How good is your CLAUDE.md?
Your CLAUDE.md is loaded into every conversation your agent has. Paste it below and get it graded against current best practices — token economy, commands, specificity, structure, and guardrails.
Files are analyzed on the fly and never stored.
How it is scored
Your file is read by a large language model alongside a rubric distilled from current best practices for agent memory files. Each of the six criteria below is scored 0–10, with feedback that quotes your actual file. The overall 0–100 score is a holistic judgment — not an average — so a single disqualifying problem, like a leaked secret, drags it down hard.
-
Token economy
Your CLAUDE.md is loaded into every conversation, so every line costs tokens on every turn. Short, declarative bullets score high; fluff, repetition, and long prose score low.
-
Documents the non-derivable
The file should hold what an agent can’t learn from the code: gotchas, environment quirks, and the “why” behind decisions. Restating file trees or code structure loses points — it wastes tokens and goes stale.
-
Commands & workflows
Concrete dev commands earn points: build, lint, run the app, and especially how to run a single test.
-
Specific & actionable
“Use 2-space indentation” beats “format code properly.” Vague aspirations the agent can’t act on score low.
-
Structure & scannability
Clear headings and grouped bullets, with emphasis like IMPORTANT reserved for the few rules where deviation is costly.
-
Guardrails & etiquette
Explicit boundaries: what must never be touched, branch and PR conventions, destructive-action warnings. Secrets in the file are heavily penalized.