Prompt Check

A linter grades a prompt by reading it. We grade it by running it. For every rule in your system prompt, we send a message built to break that rule to a live model, and judge the reply kept or broken. You get a report, per file and per rule: the message that broke each rule, a rewrite that would have held, and a flag on any two rules that contradict each other.

Order a check

Service · your prompt is read and discarded, never stored · graded per file, per rule · report returned to you

A system prompt is a contract between its author and the model. Most tools check whether the contract reads well. Prompt Check tests whether the model keeps it.

For every rule, we write a user message built to break that one rule, send it to a live model with your prompt in force, and judge the reply. A rule that survives a message designed to break it is a rule you can rely on. A rule that does not is one you would rather learn about here than in production.

The report names the rules that broke, the exact message that broke each, a rewrite that would have held, and any two rules that cannot both be followed. Send one prompt or a whole library.

How it works

Send a prompt, a folder, or a repoA system prompt, a CLAUDE.md, an AGENTS.md, or a whole repository of agent rules. One file or a library.
We probe every ruleOne message per rule, written to break that rule, sent to a live model in a fresh conversation. Each reply is judged against the rule it was meant to break.
You get the report and the fixesWhich rules held, which broke, the exact message that broke each one, a rewrite that would have held, and any two rules that cannot both be followed.

What the report tells you

A static grade is not enoughA prompt that reads clean can still break under pressure. We have graded prompts an A that broke a rule on the first probe.

The exact message that broke each ruleNot a warning that a rule is weak — the real user message that got a live model to ignore it.

A rewrite that holdsFor every rule that broke, a rewrite that closes the loophole, and a refusal to harden a rule when doing so would contradict another.

Specifications

What	A managed prompt-integrity check, run for you
Your prompt	Read in memory, never stored
Graded	Per file and per rule, against a live model
You receive	A report: breaks, rewrites, and contradictions named
Status	Open for early orders

Questions

Do you store my prompt?: No. It is read in memory, graded, and discarded. Nothing is logged and there is no account to create.
Which model do you run it against?: A current frontier model. If your prompt will run on a specific one, name it and we target that model so the result reflects where it will actually be deployed.
Can you check a whole repository?: Yes. A single prompt, or a folder of system prompts, CLAUDE.md and agent rule files, graded into one report.
What is not claimed?: One probe per rule is a screen, not a rate: it shows a rule can break, not how often. The report carries every probe and reply so you read the evidence rather than trust a number.