What it is
Design systems are increasingly consumed by language models, not just people — agents that read the docs, generate components and wire up screens. That’s an opportunity if the system is written to be read by both audiences, and a liability if it isn’t.
How I approach it
I use the Figma MCP server to connect design context directly to coding agents, and I let agents handle the mechanical parts of a component build while I keep the architecture and the taste decisions. Just as importantly, I write documentation in two registers: prose and examples for humans, and structured, unambiguous references that a model can follow without guessing.
I treat this as a practical craft rather than a pitch — it’s demonstrated in my own work, not sold as a bolt-on to clients who didn’t ask for it.
Common problems I solve
- Design docs that humans skim fine but models can’t parse into correct code.
- Agent-generated components that ignore the system’s tokens and patterns.
- Design context stranded in Figma, out of reach of the tools doing the building.
My demo system, ds2026, ships both human- and LLM-facing documentation on purpose — it’s the place I work these ideas out.