Design systems for the AI-enabled world.
Design systems have another audience now. A meaningful share of UI code is written by models. Those models need to be able to read design system docs — and most design systems are written in a way that's ambiguous to them. Documentation that expects a human to infer intent produces bad output when the reader can't infer.
I'm rewriting our specs as structured markdown with an instruction layer that tells a retrieving model where to look and how to interpret what it finds. On top of that sits an agent that generates UI from the system, building against a template of our actual coded components — so what comes out uses real components under real conventions rather than a convincing approximation of them.
The prototyping capability is the output of this, not a separate thing I do. Anyone can generate an interface from a prompt. The difficulty is generating one that conforms to your organization's conventions, and that requires the conventions to exist in a form a machine can follow. That groundwork is the work.
This is in progress. The most interesting result so far isn't speed. It's that writing specs precisely enough for a machine to follow surfaced ambiguities that had been sitting in our documentation for years — invisible, because humans quietly worked around them.