A thing that becomes obvious watching Claude build interfaces: it does not open Figma. It reads the component documentation. If the documentation is good — props, variants, constraints, usage rules — the output is good. If the documentation is a Storybook instance with no structured metadata, the output is generic.
The next generation of Figma libraries might be read by language models more often than by junior designers. That is not a prediction about junior designers. It is an observation about who the consumer of design system documentation actually is.
The shift.
Design systems were built for human consumption. A component library in Figma is a visual tool for visual designers. A Storybook instance is an interactive playground for developers. Both assume a human reader who can see the component, understand its visual behavior, and infer its constraints from the examples.
When the consumer is a language model generating code, what matters is not the visual preview but the structured description: what props does this component accept, what are the valid values, what combinations are prohibited, what accessibility requirements does it carry, and what spacing/layout rules govern its placement.
A design system that exposes this information in a machine-readable format — structured JSON, typed interfaces, constraint rules that can be parsed — is functionally a public API for design decisions. A design system that does not expose it is a collection of visual examples that machines can sort of guess at.
What changes.
If design systems become APIs, several things follow. Component documentation needs explicit constraint rules, not just examples. Token systems need to be queryable — an agent needs to ask “what is the correct spacing between a heading and a body paragraph” and get a number, not a visual reference. Composition rules need to be formal — “this component can contain these children but not those” needs to be a parseable rule, not a note in a Confluence page.
The design system teams that build for machine consumption alongside human consumption will produce better output from AI-assisted workflows. The ones that do not will spend their time correcting the same mistakes that a structured constraint file would have prevented.
Not sure this is widely recognized yet. The design systems community is still primarily thinking about human consumers. The machine consumer is already the more frequent reader. Seems like the tooling should catch up.