Claude Rebuilt a 3D Scene Without Three.js, Turning the Browser Into a Cheap Lab

Simon Willison built a Fabergé egg inspired 3D object in Three.js and asked Claude Opus 5.5 Medium, in the same chat, to port it to plain JavaScript and WebGL. The original file was 77 KB and the new one 98 KB. Both run in a browser, but they are not identical.

Claude bypassed the library, not WebGL's underlying complexity

The experiment follows Paul Ford's argument that browsers already provide a vast application platform while LLMs lower the cost of reaching unfamiliar corners. Ford cites Web Crypto, WebAssembly, WebGL and Web Audio. His TinyLittleOffice web suite still cost 0 dollars to host on Cloudflare after a surge of visitors.

Willison tested a concrete edge of that claim. Claude rewrote an existing Three.js scene without the library. He found the Three.js version slightly easier to read at a glance and acknowledged that he did not fully understand why the outputs differed.

Prototyping gets faster while code ownership trails behind

For individuals, cheap distribution now combines with cheaper access to APIs that once demanded specialization. An idea can be tested as a web file without an app store, installation package or large team. An LLM can write graphics code that the person prompting it would not normally attempt.

This mostly transforms the first stage of development. A product manager or designer can test an interaction before funding a full architecture. A developer may later inherit code whose behavior the prompt author cannot explain. Prototype speed and maintainability remain separate variables.

Two similar scenes do not establish a faithful port

Willison explicitly says the outputs differ. The experiment includes no benchmark for performance, browser compatibility, accessibility or maintenance. The 98 KB file is larger, but that fact alone says little about quality. It measures one artifact.

The LLM did not remove WebGL. It absorbed part of the interface work. A graphics bug, memory leak or device difference will still require someone who understands the rendering pipeline. A polished demo makes that boundary easy to miss.

The next test should measure changes, failures and handoff

A stronger signal will come from projects that survive several rounds of modification. Watch whether a model preserves the image while changing the scene, explains its choices and supplies tests across browsers and graphics hardware.

Ford's broader thesis does not depend on one library beating another. Browsers already had the capabilities. Exploration has become cheaper. The remaining question is whether cheap experiments mature into maintainable software or multiply impressive one offs.

Lilith's verdict

The browser already contained a whole factory. Claude merely lit the corridor, and the first breakdown will reveal who can actually reach into the switchboard.