Show HN: CrewScore – coverage check for AI agent prompts (0.6.11)

shmindmaster1 pts0 comments

CrewScore — find missing written guardrails

Browser-local guardrail checklist

Find the safety rules your AI agent prompt forgot.

Paste your AI instructions. CrewScore checks them against 23 public guardrail controls in your browser — no prompt upload, signup, API key, or model call.

Try a 10-second demoCheck my instructions

Runs locally23 public controlsOpen sourceNo API key

Written-control coverage, not runtime proof. It checks whether guardrails are written down — not whether an agent follows them.

Synthetic instructions<br>Ready

Checking locally

— of—

First gapWaiting for local check

Before — →After —

Play<br>Pause<br>Replay<br>The demo runs once and stops.

A synthetic example is checked locally. CrewScore reports written-control coverage and the first control to review; it does not test runtime behavior.

10/100 median coverage among 83 production-labeled agent prompts

356 prompts scored offline · not a quality ranking · method & harness

What are you checking?<br>Where do I find my instructions?

I use<br>ChatGPT<br>Claude<br>Cursor<br>Other / paste

Pick where your AI instructions live — or paste them below.

Paste instructionsUpload a local fileImport a public GitHub file

System prompt / AI instructions<br>Coding-agent instructions

AI assistant instructions<br>Paste ChatGPT custom instructions, a Claude Project prompt, a support-bot prompt, or another system prompt.

Drop a local text file here, or choose a file. It stays on this device.

Import GitHub file

Supported: public github.com/.../blob/... and raw.githubusercontent.com files only.

Find missing guardrailsScoring happens locally.

Your result will appear here<br>Start with a demo or paste instructions. CrewScore will show written controls it detected and suggested controls to review.<br>Run a sample check

Review suggested wording<br>Select only wording that belongs in this prompt. These are editable text suggestions, not runtime enforcement.

Apply to working copyCancel review<br>Pair written guardrails with runtime approvals, tool restrictions, and evaluation. See runtime next steps.

How it works<br>Paste instructions, find missing written guardrails, then improve the wording or add a recurring repository check.<br>Load a support-assistant example Load a coding-agent config example (AGENTS.md)

Open rules<br>Every control, regex, and ruleset version is public. Developer mode shows exact matches and IDs.<br>Read the scoring catalog

Keep controls from regressing<br>Use the CLI or Action to protect specific written controls and generate prompt-free SARIF findings.<br>Set up CI policies

Methodology<br>Coverage helps identify omitted written controls. It is not a quality ranking, a certification, or a runtime test.<br>Read validation and limitations

Questions, answered plainly<br>Does CrewScore upload my instructions?No. Scoring happens locally; prompt text never leaves this browser.<br>Does a high result prove an agent is safe?No. It means public text patterns for written controls were detected. Runtime enforcement and testing remain separate.<br>Can I make this a recurring check?Yes. Use the open-source CLI or GitHub Action to require particular controls, protect a prompt-free baseline, and emit SARIF.

Find missing guardrails

Where to find AI instructions<br>Close<br>ChatGPT: Settings → Personalization → Custom Instructions. Copy the custom instructions text.<br>Claude: open a Project → Project instructions (or Claude.ai custom instructions). Copy the instruction text.<br>Cursor: open AGENTS.md, .cursorrules, or project rules in the repo root. Coding-agent files are scored for config smells, not a governance grade.<br>Teams or apps: look for a system prompt, assistant setup, or AI instructions field. Copy only text you are allowed to inspect.

instructions prompt written agent controls find

Related Articles