What they did
The poster borrowed the prompt structure from an earlier underwater-world project and aimed it at a Brazilian running joke: a light blue 1996 Fiat Uno with an aluminium ladder strapped to the roof. The prompt ranked priorities for conflicts (visual quality, likeness to attached photos of the real car, frame rate, file size), banned engines, physics libraries, textures and models, and set a GPU budget of 11 ms average with no frame over 16.7 ms on a scripted 60-second lap. Done meant three checks passing: a fresh critic subagent each round that saw only fixed camera shots, lap frames and the reference photos and graded defects; the GPU timing; and a test that pressed every key. A second prompt added menus, gamepad support, a dashboard, a benchmark button and synthesised engine and crowd sound.
How it went
Fable 5.2 on high effort ran for 7 hours 12 minutes over 8 critic rounds and produced a 603 KB index.html. It never met its own finish line: the critic still listed 4 major defects when the poster stopped it, and the run would have cost about $348 at API prices. The same prompt on Opus 5.5 took 5 hours and 11 rounds, made a 341 KB file for about $88, and stopped itself saying the job was not finished. Players in the thread said it drives well, including on a controller.
Worth knowing
A critic that never sees the code can keep a loop running for hours, so set a round or spend cap before starting. It needs desktop Chrome or Edge with WebGPU. The post does not name the harness that ran the model. The prompts are in the repo, in Portuguese, and the code is free for non-commercial use.