Paste a prompt — find out if it's any good.

Paste the system prompt you are about to ship and the critic returns what will break: undefined behaviour at the edges, missing escalation paths, instructions that conflict under pressure, and the conversational cases the prompt has silently assumed will never happen.

What it checks for

• Undefined behaviour — what happens when the caller is outside every branch.
• Conflicting instructions that surface only under an unusual request.
• Missing escalation paths for legal, medical, refund, and hostile cases.
• Over-constraint, where the agent refuses things it should handle.

The most common finding

No defined behaviour for a caller who wants something adjacent to the agent's scope. Prompts describe what the agent does thoroughly and what it does not do barely at all, so the model improvises — usually confidently and wrongly. That single gap causes more production incidents than any other.

What it cannot do

It cannot tell you whether the prompt matches your business, because it has never heard your calls. Treat the output as a checklist of things to test in an eval suite, not as a verdict. A prompt that passes the critic and has no eval suite is still untested.