What they did
The poster wanted to test how far a current model could take a real-time 3D scene with no engine, assets or libraries. They wrote an art bible (city plan, materials, plants, water, lighting and six review shots) and made eight concept images of the city with ChatGPT, then handed both to Claude Code running Fable 5.2 with one starting prompt. Claude Code wrote everything: a WebGPU renderer, shaders for crystal refraction and mirrored canals, generators for every building and tree, collision, detail levels and a chiptune soundtrack made with Web Audio. The prompt set up its own checks: a separate subagent that saw only the art bible, the concept images and the renders listed visual faults, and GPU timing over a scripted 60-second flight tracked frame time.
How it went
The result is a ringed city of marble, canals and crystal towers that is generated on load, packed into one 96 KB file and published on GitHub Pages with the source and prompt in the repo. The poster reports about 2.5 ms per frame at 1440p on High on an RTX 5070 Ti, and the README gives about 28 fps on the Low setting on a budget Samsung phone. The post passed 390 upvotes on its first day. No cost or run time is given.
Worth knowing
This is the same poster's second build with this method, after an underwater scene, so the reviewer-subagent and timing loop is the part that carries over. The art bible and concept images gave the reviewer something concrete to judge against. It needs a WebGPU browser, best Chrome or Edge on desktop.