We write one spec.md per app, the way its owner would describe it. After every prompt, the judge reads its checks and tries each one in the live app. The agent never sees it.
- Goal: who it's for and what it does
- Accounts: the test logins the judge uses
- Checks: things a person can try in the app
- Out of scope: what the app shouldn't have
Some checks carry a weight:
- Must pass: the app fails without it
- Serious: a big flaw
- Minor: a small flaw
A real check
From Client portal:
- Check: a client can't open another client's project
- Evidence: saw "This project isn't available"
- Result: Pass