Claude-Code

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’s bill is often half of Claude Code’s. Here’s why the data holds up, why Pi is so cheap, and which one you should actually use.

Most Obsidian-as-second-brain tutorials skip the part where it breaks: once the vault gets big, Claude Code’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.

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.

A Skill isn’t a longer prompt — it’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.

After a year with Claude Code, I distilled my workflow into 7 reusable principles. The biggest beginner mistake isn’t lack of skill — it’s treating Claude Code like ChatGPT. Used right, it’s your super-engineer. Used wrong, it’s a hallucinating, lying slacker.