Building a Design System with AI
We used AI to bootstrap a design system in two weeks. The components were the easy part. The hard part was consistency.
Design systems are supposed to save time. But building one from scratch is a months-long investment most startups can't afford. We tried a different approach.
The experiment
Using Claude and a reference design (shadcn/ui as a base), we generated 30 components in two weeks. Buttons, inputs, cards, modals, tables, navigation — the full set.
What AI handled well
Individual component generation was fast and accurate. Given a clear spec and existing examples, the AI produced components that matched our design tokens and naming conventions.
What required human intervention
Consistency across components was the real challenge. The AI would make subtly different spacing decisions, use slightly different animation curves, or interpret "muted" differently depending on context.
The process that worked
We found that generating components in batches of three to five, reviewing for consistency, then feeding the corrections back as context produced much better results than generating everything at once.
Outcome
We shipped a 30-component design system in two weeks instead of the typical two to three months. About 70% of the generated code survived review unchanged. The rest needed manual consistency passes.