I gave an AI model three one-line website prompts and reviewed each first draft like a design lead would: score it against a fixed rubric, pin every issue, then redesign it. Toggle AI draft / Redesign on each case.
MessageCan a visitor say what this is in 5 seconds?
HierarchyOne clear focal point and reading order.
TypographyType fits the brand and stays readable.
ContrastWCAG AA for all text that matters.
Brand fitLooks like this business, not any business.
ActionOne obvious next step for the visitor.
Patterns I saw across all three drafts
Generic promise copy. "Unlock", "seamless", "revolutionize": the model writes for every business at once, so it says nothing about this one.
Decoration over hierarchy. Gradients, glows and patterns appear before a clear focal point exists.
Contrast lost on effects. Supporting text sits at 40–55% opacity on busy backgrounds and fails AA.
Too many equal choices. Several CTAs with identical weight instead of one primary action.
Feedback that helps the tool improve has to be specific: name the element, the rule it breaks and the fix, so the next output can be checked against it.