Perfect in code, misread by AI

Your design system can be perfect in code, and an AI will still misread it. I know this problem by heart. Most people think it comes from a lack of rigor, that you need to spec better, document better. That's wrong, and I see it on every engagement.

Since Config 2026, everyone has been plugging an AI into their design system. And many are finding that the output goes off track, without understanding why. Cristian Morales Achiardi just wrote an article that puts precise words on it.

Code has guardrails, Figma doesn't

On the code side, types, tooling and the framework constantly check that everything stays consistent. A dev can't drift too far. But those guardrails stop dead at Figma.

Nothing watches the design side. A designer who names a variant slightly differently from the code isn't held back by anything. Each side moves forward in its own corner, and the gap widens screen after screen.

The AI reads the unguarded side, and amplifies it

When you plug an AI into that, it reads this unguarded design side. It picks up the gaps, the approximations, and it amplifies them. That's why what it produces goes off track: flawless code, a clean Figma file, and yet the two don't quite speak the same language, because nothing keeps them aligned with each other.

Better specs don't change anything

The real issue is elsewhere: you need something that keeps Figma and code aligned continuously, otherwise the gap comes back on every screen, and the AI repeats it.

That's where I spend a lot of my time right now: making sure design and code point to the same reference, so that humans and AI alike read the same thing.

A design system, at heart, isn't a box of components. It's a shared language. And the whole game is making it hold on both sides.