On AI and design systems

The next consumer of your design system might not be a person. It might be an agent — and most design systems are unreadable to one. The docs are written for humans, the examples are screenshots, half the rules live in someone’s head. An AI tool walks in, guesses, and ships something that’s almost right and entirely off-brand.

I’ve spent a while now making ours legible to machines. Some of it is unglamorous plumbing: slimming our agent rules by about eighty-five percent so the model spends its attention on the right things, writing docs an agent can actually parse, and drawing one hard line — exactly one third-party component is allowed, the data grid; everything else is ours.

The part I find more interesting is judging the output. We built a small evaluation harness: real prompts from real product teams — build a section, justify a navigation choice, work inside a branch — scored against our own standards and run several times so the result isn’t a fluke. It turns an opinion (“the AI did okay”) into something closer to a measurement.

Here’s what surprised me. Almost everything that makes a system good for agents makes it better for people too. Clear names, honest examples, rules you can state out loud — humans were always quietly struggling with the same parts the machine can’t guess.

I don’t think AI replaces the design-systems team. I think it audits it. If an agent can’t use your system, that was probably never a model problem.