Every Claude Code user has a list of things they've typed more than once. "Don't add comments." "Use pnpm." "Stop apologizing." Each correction fixes one session and then disappears.
We wanted to know which corrections are universal, so we read the issue tracker.
How
We fetched every issue in anthropics/claude-code labelled area:model or data-loss (nearly 10,000), dropped pull requests, duplicates and withdrawn reports, grouped issues with similar titles, and ranked the groups by 1 + reactions per issue. Then we picked, by hand, the 20 behaviours a CLAUDE.md rule can actually change, using only issues that are still open or closed as "not planned". The fetch and mining scripts are in the repo, so the numbers are reproducible.
What people complain about most
| Behaviour | Reactions |
|---|---|
| Adds code comments, even when told to stop | 252 |
| Can't stop saying "load-bearing" | 185 |
| Runs destructive git commands | 121 |
| Replaces curly quotes and em dashes with ASCII | 61 |
| Says "belt-and-suspenders" everywhere | 43 |
| Loops when exploring instead of asking | 38 |
| Edits files in plan mode | 38 |
Uses $(...) and triggers approval prompts |
33 |
| Drifts into English after reading English output | 30 |
Passes rg -r (that's --replace) |
29 |
A caveat: each rule currently rests on one issue, so "most-reported" really means "most-reacted". Reactions are a rough signal, but they're public and anyone can check them.
How Claude Code helped build it
Claude Code wrote almost all of the code (33 of 35 commits): the script that fetches the issues with the gh CLI, the title cleanup, the similarity grouping and the ranking, plus 125 tests. We gave it a CLAUDE.md for the project (run the tests before saying done, a regression test for every bug fix) and reviewed each step.
It also got things wrong. Its first grouping threshold lumped 1,489 issues into a single group; we caught that in the output, tightened it, and the largest group is now 24. The final 20 rules were picked by a person, not the model.
The rule pack
It's free and MIT. One command, every rule linked to its issue:
curl -fsSL https://raw.githubusercontent.com/globalmpc/nomore/main/rules/CLAUDE.md -o nomore-rules.md
Then add @nomore-rules.md to your CLAUDE.md (or put both in ~/.claude/ for every project).
The universal list is half the story. The other half is what you keep repeating. The repo also has a local CLI (preview) that reads your Claude Code transcripts on your machine, with no network, and suggests the corrections you repeat as rules you choose. What surprised us on our own sessions: most repeats were tasks ("commit", "continue"), not preferences, so it shows a short list and you pick.
Repo: https://github.com/globalmpc/nomore. Which rule would you add?
Top comments (0)