Our design collection has twelve “directions”: complete concept websites a business can open, personalize, and say “build me this.” They looked like good demos. Good demos don’t sell. So we rebuilt all twelve in one working session with a team of three AI models: one architect and two builders, Model A and Model B, working in parallel on the same codebase. This is how we ran it, what each model was actually good at, and why we aren’t telling you which models they were.
The bar: 90% of the site you want
A concept site earns its keep when a med spa owner, a roofer, or a restaurant opens it and thinks “that is nearly my website.” The old pages were one template recolored twelve times: a hero, three text cards, a contact block. The new standard was a real business website: a video hero, an about section with a portrait, services with images, a gallery, a process, reviews, an FAQ, hours, and one interactive section built around how that business actually sells. A reservation desk for the steakhouse. A chair-booking flow for the barbershop. A finish selector with a square-footage estimate for the epoxy installer. A troubleshooter for the HVAC company.
You can open every one of them in the design collection. Each is clearly labeled as a fictional concept with synthetic imagery, because it is.
Step one: the architect builds the system, not the pages
The most important work happened before either builder wrote a line. Our architect model built the foundation every page would stand on:
- A content contract. One typed file per concept that controls palette, typography, hero style, section order and copy. A new site is a new file, not a new codebase.
- Shared, token-driven sections. Hero, gallery, process, reviews, FAQ, visit, footer and an inquiry drawer, all themed from each site’s tokens.
- One finished reference site. The med spa concept was built end to end to the final quality bar, so the builders had something concrete to match instead of adjectives to interpret.
- A written brief. Which files each builder owned, the quality bar, how to verify, and the honesty rules.
Step two: media, generated and labeled
Twelve premium sites need real inventory, and stock photography would have made them look like everyone else. We generated it instead: 11 new cinematic hero videos (the twelfth reused an existing clip), 86 stills, and 4 before-and-after pairs made by editing the original image, so the two frames line up pixel for pixel in a comparison slider. Every hero video loops seamlessly because the last second crossfades into the first. Total media spend, including a few generations we threw away: $29.49.
One lesson worth sharing: image models carry real-world brand knowledge. Our “before” photo of a dusty sports car came back with a manufacturer’s badge on the hood, and asking again did not remove it. We retouched it out. If you generate imagery for a business, review it for logos the same way you would review copy for claims.
Step three: two builders, one brief
Model A took eight sites, split across two agents. Model B took three. Same brief, same reference site, same assets, same codebase, with each builder restricted to its own files so nobody overwrote anybody. The architect then reviewed every page at desktop and phone widths and integrated the work. From the first line of the brief to the production push took about two and a half hours.
| Scorecard | Model A | Model B |
|---|---|---|
| Pages built | 8 (two agents working in parallel) | 3 |
| Wall-clock time | 59 and 91 minutes per agent | About 30 minutes |
| Usage | About 1.4 million tokens across both agents | About 3% of one week’s plan allowance |
| Design or copy revisions we made | None | None |
| Checked its own pages visually | Yes, in a browser at several widths | No; worked from code alone |
| Problems found in our shared code | Five, all fixed before launch | Patched around one |
| Strongest at | Atmosphere and ambitious interaction | Speed, discipline and careful copy |
What surprised us
Neither model needed a design revision
We expected to spend the afternoon fixing layouts. We didn’t. Every page from both builders shipped without a design or copy change from us. That isn’t magic; it is the reference site and the brief doing their job. When the target is concrete, strong models hit it.
Both tripped on the same bug, and that was the gift
Working separately, both builders noticed that one accent color in our shared stylesheet was too faint for small text on light pages, and both patched around it inside their own sections. Two independent workarounds for the same problem is a clear signal: the bug was ours. We fixed it once, in the foundation, by computing a readable version of each accent automatically, and deleted both patches.
Model A reviews like an engineer
Model A checked its own pages in a browser and came back with five real issues in the architect’s shared code, from a contrast rule that chose the weaker text color to stylesheets that loaded on every page instead of only where needed. All five were fixed before launch. Its best pages, the steakhouse and the barbershop, are the most atmospheric in the collection.
Model B is fast, frugal and careful
Model B finished its three sites in about half an hour, for a small fraction of its plan, and wrote the most legally careful copy of the whole project: lifespans described as industry ranges and “not a guarantee,” response times phrased as goals, a ZIP-code checker that tells you it is a preview. It worked from code alone, so the architect’s visual review was its only visual check. The pages held up anyway.
How we will build client sites with this
- The architect sets the system. Tokens, shared sections, one finished reference page, and a brief with file ownership and claim rules.
- Model B builds the bulk. Interior pages, service pages and location pages, quickly and cheaply, to the reference.
- Model A takes the showpieces. The pages where atmosphere and interaction carry the sale, plus a second set of eyes on the foundation.
- The architect reviews and integrates. Every page at desktop and phone widths, with automated checks for broken assets and unsupported claims.
- A human approves. Nothing reaches your domain without sign-off.
None of this changes what we promise clients: your scope, deposit and content come first, and the timeline in your proposal is the timeline. What it changes is how much craft fits inside it.
Why we are not naming the models
Partly because the leaderboard changes every few weeks, and a post that names this month’s winner ages badly. Mostly because the model was not the decisive factor. The reference page, the brief, the ownership rules and the review pass are what made two different models produce work that looks like one studio made it. Swap either builder for next quarter’s model and the system still holds. That system is what we sell.