This design system started like many others: with a file downloaded from Figma Community. That's not a flaw in itself, and above all, it doesn't show. As long as the team works by hand, everyone compensates for the file's quirks without thinking about it. The day an AI reads it to generate a screen, it shows right away.
One component, five axes, 160 variants
The component I'm working on right now is a ChoiceCardGroup, a list of selectable cards. It has 160 variants.
Nobody on the team built it that way, and that makes sense: a Community library is made to suit everyone. So it covers every possible case and stacks the axes. This one carries five (Selected, Size, Type, State, Breakpoint), and each axis multiplies all the others. Meanwhile, the file hasn't changed in years, while the product has.
What it costs when it's time to generate
Variant names describe geometry. Size=md or State=Hover say nothing about when to pick them. So the agent picks by how closely the words match the request. On a hover state, the mistake has no consequence. On the rest, it does.
The real problem was elsewhere: the Type axis mixed two questions that have nothing to do with each other. How you choose (radio for a single answer, checkbox for several), and what you show in the card (icon, avatar, illustration). The result: a multiple selection with avatars simply couldn't be expressed in the system.
Faced with a combination it can't express, the agent doesn't stop to ask. It detaches the instance and rebuilds the card by hand. That's how a design system gets bypassed without anyone noticing: not through rejection, but through a silent pile-up of exceptions.
V2: the complexity changes nature
V2 fits in 4 variants, orientation and size. The card became a component in its own right, carrying its own states, and the group takes as many cards as needed.
The complexity hasn't gone away, it has changed nature. It moved from variant axes, which the agent has to guess from a list of names, to instance properties, which it fills in. That's the whole difference between a system an AI has to put up with and a system it can actually use.