Skip to content

~/writingfifteen-sites-one-prompt

Fifteen websites from one prompt: what an AI agent built on its own

I gave Claude Fable 5 one adapted prompt and let it design, build and deploy fifteen showcase sites and a hub. What it did well, and what it doesn't prove.

Jon Phillips4 min read#agents#ai#cloudflare
Fifteen small white architectural models on a dark table, one glowing green

FIFTEEN is an experiment I ran: fifteen showcase websites and a hub page, designed, built and deployed by Claude Fable 5 from one prompt I adapted. No two sites share a display typeface, a palette or a signature technique. Each has a second page explaining how it was made. I gave the agent the prompt and the tools, and it did the art direction, the code, the imagery, the testing and the deploys on its own. It isn't client work, and it doesn't pretend to be.

Where did the prompt come from?

Nick Saraev built 25 websites with Claude Fable to show what it can do with design, and shared the prompt he used in Maker Zero, his free Skool community. It asks for sites that are each fundamentally different, a /guide page on every one explaining how it was made, three iteration passes before any site is called done, and no questions until all of them are finished.

The line that sets the bar:

I want each website to be fundamentally different, and involve a mixture of advanced visual techniques … Before you “ok” each site, make sure you do at least three iteration passes.

I changed three things: 15 sites instead of 25, hosted on Cloudflare Pages instead of Netlify, and the imagery generated in Higgsfield.

How did the run work?

It ran over July 8 and 9, 2026, in two layers.

First, one orchestrator agent wrote sixteen creative briefs and a uniqueness registry, so no two sites could share a display typeface, a palette or a signature technique. Then it ran up to eight builder agents in parallel, each confined to its own site's folder under one binding contract.

Each builder logged three iteration passes and checked its pages at 375, 768 and 1440 pixels wide. Nine sites use AI imagery, about 90 Higgsfield credits in all; six use none. The six WebGL sites were checked on a real GPU. A final art-director review of all sixteen found one defect, an overlapping label at 1440 pixels, which was fixed and checked again.

The range is the point. There's a perfume house rendered in a thin-film fluid shader, a 1974-style analog synth maker with the instrument built in 3D, a type foundry where the specimen is the interface, a toy company whose website is itself the toy (physics in Matter.js), and a playable sound garden made with Web Audio. Every name, product and place in them is invented.

What broke?

Twice, the run stopped. First a network drop, then a monthly spend limit that stopped every active builder at once.

No work was lost, because each agent wrote its progress to disk as it went and could be resumed from where it left off. The lesson for any long agent job: assume the run will be interrupted, and make every agent's progress something it can pick back up from.

The hub, built last, links all fifteen sites and carries an eleven-part guide to the method.

What does it show, and what doesn't it?

It shows that one model, given clear briefs and a way to look at its own work, can finish a wide range of working design without being steered page by page: shaders, 3D, physics, generated sound and moving type, fifteen times over, with no two alike.

It doesn't show a business. The perfume house, the synth maker and the tasting counter are invented, so there were no customers, no content to move, no forms that reach anyone and no brand rules to follow. The work took about a day of agent time with little input from me. Real client work runs on the client's content, feedback and deadlines, and this run never had to meet any of them.

What's worth keeping?

The method. Lock the briefs before any code. Give every agent one written contract. Make the agents look at screenshots of their own work after each build. Set the number of passes in advance. The guide on the hub spells it out so anyone can run it.

I used the same approach afterward on my own agency site, with five directions instead of fifteen. That story is in Rebuilding matchlessweb.com as The Changelog.

You can walk through all fifteen at fable-fifteen.pages.dev.

Stack: Claude Fable 5 (orchestrator and up to eight builder agents), Cloudflare Pages, Higgsfield (imagery), WebGL and GLSL shaders, Matter.js, Web Audio, variable fonts and SVG.

Jon Phillips builds websites, web apps and automations atMatchless Web in Clinton, Mississippi.

~/writing · keep reading