A reader with no manners

Shared something internally today about exposing your design system to language models. Got useful pushback within the hour, which is the right response and sharpened what I think.

Here is the shift, and I think it is the most important one in our discipline right now.

For its whole life, a design system is something people consult. A designer opens it when unsure. An engineer checks a spacing value. It is a reference, and references tolerate being slightly inconvenient, because a human who cannot find the answer will go and ask someone.

Agents are now writing a serious share of our code, and this reader has none of those manners. It will not ask. It cannot open a Figma file. It cannot authenticate through the VPN. It cannot read the judgement in a senior designer's head, which is where an embarrassing amount of any design system actually lives.

So it fills the gap. It generates a plausible token name, invents a spacing scale, writes a component that looks approximately right and shares nothing with ours. Confidently, at speed, across every file it touches.

The agent is not misbehaving. It is doing what anyone does with an underspecified brief, minus the social friction that would make a person stop and ask a question.

Which means the gaps were always there. For years they were absorbed invisibly by experienced people using judgement, and that absorption is precisely why they were never fixed. Take the people out of that step and the gaps become output — visible, immediately, in a diff someone can read.

I have spent two decades arguing that under-specification in a design system is a real cost. I could never show anyone the bill. Now the bill arrives within a minute, itemised, in a pull request.

So the vision I am now working toward is concrete: our design rules, accessibility standard and writing guidelines live as plain text in the repository, next to the components, where both people and agents read them before anything gets generated. Not documentation about the system. The system.

Within a year I expect that to be unremarkable here, and I expect the teams that got there first to be visibly faster than the ones still maintaining a beautiful library nobody's tooling can read.