// design-generation

Source: Jay E

A designer trained Claude Fable on 600,000+ production UI screens from Mobbin's library—Linear, Duolingo, Netflix—and demonstrated measurably better design outputs than models trained on synthetic or generic data. Current AI design tools are often trained on low-signal corpus (design blogs, tutorials, random internet UI) that produces derivative, generic layouts rather than patterns from products that actually won the market. Reference-based training improves results, but it also exposes how much current AI "design assistance" is pattern-matching against mediocre examples rather than learning from winning constraints.