claude-md-review

# CLAUDE.md Review Audit a `CLAUDE.md` against the failure modes that actually cost output quality. **Premise:** `CLAUDE.md` loads into context on every single session, so every line is either paying rent or costing you tokens on every turn forever. Most "the model isn't following instructions" problems are instruction problems. ## Steps 1. Read the target file (default `./CLAUDE.md`; also check `~/.claude/CLAUDE.md` and any nested `**/CLAUDE.md` if the project has them, since the closest one wins). 2. Score each dimension below and quote the specific lines that fail. 3. Output the report format at the bottom. Propose concrete rewrites, not "consider being more specific." ## What to check **Specificity** — the highest-leverage dimension. - Flag unfalsifiable directives: "write clean code", "follow best practices", "be careful", "use good naming". - Every rule should be checkable by reading a diff. `"Refactor functions over 40 lines"` is checkable; `"keep functions short"` is not. **Named anchors.** - Rules that reference "the config", "our API layer", or "the usual pattern" force rediscovery every session. Replace with real paths: `src/config/env.ts`, `src/api/client.ts`. - Verify every path, command, and filename mentioned still exists. Report the dead ones — a `CLAUDE.md` pointing at a deleted file actively misleads. **Staleness.** - Version numbers, model names, and tool commands that no longer match the repo. - Instructions for a framework, script, or directory th...