The design system the AI actually reads
The design system the agent applies isn't a Figma file or a brand PDF. It's three text files, and it reads them better than I do.
For years my design system lived in Figma: components, variants, a brand PDF nobody opened twice. When I started building with an agent, I found out it barely reads any of that. Images are opaque to AI. What it eats up is text.
Three files, not a library
The system the AI applies fits in three text files:
tokens.md— the usage dictionary. Padding, shadow, typography, spacing, and when to use each one.tokens.css— the real implementation, so the code matches the manual instead of drifting from it.decisions.md— the brain. The rules I don't want to repeat in every prompt.
The decisions.md is where the taste lives
Anyone can export tokens. What makes the AI sound like me are the decisions written in plain text: "prefer a contextual progress bar", "always an orange arrow on the action button", "square corners, never rounded". Without that, the agent hands back the average look of the internet, that generic AI slop. With it, it applies my own bar on a project it has never seen.
In this portfolio
This site runs exactly like that: a DESIGN.md with the decisions and a tokens.css with the tokens. When I ask for a change, the agent reads both before touching a single line of CSS. The result is born inside the system, not a patch glued on top afterwards. The difference isn't speed, it's coherence: the taste is recorded in a place the machine reads every time.
// Takeaway The design system became a prompt. Whoever writes the decisions.md teaches the AI to have your taste, and that's the part you can't outsource.