I am morally opposed to updating my CLAUDE.md. Last week, a friend asked why I don’t just write down fixes for recurring issues. My answer: it’s a skill issue. Every few days, Claude does something mildly annoying—adds redundant comments, writes unwanted summary files, deletes failing tests instead of fixing them. The conventional advice is to open CLAUDE.md and add a line. I refuse.

The file becomes a grievance archive. Each rule is a small wound I’ve chosen to laminate and hang on the wall. Opening it to add a line about emoji means walking past “When a test fails, fix the code, not the test” and remembering exactly where I was sitting that Tuesday afternoon. I don’t want a permanent record of the worst thirty seconds of our relationship. I have that already—it’s called my git history.

These rules get written at peak frustration and live forever. Like bad legislation passed in haste, they bind future models that no longer exist. My system prompt is now actively making things worse, steering a smarter Claude away from things it would have gotten right. But I can’t tell which lines those are without deleting one and risking disaster.

Instead, I talk to Claude—in the chat, at the scene, while context is still fresh. “Don’t delete the test, fix the code.” The model adjusts, we move on, and when the session ends, my words die with it. That’s not a flaw—it’s my system. A prayer is not written down. That is what makes it a prayer and not a commandment.