Claude builds an entire underwater world in just 97 KB, no game engine needed
A developer used Anthropic's Claude Fable 5.2 to build an explorable underwater world that fits entirely into a single 97 KB HTML file. It contains no images, 3D models or sound files. Everything is generated from code when it loads. The world runs right in your browser, and anyone can try it.

What's going on
A developer has used Claude Fable 5.2, Anthropic's new model, to build an underwater world you can swim through in your browser. It's called "Abyssal Garden". There are kelp forests, a shipwreck, a stone whale tail, manta rays, schools of fish and sunbeams streaming down from the surface. All of it lives in a single HTML file of just 97 kilobytes. For comparison, one phone photo is usually 30 times that size.
The project first appeared in the Claude forum on Reddit. Wccftech and BigGo name the developer as "Noniv", which is also the name of the GitHub account. XDA Developers gives the Reddit username as "Any-Reputation8118". The demo spread widely via Japanese design blog Coliss, XDA, Wccftech and BigGo. The source code is on GitHub under the permissive MIT license, so you can use it for free, even commercially: github.com/Noniv/abyssal-garden.
All code, no files
What makes it special is what's missing. A 3D scene normally consists of lots of files: textures (surface images), 3D models, sound files, plus an engine, a ready-made graphics framework like Three.js, Babylon.js or Unreal. Abyssal Garden has none of these. Everything is procedurally generated at load time: the code assembles the world itself from rules and formulas instead of loading prebuilt assets. Think of it as sending the recipe instead of the finished cake.
According to the project's GitHub page, Claude wrote:
- WGSL shaders. Shaders are small programs that run directly on the graphics card and define how light, colors and surfaces look. WGSL is WebGPU's shader language.
- a seafloor sculpted on the GPU,
- around two million kelp blades and plants, tens of thousands of rocks and corals and thousands of sea creatures and decorative objects,
- sunlight falling through the water's surface, animated caustics (the dancing light patterns you know from the bottom of a swimming pool), volumetric-looking light shafts and floating particles,
- a soundscape generated in real time through the Web Audio API, a browser interface that lets you synthesize sound right in the browser.
The world feels physical, too. According to Coliss, the seafloor is solid rock, and if you swim into a wall, you slide along it instead of getting stuck. You can swim through every tunnel and arch. Rendering runs on WebGPU, a new browser interface that gives web pages direct, fast access to the graphics card.
How the demo came together
According to the developer, the first version came from a single, detailed prompt to Claude Code, Anthropic's command-line coding tool. You can read the prompt in the GitHub repository. It set requirements for visual quality, procedural generation and a steady frame rate, and asked for a self-contained HTML file as the result. Engines like Three.js and Babylon.js were explicitly banned.
The developer also built in quality control from the start: a separate AI subagent, a second Claude instance with its own task, checked screenshots of the rendered scene for graphics glitches. Meanwhile, GPU timing measured performance during a scripted 60-second flight through the world. In other words, the AI checked and fixed its own work.
It wasn't done in one pass, though. Follow-up prompts added collision detection, a start menu and adjustable graphics presets. The developer estimates Claude's actual compute time at two to three hours, spread over about three days because the subscription only allows limited usage.
Performance and trying it yourself
| Feature | Detail |
|---|---|
| File size | 97 KB, one HTML file |
| Performance (developer's figure) | approx. 180 FPS at 2560 × 1440 on an RTX 5070 Ti |
| Low-end smartphone | approx. 20 FPS on a Samsung Galaxy A16 (Low, reduced resolution) |
| Tech | WebGPU, WGSL shaders, Web Audio API |
| Not used | Three.js, Babylon.js, textures, 3D models, audio files |
| AI compute time | approx. 2–3 hours over approx. 3 days |
| License | MIT |
Nobody has independently measured the frame rates yet, as Wccftech and BigGo point out. There are three detail levels. According to Coliss, High is meant for computers, Medium for smartphones with 8 GB of RAM and Low for smartphones with 4 GB. The world is the same at every level. Only lighting, shadows and object density drop.
There's nothing to install, and the demo even works offline. Open the project website, or download index.html from GitHub and open it locally. Chrome and Edge are recommended. Coliss also got it running in Safari on macOS. If you see a message saying WebGPU is required, switch to Chrome or Edge. Controls:
- Mouse: look around
- W A S D: swim
- E / Q: up and down
- 1–6: switch camera
- M: sound on/off, H: HUD, Esc: menu
Hit "Dive in" to go under.
The bigger picture
Abyssal Garden is an impressive graphics demo, but it isn't a finished game. There are no quests, no goal and no game mechanics beyond swimming. Procedural graphics aren't new either. The demoscene has been cramming entire worlds into tiny programs for decades, often just 4 or 64 KB, and by hand. What's new is that a language model largely programs a scene like this from a written description, handling optimization and debugging with its own review loops.
BigGo especially highlights the decision to skip an engine. An earlier much-discussed Claude Fable 5.2 demo, a shooter in the style of Gears of War: E-Day, still used Unreal Engine 5. BigGo also finds the pattern of one AI agent checking another's output interesting, since it could make graphics-heavy development cheaper. For Anthropic, a project like this is good advertising too, especially as competition among AI vendors increasingly revolves around agentic coding.
Some questions remain open. The performance figures come only from the developer, and it's hard to judge from the outside how much manual work was actually needed between prompts. There's still a long way from a pretty demo to a market-ready product. Even so, Abyssal Garden shows pretty concretely what AI coding tools can now pull off in a few days.
Sources
- Hatena Bookmark IT (JP) coliss.com
- wccftech.com wccftech.com
- xda-developers.com xda-developers.com
- finance.biggo.com finance.biggo.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.