How much of your CLAUDE.md can a tool actually check?
A CLAUDE.md is project memory, not just a rulebook — stack, commands, folder layout and conventions all belong in it. But only the parts that are actual instructions can be verified afterwards. Across 559 real public files, 63.2% of the contents carried no instruction at all (how that figure has changed). Paste yours and see the split.
● Runs entirely in your browser. Your file is never uploaded, stored, or sent anywhere.
0% of your file is context, not instructions
—
0
items parsed
0
context & docs
0
instructions
0
checkable mechanically
0
need a human
What this doesn’t tell you. It counts what your file contains, not whether your agent followed any of it. That needs the actual session:
npx rulereceipt check
Free, runs locally, no account. There’s a sample report and the session it was generated from, so you can verify one before installing anything.
Now drop a real session and see what your agent actually did
● Also entirely in your browser — the session file is never uploaded. Same checkers as the CLI.
This is the same check npx rulereceipt check runs locally — here it ran on your dropped file, in the page, with nothing uploaded.
Context is not waste — a CLAUDE.md is meant to carry stack, commands and layout, and Claude Code loads all of it. The split matters for a different reason: only instructions can be checked after the fact, and a long file spends attention on both. Same classifier the CLI uses, compiled to run in the page — so this and the tool agree by construction. It looks for whether a line contains a directive, not for a particular file format. Method and known gaps: what we got wrong.