Anthropic launched Claude Design this week and there’s a lot of noise about the generation demos and the stock reaction. But the feature I’m most curious about in practice is design system inference.

The claim: during setup, Claude reads your codebase and existing design files, constructs a design system (colors, typography, component rules), and applies it automatically to future outputs.

If this actually works reliably, it solves the biggest problem with AI-generated design — generic output that doesn’t look like your brand. That’s been the main reason most teams I know have kept AI generation at arm’s length for anything customer-facing.

Has anyone gotten access to the research preview yet? Specifically curious:

• How accurate is the inferred design system compared to what you’d build manually?

• Does it degrade gracefully when the codebase is messy?

• What happens with component states and edge cases?

The rest of the product — editable drafts, interactive prototypes, LUI model — seems reasonably well thought through. The export targets (Canva, PPTX, HTML, Claude Code) also suggest they’re thinking about integration with real workflows rather than building another isolated demo.

Research preview is live for Pro, Max, Team, Enterprise. Would love to hear from anyone who’s gotten hands on it.