Build a drivable procedural car game with a visual critic subagent

A Redditor had Fable 5.2 build a drivable 1996 Fiat Uno with a roof ladder on a race track in one WebGPU HTML file; it ran 7 hours and 8 critic rounds, about $348 at API prices.

Done withClaude· Nov 16, 2026

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.

Try it

With Claude· 6 parts to fill in

Build [a drivable vehicle and track] as one HTML file using WebGPU and WGSL, with no engine, physics library, textures or models: generate everything in code. When goals conflict, rank them [visual quality, likeness to the attached reference photos, frame rate, file size]. Hit [11] ms average GPU time on a scripted [60]-second lap on [my machine]. Each round, give a fresh subagent only rendered camera shots and the reference photos, and have it list defects as major, medium or minor. Also run the timing lap and a test that presses every control. Stop when no major or medium defects remain and the budget is met, or after [number] rounds, and report what is left.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related