# CrewScore

> CrewScore is a free, open-source prompt-hygiene checker. It scores written prompts against 23 published guardrail controls across 8 dimensions and reports the controls a prompt is missing. It is a checklist, not a benchmark or certification: a score does not prove runtime quality, safety, or compliance.

## CrewScore

- [CrewScore](https://shmindmaster.github.io/crewscore/): prompt-hygiene checker and scoring tool
- [Documentation](https://shmindmaster.github.io/crewscore/docs/): usage, rules, and configuration
- [Validation study](https://shmindmaster.github.io/crewscore/docs/validation-corpus.md): reproducible corpus report
- [Vendor checklist](https://shmindmaster.github.io/crewscore/vendor-checklist/): vendor prompt-hygiene checklist
- [Security](https://shmindmaster.github.io/crewscore/security.html): security posture
- [Privacy](https://shmindmaster.github.io/crewscore/privacy.html): analytics and privacy boundary
- [Repository](https://github.com/shmindmaster/crewscore): source code
- [PyPI package](https://pypi.org/project/crewscore/): installable package

## How to use

Install: `pip install crewscore`
Run: `crewscore scan .`
Require a human gate: `crewscore scan . --require human_gate.approval_required`

## Key facts

- 23 published written-guardrail controls across 8 dimensions.
- Validation: 83 production-labeled prompts median 10/100; 273 general GPT-store prompts median 0/100; separation delta 0.614 (p = 0.0001).
- Scoring: each dimension scores its covered controls; the overall score is the floor of the mean of the 8 dimension scores. One control stated in each dimension scores 36, all 23 controls score 100. The number measures whether controls are written down, not runtime behavior.
- Configuration-smell advisories are a separate track and never feed the score.
