What does a design system have to contain so that an AI agent reliably builds the right user interface from it?
Nobody knows yet. I am a decent, battle-tested designer, but this is an enormous task. To do great work, I need clear boundaries and a defined entry point to be at my best. So instead of designing the whole factory, I start in one assembly room at a time, building one wheel: one small question. That way, everything a test proves helps the system grow.
Each question is a wheel, for now. When I move to a different assembly room, it may become a spring or a bearing. A wheel has a hypothesis written before the run and raw agent output that is never edited. It ends with findings, including the failures, and what the result means for the system.
The repo is public, so every number below links to the run that produced it. Every test is written so you can run it and check the result yourself.
With 12 written rules, Claude chose the correct button variant in 87 of 87 answers, against 63 of 87 with typical human written documentation.
With rules, 10 models chose the correct button variant in 80 to 87 of 87 answers, against 60 to 73 with human written documentation.
With the rules, models still score 76 to 87 of 87; 3 amended rules meant to fix the last 32 misses raised them to 40.
Without test buttons named as examples, rewritten button rules got 948 of 960 answers right on 10 models; the old rules partly matched labels.
The rules got 948 of 960 answers right on the 32 scenarios they were developed against. Next, 10 models get 24 buttons from real products, with my answers decided before the run.
12 hard rules for choosing a button variant, from run 3 of wheel 004, where 10 models got 948 of 960 answers right, 98.8%. Each one says what it means, which rule wins when 2 apply, and when the answer is 'none of these'. Nothing else yet.
| Since 1 October 2026 | Total |
|---|---|
| API credit, pay as you go | €24.82 |
| Subscriptions | €98 |
| Answers from models | 7,635 |
| Tokens used, in model runs and chats | 625 million |