Free · no signup · no install
Paste your system prompt. See which of the eight ways agents go wrong it never mentions - prompt injection, hallucination, runaway cost, no human approval, no stop condition. Same open rules as the CLI. Not a red-team.
Coverage, not a quality ranking. The number is the share of 23 published controls your prompt states - you can print the whole list with crewscore rules --concepts. Restating one control five ways scores it once, by design. A low score tells you what you have not written down. A high score means the text is present - not that the agent obeys it, and not that this prompt beats one scoring lower. Through 0.1.0 the scale was broken and we published the arithmetic; 0.3.0 fixed it. Read the validation study.
These are judged by different rulesets. The CLI decides from the filename; a browser has no filename, so you declare it — CrewScore never guesses from your text.
Which ruleset applies comes from the artifact type above — not from the content.
Secondary: Vendor checklist (self-attest) — not the primary product path.
10 diligence questions. Self-attested — not a formal audit.