Agent design

Overview

I built a small QA agent…

…with Claude and Cursor, that reviews design prototypes against a seven-item rubric I wrote and reports only what it can prove.

I built it to answer one question: how do you make an AI checker's findings trustworthy, when trust has to live in code, not copy?

What I learned: grounding has to be enforced in code, not just requested in a prompt. Failure handling only matters if you actually test it, so I broke this four ways on purpose. And the biggest lesson: one clean run proves nothing. Measured across multiple runs, it swung from catching none of the real issues to catching all of them. That's the whole point: measure it, don't assert it.