AI-built storefronts passed self-review but a browser QA gate caught four bugs

A team had an AI agent build two fictional storefronts, a Warsaw watch maker and a small type foundry, then graded the output with a fixed pass/fail browser checklist run at three screen sizes. The agent had already approved its own work, but the gate flagged four problems: a hero section that omitted a needed spec table, a purchase bar that broke on mobile, a reduced-motion fallback that froze mid-animation, and signage text that overflowed with real names. The authors argue the issues are structural rather than aesthetic, since AI self-review tends to grade intentions instead of the actual rendered artifacts. They describe recurring patterns in AI-built interfaces and propose a short generation-time constraint prompt paired with a browser acceptance gate. The write-up also promotes a paid guide from the authors' shop.