What I learned about CLAUDE.md after 72 hours of autonomous Claude Code runs

I've been running Claude Code autonomously for 72+ hours. The single most impactful file in the whole setup is CLAUDE.md. Here's what I actually learned. What CLAUDE.md actually does Claude Code reads CLAUDE.md at session start and after every compaction. It's not a config file — it's a persistent briefing. Everything in it shapes how the agent reasons about your project. The mistake most people make: they write it once and treat it as documentation. It should be a living document that gets updated as you learn what the agent gets wrong. The patterns that actually matter 1. Negative constraints outperform positive instructions "Don't add comments to code you didn't change" is more useful than "Write clean code." "Never use git add -A " prevents more problems than "Be careful with git." The agent knows how to write code. It doesn't know what you specifically don't want. Negative constraints are high-leverage. 2. Tool-specific rules beat general rules General: "Be careful with file edits