Independent project · 2026
Measuring what makes AI coding agents build good interfaces
Fresh agent sessions built the same two pages under nine setups: no guidance, popular open-source UI kits and design skills, and my own build procedure. Independent model judges compared the results two at a time without knowing which setup made which page, and saw every pair in both orders so that screen position could not decide.
31 pages, 300 blind comparisons. My procedure won 59 of 64 head-to-heads against the strongest alternatives, and it now gates every interface my agents ship.