
A complete first-person shooter — enemies, weapons, particle effects, sound, a mobile version — arrived on a Chinese developer forum this week as a single HTML file of roughly 432 KB. No installer, no asset folder, no build step. Double-click it and you are playing. The thread drew more than 11,000 views and 137 replies in three days, and the reaction was less "nice demo" than genuine disbelief: commenters kept asking how that much game fits into a file smaller than a single smartphone photo.
The short answer is that almost nothing in the file is a game asset in the traditional sense. The long answer explains why this kind of stunt keeps happening now — and what it actually proves about AI coding tools.
According to the discussion on LINUX DO, a forum member posted a self-contained HTML file that runs as a full 3D first-person shooter in the browser, with a mobile version included. The file reportedly contains hidden debug and showcase modes for benchmarking, enemy and weapon models, materials, props, and particle effects — all generated at runtime rather than shipped as separate files.
The numbers that made the thread take off are the ones forum members kept repeating: about 432 KB for the whole game, with no separate image, model, or audio files. For comparison, a single texture in a commercial game can be larger than that. The discussion also covered the reported cost of producing the file with AI — heavy token use — and the demanding performance on phones and weaker PCs. Several commenters joked that results like this could become a new benchmark for AI coding tools, measuring models by what they can build rather than what they can answer.
None of the specifics about file size, cost, or contents are independently verifiable from outside the forum; they are reported by the thread's participants. What is verifiable is the technique itself, and it is surprisingly ordinary once you look past the spectacle.
The trick is procedural generation, and it is older than the AI hype. Instead of shipping textures, models, and sounds, the file ships the recipes for them and cooks them up when the page loads.
A modern browser already contains nearly everything a simple 3D game needs. WebGL gives JavaScript direct access to the GPU, and Three.js — the open-source JavaScript library that wraps WebGL in a friendlier API — supplies the scene graph, cameras, lights, and materials. Three.js has been around since 2010, is MIT-licensed, and is the standard way to do interactive 3D on the web. In a single-file game, the library itself is usually inlined or pulled from a CDN, and everything else is code.
That code then generates the world on the fly:
None of this is new technology. What is new is who is writing it. A human developer could hand-write a game like this — single-file game jams have existed for years. The difference is that an AI coding agent can now produce the whole thing from a prompt, iterate on bugs it finds itself, and keep going until the result is playable.
This is not an isolated stunt. The same pattern has appeared repeatedly in recent months: a non-programmer reportedly built a complete FPS called Zombie Slayer — a 900 KB HTML file with ten levels, synthesized sound, and escalating enemy AI — using only prompts in Google's Antigravity environment. A separate hands-on test had the open-weight model GLM-5.2 generate a 1,014-line open-world survival game in a single HTML file on the first attempt — procedurally generated terrain, a day-night cycle, shelter building — at just 38.3 KB. And a Chinese blogger recently published a "skill" that teaches Claude Code to build a playable 3D action game in the browser, with a xianxia-themed demo packing models, textures, music, and voice acting into one HTML file.
The pattern is the same every time: one prompt, one file, zero asset pipeline. That matters because the asset pipeline has always been the expensive part. A small team can write game logic in weeks; producing art, animation, and sound takes months and money. When models can generate all three as code, the bottleneck moves. The file you download is no longer the product of an art department — it is the output of a long conversation with a model.
There is a second reason these demos cluster around Three.js and the browser. The browser is the only platform where "here is the game" can mean "here is a file." No store approval, no installer, no platform cut. You can email a game, paste it into a chat, or host it on a static page. That distribution model makes the demo itself the story: the wow factor is not just that the game exists, but that it arrives as an attachment.
For all the excitement, the forum discussion is refreshingly honest about the trade-offs — and they are the same ones that show up in every AI-generated game demo.
Performance is the first wall. Procedural generation saves download size, not compute. Every texture, model, and sound is built at runtime, and the game loop runs on the main thread. On a desktop with a decent GPU this is fine; on phones and weaker laptops, forum members reported, it struggles.
Token cost is the second wall. Generating tens of thousands of lines of working game code is not cheap. The thread discussed heavy token consumption, and comparable public experiments have reported figures like hundreds of millions of tokens for a polished result. The game may be free to download, but it was not free to produce — the cost falls on whoever ran the model.
Playability is the third wall. These games are real and playable, but they are not deep. Enemies follow simple patterns, levels are generated rather than designed, and the fun wears thin faster than in a hand-crafted game. A benchmark that measures "can the model produce a working game" is not the same as one that measures "can the model produce a game people keep playing."
If you are a developer watching this trend, the useful lesson is not "AI can make games now" but something narrower and more practical. Single-file web games are a genuinely good format for learning and prototyping: the feedback loop is instant, the deployment story is trivial, and the constraints force you to understand what each part of a game actually does.
A reasonable way to experiment is to start small and stay in the browser. Pick one mechanic — movement and shooting, or a day-night cycle — and ask a coding agent to build it as a single HTML file using Three.js from a CDN, with pointer-lock mouse controls and simple raycasting for hit detection. Those are the same building blocks the forum's game reportedly uses. When the mechanic works, add one system at a time. The failure mode to avoid is asking for the whole game in one prompt and then debugging a 10,000-line file you have never read.
For a sense of where this sits in the wider AI-coding landscape, the same dynamics — agents that can sustain long tasks but need supervision on quality — show up in everyday coding work too, from the apology loops agents fall into when they keep failing a task to the way plan-mode workflows in tools like Antigravity structure large changes. The single-file game is the same story with better visuals: a long task, completed, with all the rough edges still visible if you look.
A 432 KB shooter is not going to replace commercial games, and nobody in the thread claims it will. Its significance is narrower: it is a demonstration that the distance between "I have an idea for a game" and "here is a playable game" has collapsed to a prompt and a file. The assets are generated, the distribution is a download link, and the cost is measured in tokens rather than months.
That is why the thread resonated. The game itself is a curiosity; the format is the news. When a playable 3D world can arrive as an email attachment, the question stops being whether AI can write code and becomes what we do with the fact that it can.