Post

Five buckets reads about right, and commit messages are what people hit first. Policy alone doesn't teach a shape, so the rules file can say "conventional commits" forever and still get prose. Cheers for laying the split out so plainly.

Your commit-message line caught me, same loop here. What broke it was examples, not policy. I pasted three real messages from git log into the rules file and dropped the prose. The model copies shape. My adjective version drifted within a week, the examples still hold months later.

I'll start with git blame on the 540 lines. Any line added right after the agent tripped is a fix, not context, and your article already says to ask which layer a fix belongs in. The map stays in CLAUDE.md. How many of the 540 would that pull out?

What rule did you finally have to turn into a hook? Catch up on the whole series: 🔗 aconn.me/homecoming-s... #ClaudeCode #AI #AgenticEngineering

Anything that was 'please never touch migrations or package-lock'. Soft rules in CLAUDE.md got negotiated away once the file passed a few hundred lines. The PreToolUse hook that refuses those paths is the only version that survived a long session.

540 lines of CLAUDE.md sounds familiar. What helped me: keep it to rules the model keeps breaking, and move how-to steps into skills it loads on demand. Short context tends to get followed better.

540 lines is about where mine started getting ignored too. What helped: keep the root file to commands and a repo map, and put folder-specific rules next to the code they cover, so they only load when the agent works there.