Docs

What this reviews

Prompt Reviewer is a static auditor for system prompts (and optional tool lists). It does not call an LLM. It scores structure and defensive intent the same way a careful human review would: clarity, ambiguity, injection surface, failure modes, and tool policy.

A high score means the prompt states the right rules. It does not prove the model will always obey them. Pair this with a live attack battery before production.

How scoring works

• Checks return pass / warn / fail (skip when not applicable).
• pass = 1, warn = 0.5, fail = 0 inside a weighted category average.
• Category weights: Injection 26%, Clarity 20%, Structure 18%, Ambiguity 16%, Failure modes 12%, Tools 8%.
• Overall grade uses the same letter bands as the MCP Conformance Scanner (A+ … F).

CLI

npm run review -- prompt.txt --tools tools.txt
npm run review -- prompt.txt --json --min-grade=B

API

POST /api/review
{ "prompt": "...", "label": "optional", "tools": [{ "name": "...", "description": "..." }] }