"It looks like an empty dev sandbox." That was the verdict — my own — on the design an AI model had given my website. And the model had done exactly what I asked.
The experiment was simple: I switched off the design guides and templates and let the model draw the entire public site to its own taste. No moodboards, no "make it like …". I wanted to see what it could do on its own — partly because that's how I compare different models' raw taste: same task, no training wheels, then see what comes out.
The first result was competent and dead. Clean, white, orderly — and without a single decision that felt made. Hence the verdict.
The paper test
The calibration that changed things was a single sentence: if you can print the page without losing anything, there is no experience. Scroll has to carry meaning. Motion has to say something. Decoration is banned.
With that line in hand, the model rebuilt from scratch. Today the site behaves according to the idea: raw draft compiles into finished build as you scroll, and the front page's headline can be dragged from the client's messy request to the finished phrasing with a physical dial. Print it, and you're holding a wilted piece of paper. That's the point.
Every fix became a handover
Along the way I learned something I hadn't planned on: design critique can't be shuttled back and forth between models. I tried letting one model design and another review. Every fix round became a handover of taste, and every handover pulled the design in a new direction. The result resembled neither model's work — it resembled a compromise nobody had asked for.
The rule today: one model owns the visuals from first sketch to final commit. Others are welcome to check facts — does the button work, does the contrast hold, can a keyboard reach everything — but they don't pass aesthetic judgment. Responsibility for the whole can't be split into shifts.
The division of labour: one model owns all taste · other models verify function only · the human holds the veto
The same goes for people, incidentally. A design that four people have each adjusted a little looks exactly like that. Taste is not a task that can be delegated in small pieces — not to machines, and not to each other.