I had Claude build me a landing page. Two versions. They looked clean. I almost shipped — and then I started reading the code.
Here’s what I caught:
- Every CTA button was invisible.
.brand-page button { background: none }wiped them all — white-on-white. A CSS specificity bug:.brand-page button(0,1,1) beat.ai-form-submit(0,1,0). - The brand font silently never loaded. The
@importfor Poppins sat after:root {}. The CSS spec drops an@importthat isn’t the first rule — so the whole page rendered in system-ui. - A contact form that would have routed 100% of leads to the wrong CRM queue. Wrong campaign ID, a missing auth token for a required compliance field, no E.164 phone normalization, no real country dropdown. The skeleton was right. The contract wasn’t.
- Eight FAQs that Claude invented because the source doc didn’t include any.
A 14-check static QA script caught the rest: nested HTML comments, em dashes in CSS comments, duplicate analytics labels, z-index drift. None of those would have failed a visual review.
The pattern
AI gets to 80% fast. The last 20% — wrong campaign ID, broken specificity, a missing auth token — needs domain knowledge.
Use it like a fast junior on a senior team.
What’s the last AI-generated thing you caught right before it shipped?