You wrote in confused about something that felt like a bug: you generated a restaurant site for the new place, didn't love it, regenerated it, and got something that looked suspiciously like the same site with the sign changed. Terracotta walls, cream cards, a serif logotype that could belong to any bistro from Portland to Prague. You asked, reasonably, whether the tool was just recycling one template under the hood.
It isn't, and I want to actually answer the question instead of reassuring you past it.
What you ran into is real, and it's not particular to your build — it's a known failure mode of anything that generates from a large training distribution without being told to do otherwise. Ask a model to design a restaurant site with no further direction and it reaches for the statistical middle of every restaurant site it's ever seen, and the middle is terracotta and cream, because that's what dominates the distribution. Same story for SaaS landing pages (centered gradient hero, three rounded feature icons) or crypto dashboards (the one indigo-to-violet gradient every model seems to agree on). It's not that the model can't do something else — ask it to design like a 1970s ski resort brochure or a Soviet transit map and it'll usually pull it off, often well. It's that nothing forced a decision, so it defaulted to the average.
What actually changed under you
The reason your second attempt should have looked different — and if it didn't, that's the part worth us fixing directly rather than you working around — is a pass we call the design director. It runs before the builder writes any markup, and its only job is to make choices the builder can't wriggle out of later: a palette with a point of view, specific type, a mood stated plainly enough to be checked against. Concretely, that's:
- Five to seven named colors with roles attached — not "primary/secondary" but "the color reserved for danger states," "the color that only shows up on hover"
- Two typefaces, each with a reason it's there
- Two or three sentences of mood language, the kind an actual art director would pin to a board
The test we hold a brief to is whether it can be disobeyed. "Warm and inviting" can't be — there's no way to violate it, which means there's no way to enforce it either, and the builder drifts straight back to terracotta-and-cream regardless of what the brief says. "Warm hand-drawn autumn palette, ink outlines on cream paper, nothing glossy, no drop shadows" is a real commitment, and a glassy gradient card rendered against that brief is visibly wrong, which is exactly what makes it useful.
If your build came back generic, the honest possibility is that your brief in the prompt was underspecified enough that the director had nothing firm to commit to. Naming a reference — a place, an era, a material, even just "no beige" — gives it something to push against.
Why I trust this and not just the screenshots
I'll tell you what convinced me, because it wasn't eyeballing three nice-looking examples side by side — that's how you talk yourself into a win that isn't there. We ran a few hundred builds per category, before the director existed and after, and measured pairwise color distance between same-category builds using Lab-space delta-E rather than raw RGB, because RGB will insist two similarly-saturated blues are far apart when they read as identical on screen. Restaurant sites were the worst offenders pre-director — the overwhelming majority clustered inside a tight delta-E band, all reaching for that same terracotta.
I won't oversell it to you either. Color distance measures "looks distinct," not "looks good," and it says nothing yet about layout or type variety, which we haven't instrumented nearly as carefully. A palette can be maximally distinct from its neighbors and still be ugly. We think distinct-and-occasionally-odd beats safe-and-uniform for what you're building, but that's a judgment call, not a proof.
One thing to watch for on your end
There's a failure mode that runs the other direction, and it's worth flagging before you hit it: sometimes "clean and modern" isn't underspecification, it's the actual request. Someone building an enterprise SaaS dashboard often wants the boring pattern on purpose, because their users expect it and deviating from it generates support tickets, not delight. A director that's too eager to differentiate will occasionally hand back something more art-directed than you asked for. We weight your explicit constraints above the director's instinct to differentiate, so "boring and enterprise" should get you tasteful boring rather than a rebellion against your own brief — but if you already have brand colors, a logo, reference photos of the space, put them in the prompt or attach them directly. Locked inputs beat the director inventing its own direction every time, and it'll treat them as constraints, not suggestions.
For what it's worth, this matters even more if you ever build one of our games instead of a site — a game has far more surface for a director to leave undirected than a landing page does, since the brief there has to lock sprite style, animation feel, UI chrome, and the audio palette together in one pass (more on the audio side specifically in Sound that isn't beeps), because inconsistency inside one build reads as unfinished in a way that inconsistency across two unrelated builds never does.
