Guide
Check AI agent instructions for missing written guardrails
A useful prompt review starts by separating what is written from what is enforced. Written guardrails clarify intent; runtime tools, approvals, and tests enforce behavior.
Start with the instructions you own
Paste a system prompt, custom-instructions field, support-bot setup, or agent instruction file into the local checker. It reports 23 public controls across injection handling, evidence, sources, cost limits, human approval, stopping, records, and sensitive data.
Review omissions one control at a time
Do not append every template just to raise a number. Choose wording that fits the agent's actual tools and authority, then make the runtime behavior match it. For a bot that sends messages, a written human-approval rule should be paired with a real approval hook.
Make meaningful controls recurring
For repositories, use `crewscore init .` to create a prompt-free baseline and pull-request workflow. Require precise public controls such as human_gate.approval_required instead of interpreting an average coverage number as a safety threshold.
Know the limit
CrewScore uses deterministic text matching. It does not execute tools, attack a model, certify compliance, or prove that an agent follows the instructions. Use evaluation and runtime controls alongside prompt review.