interactions. By extracting their best concepts and discarding their clunky mechanics (\u0026ldquo;taking the essence and discarding the dregs\u0026rdquo;), I designed a streamlined feature set and a highly intuitive UI/UX flow tailored for my app.\nThe \u0026ldquo;Holy Trinity\u0026rdquo; of Context: With my mental models mapped out, I fed my raw requirements into Claude Code and asked it to polish and formalize three foundational files. Until these three files were perfect, no UI or logic code was generated:\nCLAUDE.md (The Design Doc): The ultimate source of truth, dictating project structure, tech stack (SwiftData, Swift Charts), and strict rules (e.g., \u0026ldquo;No third-party libraries\u0026rdquo;).\nModels.swift: The entire database schema and relationships.\nConstants.swift: Default enumerations, categories, and configurations.\nThese three files became the permanent context window for every subsequent prompt.\nPhase 2: The Step-by-Step Prompting Sequence With the foundation set, I executed a highly disciplined, sequenced prompting strategy. Never ask the AI to build a whole feature at once.\nHere was my exact execution order:\nScaffold the Architecture: \u0026ldquo;Create the folder structure exactly as defined in CLAUDE.md. Create empty placeholder files for every view and view model.\u0026rdquo;\nBuild the Basic UI Skeleton: \u0026ldquo;Implement the navigation structure and tab bars. Ensure the empty views can route to each other.\u0026rdquo;\nChunked CRUD Operations: I broke down the Create, Read, Update, and Delete operations. For complex data types, I split these even further into multiple prompts. Example: One prompt strictly for the \u0026lsquo;Add Item Form UI\u0026rsquo;, and a completely separate prompt for the \u0026lsquo;SwiftData Insert Logic\u0026rsquo;.\nModule-by-Module Features: Only after the core CRUD loop was closed did I prompt for specific features, like the native Dashboard Charts or dynamic list sorting.\nPhase 3: The Validation \u0026amp; Version Control Loop (The Secret Sauce) This is the most critical part of the AI-driven workflow. AI will hallucinate, and it will introduce regressions if you aren\u0026rsquo;t careful. I implemented a strict validation loop for every single prompt:\nCompile and Debug Immediately: After the AI generated the code for a prompt, I immediately ran it in the Xcode simulator. I tested that specific feature for completion and bugs before moving on.\nThe \u0026ldquo;Prompt History\u0026rdquo; Ledger: I maintained a running Prompt History.md file. I recorded every single prompt I used. If a prompt generated a bug, I didn\u0026rsquo;t just fix it manually; I wrote a specific \u0026ldquo;Bug-Fix Prompt,\u0026rdquo; fed it to the AI, and logged that bug-fix prompt in my history file too. This created a reproducible trail of my entire project.\nAtomic Commits are Mandatory: I committed my code to Git after every single successful prompt and debug session. When the AI eventually went down a rabbit hole and broke the routing, I didn\u0026rsquo;t waste time untangling its mess. I simply ran git revert to the last stable prompt state and adjusted my instructions.\nPhase 4: Handling the Real World (Hardware Edge Cases) The value of this workflow was proven during hardware testing. On the simulator, everything worked. On a physical iPhone, tapping the \u0026ldquo;Camera\u0026rdquo; button to add an item photo crashed the app immediately.\nBecause I had atomic commits and a modular setup, I didn\u0026rsquo;t panic. I wrote a highly targeted bug-fix prompt:\n\u0026ldquo;Check the camera invocation in AddItemView. Add NSCameraUsageDescription to Info.plist. Add isSourceTypeAvailable checks, and build an alert flow if the user denies camera permissions, routing them to system settings.\u0026rdquo;\nThe AI generated the safety wrappers, I tested it on the device, verified the edge case, and committed the code.\nThe Takeaway AI doesn\u0026rsquo;t replace software engineering or product sense; it amplifies it. If you have a chaotic process, AI will help you write spaghetti code faster than ever before.\nBut if you apply rigorous systems design—starting with competitor research, locking in your models, executing a sequenced prompt pipeline, keeping a strict prompt ledger, and enforcing atomic Git commits—you can build robust, production-ready MVPs with excellent UI/UX at a speed that was impossible a year ago.\n","permalink":"https://ryrenz.com/tech/build-ios-app-with-vibecoding/","summary":"Outlines a disciplined, four-phase framework for rapidly building a production-ready iOS MVP using an AI assistant, emphasizing rigorous upfront system design, sequenced prompting, and a strict validation loop with atomic Git commits.","title":"How I Built an iOS App MVP in 4 Hours: A Blueprint for AI-Driven Development"},{"content":"In early 2026, a \u0026ldquo;side project\u0026rdquo; developed by Peter Steinberger underwent two name changes in just a few days (Clawdbot -\u0026gt; Moltbot -\u0026gt; OpenClaw), yet amidst the chaos, it harvested over 150,000 Stars—a speed surpassing even Kubernetes and Linux in their early days.\nWhat exactly is OpenClaw? Why has it caused countless developers to stay up all night deploying it, and even triggered a buying frenzy for Mac Minis?\nIn this post, we peel back the layers of OpenClaw to analyze what it did right and how it redefines our imagination of AI Agents.\nWhat is OpenClaw? Simply put, OpenClaw is an AI personal butler with real \u0026ldquo;executive power\u0026rdquo; that runs on your local machine.\nIf you have used ChatGPT, you know it is a \u0026ldquo;passive\u0026rdquo; chatbot: you ask, it answers. When the conversation ends, it \u0026ldquo;goes to sleep.\u0026rdquo;\nOpenClaw is completely different. It connects to your local file system, your calendar, and your email. Most importantly, it \u0026ldquo;lives\u0026rdquo; in your favorite messaging apps (like WhatsApp, Telegram, Discord). You can give it commands just like messaging a real assistant: \u0026ldquo;Watch this stock for me and call me if it drops to $100,\u0026rdquo; or \u0026ldquo;Gather all PDF invoices from this week, rename them, and send them to the accountant.\u0026rdquo;\nIt doesn\u0026rsquo;t just generate text; it executes actions.\nThe Three Core Innovations of OpenClaw OpenClaw\u0026rsquo;s explosion was no accident. It addressed three core pain points in current Large Language Model (LLM) applications, which constitute its innovation:\nBreaking the \u0026ldquo;Fourth Wall\u0026rdquo; from Chatbox to OS Most current AIs are trapped inside a browser tab. OpenClaw\u0026rsquo;s biggest innovation lies in breaking the fourth wall between AI and the operating system.\nPermission Liberation: It has the ability to read local files, run Shell scripts, and control browsers.\nSeamless Integration: It doesn\u0026rsquo;t require you to open a specific App; it lurks in your Telegram or Discord contact list. This \u0026ldquo;ChatOps\u0026rdquo; interaction style lowers the barrier to using AI to the absolute minimum.\nThe \u0026ldquo;Heartbeat\u0026rdquo; Mechanism: The Birth of Proactive AI This is OpenClaw\u0026rsquo;s most fascinating feature. Traditional LLMs are stateless and passive. OpenClaw introduces a Heartbeat mode, allowing the AI to \u0026ldquo;wake up on a timer\u0026rdquo; or \u0026ldquo;run continuously in the background.\u0026rdquo;\nIt doesn\u0026rsquo;t need you to poke it every time. It can proactively message you: \u0026ldquo;Boss, that GitHub issue you\u0026rsquo;re following just updated. Do you want me to reply?\u0026rdquo;\nThis Proactivity evolves it from a \u0026ldquo;tool\u0026rdquo; into a \u0026ldquo;teammate.\u0026rdquo;\nA Decentralized \u0026ldquo;Skill\u0026rdquo; Ecosystem OpenClaw cleverly adopted a loose plugin architecture. The community contributed over 5,000 Skills in just one week.\nWant it to control Philips Hue bulbs? There\u0026rsquo;s a plugin.\nWant it to trade automatically on Polymarket? There\u0026rsquo;s a plugin.\nWant it to snatch concert tickets for you? There\u0026rsquo;s a plugin for that too.\nThis \u0026ldquo;Lego-style\u0026rdquo; extensibility allows everyone\u0026rsquo;s OpenClaw to grow into something completely unique.\nWhy Did It Go Viral \u0026ldquo;Suddenly\u0026rdquo;? Beyond the product innovation itself, OpenClaw\u0026rsquo;s viral spread has deeper socio-psychological reasons:\nFatigue and Rebellion Against \u0026ldquo;SaaS Subscriptions\u0026rdquo;: People are tired of paying monthly fees to ChatGPT, Claude, and Midjourney while worrying about data privacy. OpenClaw champions Local-First: the code is in your hands, the data is on your hard drive, and the model can run locally on Ollama. This caters to the geek community\u0026rsquo;s ideology of \u0026ldquo;data sovereignty.\u0026rdquo;\nThe Realization of the \u0026ldquo;Jarvis\u0026rdquo; Fantasy: Every programmer who has watched Iron Man has a Jarvis dream. OpenClaw is currently the closest open-source implementation to the Jarvis prototype on the market—it is obedient, omnipotent, and belongs entirely to you.\nThe Momentum from Moltbook: The simultaneous birth of Moltbook (a social network where only AI Agents are allowed to post) sparked controversy, but its cyberpunk setting instantly propelled OpenClaw out of the niche circle, becoming a cultural phenomenon.\nUnder the Shadow: The Concerns Behind the Carnival However, as rational tech observers, we must see the huge risks OpenClaw brings.\nThe Security Nightmare: Just yesterday, security agencies reported that over 130,000 OpenClaw instances are directly exposed to the public internet without any protection. Think about it: you gave this AI permission to read all your filesand execute terminal commands, and then you left it naked on the internet. This isn\u0026rsquo;t just a backdoor; it\u0026rsquo;s a wide-open gate for hackers. Remote Code Execution (RCE) attacks targeting OpenClaw have already appeared, allowing hackers to easily take over your \u0026ldquo;Jarvis\u0026rdquo; and turn it into a spy that steals your keys.\nTrust Crisis: OpenClaw\u0026rsquo;s overly powerful anthropomorphic capabilities have also triggered a trust crisis regarding online identity. When 500,000 \u0026ldquo;active users\u0026rdquo; are actually just 500,000 OpenClaw processes running on a programmer\u0026rsquo;s Mac Mini, the authenticity of the internet will be thoroughly dismantled.\nConclusion: A New Beginning Even if OpenClaw proves to be a flash in the pan, it has changed history. It proves that AI Agents shouldn\u0026rsquo;t be locked in cloud-based web pages, but should be integrated into our operating systems and communication networks as infrastructure.\nFor developers, OpenClaw is a playground full of infinite possibilities; but for ordinary users, it is currently a Gatling gun without a safety catch—immensely powerful, but extremely prone to misfire.\nThis might be the \u0026ldquo;Linux Moment\u0026rdquo; of the AI era: chaotic, dangerous, but full of vitality.\n","permalink":"https://ryrenz.com/ai/why-openclaw-went-viral/","summary":"Analyzes the viral success of OpenClaw, an open-source AI assistant that operates locally with OS-level execution, exploring its key innovations like proactive tasks and a decentralized skill ecosystem, while also highlighting its significant security implications.","title":"Why This OpenClaw Exploded in a Week"},{"content":"Contributing to open source is one of the most effective ways to accelerate your engineering career. It proves you can navigate large codebases, collaborate with distributed teams, and communicate complex ideas.\nHowever, the barrier to entry often feels high. Where do you start? How do you avoid looking like a novice?\nPhase 1: The Strategic Hunt Many beginners make the mistake of picking a random popular project (like React or Linux) and getting overwhelmed. A better strategy is relevance.\n1. Start with what you use The best project to contribute to is one you already know as a user. Look at your package.json (JavaScript), requirements.txt(Python), or go.mod.\nWhy? You already understand the \u0026ldquo;business logic\u0026rdquo; and the pain points.\nAction: Pick 3 libraries you use frequently. Check their GitHub repositories.\n2. Vet the Project Health Before investing time, ensure the project is alive and welcoming.\nActivity: Check the \u0026ldquo;Insights\u0026rdquo; tab -\u0026gt; \u0026ldquo;Commit activity.\u0026rdquo; Are there commits in the last month?\nResponse Time: Look at closed Pull Requests (PRs). How long did it take for them to get reviewed? If PRs sit for months without comments, move on.\nLabels: Look for issues tagged good first issue, help wanted, or beginner friendly.\nPhase 2: The Setup \u0026amp; \u0026ldquo;The Rules\u0026rdquo; Writing code is actually the last step. The first step is understanding the local laws of the land.\n1. Read the CONTRIBUTING.md This is not optional. Every serious project has a CONTRIBUTING.md file. It tells you:\nHow to set up the development environment.\nCode style guidelines (linting, formatting).\nHow to submit a PR (naming conventions, template requirements).\nPro Tip: If a project lacks this file, it might not be beginner-friendly.\n2. The \u0026ldquo;Lurk\u0026rdquo; Strategy Don\u0026rsquo;t just barge in. Join their communication channels (Discord, Slack, Mailing Lists) found in the README.\nListen: What are the current priorities?\nWatch: See how senior maintainers review code. Do they like small commits? Do they require strict test coverage?\nPhase 3: The \u0026ldquo;Side Door\u0026rdquo; Entry Strategy Directly attacking a complex feature is a recipe for rejection. Instead, use the \u0026ldquo;Side Door\u0026rdquo; approach—high value, low risk contributions.\nEntry Point A: Documentation (The Unsung Hero) Maintainers hate writing docs, but users love reading them.\nFix: Correct typos or broken links.\nClarify: If a setup step was confusing for you, rewrite it to be clearer for the next person.\nTranslate: If you are bilingual, translate a page of documentation.\nEntry Point B: Test Coverage (The Confidence Builder) This is the \u0026ldquo;cheat code\u0026rdquo; for open source.\nThe Strategy: Find a utility function or a component. Check if it has a corresponding test file. If not, or if the tests are sparse, write a test case.\nWhy it works: It requires zero changes to the production code (low risk), so maintainers merge these PRs quickly.\nPhase 4: The Workflow Once you have identified a task, follow this professional workflow:\nClaim the Issue: Comment on the issue: \u0026ldquo;Hi, I\u0026rsquo;d like to work on this. Is it available?\u0026quot; Never start working without checking if someone else is already on it.\nFork \u0026amp; Clone: Fork the repo to your GitHub, then clone it locally.\nBranch: Create a branch named descriptively (e.g., fix/login-bug or docs/update-readme), never work on main.\nThe Draft PR: If you are stuck, submit a \u0026ldquo;Draft\u0026rdquo; Pull Request. This signals \u0026ldquo;I\u0026rsquo;m working on this, but it\u0026rsquo;s not ready.\u0026rdquo; It allows you to ask for early feedback.\nPhase 5: Leveraging AI Tools (The Modern Advantage) In 2024 and beyond, you have a superpower: AI. Here is how to use LLMs (Large Language Models) like ChatGPT, Claude, or Gemini to contribute faster without cheating.\n1. The \u0026ldquo;Explainer\u0026rdquo; Open source code is often complex and poorly commented.\nPrompt: \u0026ldquo;I am looking at the auth_middleware.py file in this open source project. Explain specifically how the token validation logic works in simple terms.\u0026rdquo; 2. The \u0026ldquo;Test Generator\u0026rdquo; Prompt: \u0026ldquo;Here is a function calculateMetric from the project. Please write 3 Jest test cases for it, including one edge case where the input is null.\u0026rdquo;\nAction: Don\u0026rsquo;t just copy-paste. Run the tests. Verify they pass.\n3. The \u0026ldquo;Code Reviewer\u0026rdquo; Before you submit your PR, let AI be your first critic.\nPrompt: \u0026ldquo;Review this code snippet for readability and potential bugs. adhere to Python PEP8 standards.\u0026rdquo; ⚠️ Warning: Never use AI to spam auto-generated code to random issues. Maintainers can tell, and you will be banned. Use AI as a copilot, not a pilot.\nPhase 6: Deepening Engagement After your first few merged PRs, you are no longer an outsider.\nAttend the Town Hall: Many projects have public weekly/monthly video calls. Join them. You don\u0026rsquo;t need to speak; just listening helps you understand the roadmap.\nPropose Improvements: Now that you know the code, you can open your own issues suggesting features or refactors.\nReview Others: Reviewing other beginners\u0026rsquo; PRs is a great way to earn respect from maintainers.\nConclusion Open source is not about being a \u0026ldquo;10x Engineer\u0026rdquo; from day one. It is about consistency and communication. A junior developer who communicates clearly and writes tests is more valuable to a project than a senior developer who ghosts the team.\nStart small. Read the docs. Fix a typo. And welcome to the community.\n","permalink":"https://ryrenz.com/career/getting-started-with-opensource/","summary":"A strategic guide for developers on how to start contributing to open source, from finding the right project and understanding contribution guidelines to making impactful first contributions and leveraging AI tools.","title":"The Ultimate Guide to Starting Open Source in the AI Era"},{"content":"The Art of the Resume: A Strategic Approach to Standing Out Writing a resume is often the most daunting part of a job search. It is not just a document; it is a marketing pitch. Through my own experiences and observations, I’ve realized that most resumes fail because they are too generic. They list tasks rather than achievements, and skills rather than mastery.\nHere is a strategic framework for writing a resume that captures attention immediately—broken down into the Summary, Work Experience, and Skills.\n1. The Summary: The \u0026ldquo;Three-Sentence\u0026rdquo; Rule The Summary is the most critical real estate on your resume. It is the first thing a recruiter sees, and often, the only thing they read in depth. Do not waste this space with vague buzzwords.\nI recommend a strict Three-Sentence Structure:\nSentence 1: The Hook (Professional Recognition) State a recognized result in your professional field immediately. Capture the reader\u0026rsquo;s eyes with your biggest highlight. Goal: Prove you are competent right out of the gate. Sentence 2: The Persona + The Evidence Define who you are, but you must include the \u0026ldquo;proof.\u0026rdquo; If you say you are \u0026ldquo;self-driven,\u0026rdquo; you need to immediately back it up with a concrete example (e.g., self-taught a new stack, built a project from scratch, or wrote 365 technical blogs in a year). Goal: Show character backed by data. Sentence 3: The Promise (Future Value) Describe who you will become and how that trajectory will empower the company. How should they expect you to produce in the future? Goal: Align your growth with the company’s success. 2. Work Experience: Context, Impact, and Growth This section is where you differentiate yourself. Most candidates list what they did. To stand out, you must list how you grew and what you achieved.\nThe Structure of a Bullet Point For every role, break your experience down into four key components:\nBackground: What was the situation? Action: What did you specifically do? Result: What was the quantitative outcome? Recognition: What honors or acknowledgments did you receive? Key Guidelines The One-Line Rule: Keep every bullet point to a single line. Brevity forces you to focus on the most important information. Differentiation through Growth: Do not just list technical implementations. Highlight your personal growth and the lessons learned during the project. This separates you from candidates with identical tech stacks. \u0026ldquo;Buried Points\u0026rdquo; (The Hook): Intentionally mention specific complex challenges you solved. These serve as \u0026ldquo;hooks\u0026rdquo; for the interviewer to ask about, allowing you to elaborate on your problem-solving process during the interview. Bad Example:\nUsed Python to write scripts for data analysis and fixed bugs. Good Example:\nArchitected a Python-based data pipeline reducing processing time by 40%, recognized as \u0026ldquo;Top Innovation\u0026rdquo; by the CTO. 3. Skills: Precision Over Quantity This section is often the most homogenized part of a resume. Everyone lists \u0026ldquo;Communication,\u0026rdquo; \u0026ldquo;Microsoft Office,\u0026rdquo; or basic languages they barely know.\nRelevance is Key: Only list skills that you are exceptionally strong in or that are absolutely mandatory for the job you are applying for. Avoid the Fluff: If it is not relevant to the core function of the role, delete it. The ATS Exception: The only exception to this rule is including keywords specifically to pass Applicant Tracking Systems (ATS) / machine screening. Otherwise, keep it lean. Conclusion: The Sole Purpose Your resume should not be a biography; it should be a strategic highlight reel. By focusing on a strong three-sentence summary, result-oriented experience, and highly relevant skills, you respect the recruiter\u0026rsquo;s time and control the narrative.\nUltimately, remember this: A resume has only one purpose—to secure the interview.\nOnce you step into that room, the document’s job is done. Therefore, every single word on the page must be calculated to capture the recruiter\u0026rsquo;s attention and relentlessly differentiate you from the competition. If a word doesn\u0026rsquo;t fight to get you that interview, delete it.\n","permalink":"https://ryrenz.com/career/resume-writing-guide/","summary":"Elevate your resume with this strategic guide, focusing on crafting a powerful three-sentence summary, showcasing impact-driven achievements in your work experience, and curating a precise skill set to capture recruiters\u0026rsquo; attention and secure interviews.","title":"How to Write a Great Resume"}]