# The gaps we never wrote down

Vercel's design.md didn't just improve their AI output — it measured how much of their design system was never documented.

Vercel loaded six test pages into an agent without constraints. The model generated something that looked like a generic SaaS dashboard — wrong hierarchy, wrong type roles, wrong evidence placement. They shipped a 4,000-word design spec and ran the same tests. Failures dropped from 91 to 39.

The number that matters isn't the drop. It's the 91.

Ninety-one mechanical design failures across six pages, in the absence of written rules. That's not a measurement of what models don't know. It's a measurement of what a design system had never actually documented.

Design docs were written for people. People fill in gaps — from experience, from osmosis, from asking someone. A model can't do that. When you write rules it can follow, the gaps appear exactly where they are.

Which might be the most useful thing models do for design systems: not generate pages, but show you precisely where the system ends.