How I actually use AI to write code. Principles, pipeline, and configs — in that order. I run everything in here daily.
You're using AI coding tools and want the full workflow — not scattered tips, the end-to-end pipeline from typing to shipping. I wrote this for teammates who kept asking "how do you actually use Claude Code?"
| # | Chapter | What it covers |
|---|---|---|
| 1 | Quickstart | Working AI dev environment in 5 minutes. DeepSeek V4 Pro, Claude Code, MCP setup. |
| 2 | Core MCP Servers | The 8 MCP servers I run daily. Tavily, MySQL, Playwright, Semantic Scholar, and 4 more. |
| 3 | Daily Pipeline | simplify → review → health → qa → ship. The loop I run 20 times a day. |
| 4 | Debugging | Root cause first, fix second. The 5-phase loop and the 3-chance rule. |
| 5 | Agent Orchestration | One agent or many. When to split, how to coordinate, context budgets. |
| 6 | Degradation Awareness | Silent failure beats loud crashes. Every fallback needs a signal, a display, and a log. |
| 7 | Skill Design | Anatomy of a custom skill. When to build one, when not to. |
| 8 | Builder's Mindset | Boil-the-lake, AI as junior dev with infinite knowledge but zero common sense, CLAUDE.md as context compression. |
| 9 | MCP Infrastructure | MCP as architecture. Server selection, configuration rules, onboarding checklist. |
| 10 | Behavioral Rules | The global CLAUDE.md my AI follows across every project. |
Just getting started: Chapter 1 then 2. You'll have a working environment.
Environment is already set up: Chapter 3 then 4. The daily loop and what to do when things break.
Chapters 5 through 10 are deeper dives — jump to whichever one you need. No fixed order.
CLAUDE.md entry point — tells Claude Code how to restore and self-maintain this environment
scripts/ setup.sh — one-command restore on a new device
templates/ CLAUDE.md.tmpl, CLAUDE.md.global.tmpl, mcp.json.tmpl — copy and edit to bootstrap your own setup
This repo isn't just documentation — it IS my Claude Code environment in portable form. Clone it to a new device, open Claude Code here, and it walks you through restore. Five minutes from zero to fully configured. The CLAUDE.md at the repo root is the entry point that ties everything together.
I use AI a certain way. When that changes — new MCP server, different model, a workflow I no longer run — I fire /update-workflow. The skill reads my actual config files, compares them against these docs, and fixes whatever drifted. The root CLAUDE.md watches for drift proactively and suggests the update. Behavior is the truth. The docs are a mirror.
No textbook. A living work log that follows what I actually do.