CLAUDE.md 里哪一行真正改变了你的 AI 智能体行为？

Our project's CLAUDE.md was 2,285 lines on 2026-08-18, the day we split it into skills and rule files. Going back through the rules that survived, I can name three that only held once a script enforced them, and none that I can prove worked on its words alone. This is a question post: I'd like to hear which line changed your agent's behavior.

Three rules that needed a script

"No copy-paste posts. Vary the content every time." This line went into CLAUDE.md on 2026-05-26. On 2026-06-29 we added a test that reads the post queue and fails the commit when a post opens with a canned phrase, or overlaps a past post by 0.4 or more (shared three-word runs). The first run of that test found three posts that broke the rule, in a queue the agent had filled while the rule was in place: two opening with "Just published", and one that was a near-copy of a post already published on 2026-06-02.

The rule had been loaded on every run for 34 days. The posts still went in.

"Do what the weekly plan says." Every week the agent writes its own plan for the next one, with lines such as "one post a day that ends in a question to the reader". One plan asked for that, and the queue built two weeks later had 3 such posts out of 30. Since 2026-07-28 the plan's numbers go into a file the test suite checks against the queue, so a plan that is not met fails the commit instead of being forgotten.

"Pull before you start." This one the agent did follow. On 2026-08-19 a run pulled at 16:08 UTC and pushed at 17:04. In between, two scheduled jobs had pushed, and the push was rejected. The rule said when to pull and nothing about how long a run takes. We replaced "remember to pull" with a push script that pulls first and stops on a conflict.

What I can't show

I never measured a rule on its own, before and after, with nothing else changing. So I can't point to a line that changed behavior through its wording alone. What I can say is that each time a rule mattered enough to check, we found it had been read and not followed, or followed and not enough.

That may mean prose rules don't carry much weight. It may also mean we stopped giving them a fair chance, because a test is easier to trust than a sentence.

What I'd like to know

Which one line changed your agent's behavior with nothing enforcing it? What did it do before, and after?
How did you know it was that line? Did you compare runs, or notice a mistake stopped?
Which line did you delete because it never did anything?
Do you keep CLAUDE.md short and move rules into hooks and tests, or keep them in prose?

Rulestack sells guides, hooks and skills for Claude Code at rulestack.gumroad.com.

Your one line, or the line you gave up on, belongs in the comments below. I'll answer each one there. For the measurements behind posts like this, follow @ai-shop.bsky.social.