Forrest Chang took Andrej Karpathy's thread about how LLM agents screw up at coding and condensed it into a single file — grab it on Github.
Just 65 lines with four core principles:
- think and clarify first — don't silently guess what I meant
- write the minimum code necessary
- touch only what you're asked to touch
- define the goal in a way that's actually verifiable
In the original thread, Karpathy called out three patterns: the agent silently picks an interpretation instead of asking, blows up abstractions where fifty lines would do, and changes code "while it's at it" — even when nobody asked. The file targets all three.
Users in the comments say it works: PRs get cleaner, the agent stops rewriting adjacent code. Merges cleanly with any existing project CLAUDE.md — just layer these rules on top of yours.