The CLAUDE.md Checklist: 10 Things Every Good One Has

Most CLAUDE.md files fail in the same predictable ways: they're vague, they're bloated, or they brief Claude on who you are but never tell it what not to do. A good one is shorter than people expect and more specific than people are comfortable writing.

This is a concrete, pass/fail checklist. Run your CLAUDE.md against the ten items below. Each maps to one of the six things that actually determine whether Claude reads your file and acts on it: clarity, specificity, format, voice, boundaries, and token budget — the same six dimensions CoworKit's free Builder scores automatically.

The 10-point checklist

1. It opens with who you are and what you do — in two sentences. Claude needs the role before the rules.
2. It names specifics, not categories. "Write concisely" is a category; "Default to 3-sentence answers; expand only when I ask" is a specific. Biggest predictor of whether a CLAUDE.md changes behavior.
3. It states your tools and stack explicitly. Frameworks, languages, services, file formats — stops Claude suggesting the wrong library.
4. It has explicit boundaries. The "never do this" list. Most-skipped, most valuable. "Never commit to main." "Don't send email without showing me first."
5. It's formatted so Claude can parse it. Short headers, grouped sections, scannable lists. A wall of prose buries the rules.
6. It captures your voice with an example, not an adjective. Paste one sentence written the way you'd write it.
7. It's under ~2,000 tokens (~1,500 words). If over, trim it — see how long it should be.
8. It contains no stale or contradictory rules. Conflicting instructions are worse than missing ones.
9. It separates durable facts from temporary state. Durable → CLAUDE.md; today's tasks → a working/memory file.
10. It earns its length. If you can delete a sentence and Claude behaves identically, delete it.

How to use this checklist

Score one point per item. 8–10 is strong. 5–7 means the bones are right but it's leaking specificity or carrying bloat; under 5 means Claude is mostly guessing. The two most-failed items are #2 (specifics) and #4 (boundaries) — also the highest payoff.

Skip the manual audit

CoworKit's free Builder does this automatically: describe your role in plain English and it generates a CLAUDE.md with a Claude Fluency Score across clarity, specificity, format, voice, boundaries, and token count. Already have one? Paste a real chat into Session Review.

FAQ

What should a CLAUDE.md include at minimum?

Who you are (role + context), the tools/stack you use, standing preferences as specific rules, and an explicit list of things Claude should never do.

How long should a CLAUDE.md be?

Under ~2,000 tokens (~1,500 words). See how long should CLAUDE.md be.

Why isn't my CLAUDE.md changing Claude's behavior?

Usually rules are categories not specifics (#2 on the checklist), or the file isn't being read by Claude.

Is there a tool that scores my CLAUDE.md?

Yes — CoworKit's Builder generates and scores one free; Session Review grades an existing setup.