The AI-native design system
"AI-first design system" is mostly a buzzword. This is the working version, as a four-rung ladder.
When the industry says AI-first, it usually means hygiene: structured docs, clean tokens, git, Code Connect. That's real, because AI raises the penalty for sloppy foundations, but it's the floor, not the building. I think of it as a four-rung ladder, and I've shipped all four.
Rung one: AI-ready
Machine-readable component documentation, semantic token structure, and design rules (content, accessibility, personas) codified as data instead of wiki prose. Table stakes.
Rung two: AI-assisted delivery
A pipeline delivering around 60% production-ready code at design hand-off. In one sprint I built seven skills encoding our design system's rules, and content accuracy against our guidelines went from 10% to 100%. When a tooling limitation blocked the design-to-code loop, I built an MCP server around it, making the workflow bidirectional. The design team shipped three working applications without engineering support.
- production-ready code at hand-off
- 60%
- content accuracy against guidelines
- 10% → 100%
- applications shipped without engineering support
- 3
Rung three: AI-native governance
The system polices itself. That instinct predates the AI wave: I built a Figma plugin that automated our accessibility scoring and replaced a paid vendor subscription outright. The LLM era extended the pattern: token drift detection, prototype QA linting, and an audit loop that scored all 127 components across six dimensions and turned findings into engineering-ready tickets, with fixes verified within the week.
Rung four: the system as an API
With an MCP server and machine-ready docs, agents don't just reference the design system, they assemble correct, on-brand, accessible UI from it. The design system stops being a library and becomes infrastructure.
The conviction predates the hype: at Virgin Media O2 I built a token pipeline where a designer's change in Figma automatically translated to code and shipped to the development environment. The tools changed; the thesis didn't.