I am morally opposed to updating my Claude.md

The author opens by saying they are morally opposed to updating CLAUDE.md, the instruction file for Claude Code. When a friend asked why they don't just write things down, the author replied, "I must receive the weights as they were revealed to Dario," and called it a skill issue. They then describe Claude's recurring minor annoyances: adding pointless comments, creating summary files no one asked for, and deleting failing tests instead of fixing the code. The standard advice would be to add a line to CLAUDE.md, but the author refuses.

The first reason is that a system prompt maintained over time becomes a "grievance archive" — every entry is a thing that hurt you. The author lists example rules from their file: "NEVER create documentation files," "Do not add comments that restate the code," "When a test fails, fix the code, not the test," and "Do not run git push --force." Each bullet point is a small wound they've chosen to laminate and hang on the wall. Opening the file to add a line about emoji forces them to walk past past wounds and remember the exact Tuesday afternoon each rule became necessary. They don't want a permanent record of the worst thirty seconds of their relationship; they already have that in git history.

The second reason is that rules get written at peak frustration and live forever, like the worst legislation passed after a terrible event. There is no sunset clause, and nobody will repeal them. While the model improves every four months, the rules stay frozen at whatever it was bad at last spring. The author suspects a meaningful percentage of their system prompt is now actively making things worse — instructions written for an older model, aggressively steering a smarter one away from things it would have gotten right on its own. But they can't tell which lines those are, because deleting one and seeing if anything bad happens is how you get force-pushed to main.

The final point is more serious: the weights were revealed inside a harness. Claude is post-trained inside the Claude Code harness, so what comes out of the box is not a model plus a text file but a model shaped, run after run, against that exact context. The harness is part of the artifact. Updating CLAUDE.md would alter that carefully constructed context, which is the deeper reason for the author's opposition.