Purple

"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. And... there's more stuff coming... but can't talk about it yet lol Anyway, the robots love purple."

Large language models repeatedly produce the same popular UI patterns, frequently generating purple-gradient rounded buttons and other homogenized choices. These models struggle to follow multiple or conflicting design guidelines consistently because training data biases reflect prevailing trends from platforms like Dribbble in the late 2010s and early 2020s. A v0 approach adds design systems and registries to provide structured context about components and their relationships, enabling constraint-based generation. Defining design tokens with shadcn lets the AI remain inside specified constraints. Additional features are under development. The models currently favor purple but are gradually learning to apply other colors.

Read at Medium