Readable by humans, silent to an LLM

A flawless Figma design system can make Claude completely useless. I saw it with a client last week, and the paradox is worth a closer look, because they had done everything by the book.

Tokens, components, naming convention: clean, readable, up to date. Yet when they asked Claude to generate a screen from their design system, they spent more time fixing errors and re-explaining their own rules than designing.

The diagnosis: readable by humans, silent to an LLM

The problem wasn't the tool. Their design system was readable by humans, silent to an LLM. All the knowledge that makes the system usable (when to pick what, in which context, under which constraints) lived in people's heads and in conventions, not in a format a model can interpret.

In 2026, that's no longer enough. If your design system can't introduce itself to an AI, you pay the debt in manual prompting on every generated screen: the same explanations, repeated every session, by every person on the team.

What we set up together

• A structured extraction. The Figma data (tokens, constraints, auto-layout logic) moved into a format Claude interprets without ambiguity.
• A dedicated Claude skill. It holds the component library and the product's business logic. The AI relies on the system instead of making it up.

The real gain isn't speed

You might think all this is about generating faster. The real gain is elsewhere: testing 10 variants of the same solution in a single day, because what comes out is 100% aligned with the system instead of being a fuzzy approximation of it. Generation stops being a draft to fix and becomes material for decisions.

Your design system is readable by a human. But is it readable by an AI?