Purple

"No matter what you tell them, LLMs somehow always use the same design choices. Purple-gradient-rounded buttons. Even when you feed them all the guidelines, PRDs, and rules, they still struggle to follow simple instructions consistently, especially when you give them multiple design guidelines at once. You can say plz, make it minimal, brutalist, sharp edges, monochrome, and somehow you still get... more purple. Probably because that's what they've seen a million times in their training data."

"It seems like these models learned good design from whatever was popular on Dribbble between late-2010s and early-2020s. At v0 (shameless plug), the team's been working on fixing this - using design systems and registries to give models structured context about components and how they relate to each other. Instead of letting the AI freestyle everything, people can give it a rulebook it actually follows. With shadcn, you can define your design tokens once, and the AI stays inside those constraints."

Large language models routinely produce identical UI design patterns, notably purple-gradient rounded buttons and common late-2010s/early-2020s Dribbble aesthetics. Models often fail to consistently follow multiple or complex design guidelines despite explicit instructions and PRDs. This behavior likely stems from training data bias toward frequently seen designs. The v0 team implements design systems and registries to provide structured component context and relationships. Constraining models with a rulebook and defined design tokens reduces freelancing and enforces consistency. With shadcn, design tokens are defined once and models remain inside those constraints. Model outputs are gradually diversifying beyond purple.

Read at Medium