You’re about to read about a hands-on experiment that tries to answer a simple question: can AI be the language you use to make a game without driving you crazy with technical details? Hugging Face publishes a piece about VibeGame, a web engine designed so models like Claude Code can “vibe code” simple games and keep the project manageable as it grows. (huggingface.co)

El problema que aparece cuando la IA empieza a ayudar

At first everything works: you ask the AI to generate code, you see results, and you get excited. What happens when the project grows? The model’s context window fills up, the AI loses track of what’s important, and coherence breaks down. That’s the starting point of the article: context management is the bottleneck to using AI as a “language” in larger projects. (huggingface.co)

What’s the typical solution? Ad-hoc shortcuts in project files or complex tools to manage context. The author built a lightweight tool called Shallot to automate loading and updating context during conversations with Claude Code. It’s a pragmatic approach: not everything needs enterprise architecture; often order and good abstractions are enough. (huggingface.co)

¿Qué es "vibe coding" y por qué importa?

Vibe coding comes from an Andrej Karpathy tweet and, in practice, means letting the AI act as a high-level language. It’s not magic: you write instructions at a conceptual level and let the model generate the implementation. Sounds irresponsible? Yes and no. It works well if you keep the project within what the AI can “understand” and you’ve defined clear boundaries. (huggingface.co)

Exploración: tres plataformas, tres resultados

The author tried three paths to make games with AI:

• Roblox MCP: great abstraction and ready-made mechanics, but lacks file management and is a closed ecosystem. Ideal if you work inside Roblox, less so for open projects. (huggingface.co)

• Unity MCP: full access to the file system, but too many ways to do the same thing and changing versions that confuse the AI. It requires deep engine knowledge. (huggingface.co)

• Web (three.js + rapier + bitecs): the AI performed better here, probably because of the volume of web data in its training. It’s open and controllable, but low level: you almost end up building an engine before building the game. (huggingface.co)

Practical result: the web won for the model’s ability to understand and generate code, but we needed an abstraction layer that kept the project light and understandable to the AI. (huggingface.co)

La solución: VibeGame, un motor declarativo pensado para IA

VibeGame is the answer: a high-level declarative engine on top of three.js, rapier and bitecs. The idea is to combine Roblox-style abstraction with the openness of the web. It includes:

• XML-like declarative syntax to describe the world and objects, which feels familiar to models trained on HTML/CSS.
• Entity-Component-System architecture to keep code modular and scalable.
• Documentation designed to be included in the model’s context (file llms.txt) to improve the AI’s coherence. (huggingface.co)

A minimal example in the article shows a scene with a ground and a ball using syntax like this:

<world canvas="#game-canvas" sky="#87ceeb">
  <static-part pos="0 -0.5 0" shape="box" size="20 1 20" color="#90ee90"></static-part>
  <dynamic-part pos="-2 4 -3" shape="sphere" size="1" color="#ff4500"></dynamic-part>
</world>

<canvas id="game-canvas"></canvas>

<script type="module">
  import * as GAME from 'vibegame';
  GAME.run();
</script>

That level of abstraction makes it easier for a model to generate simple scenes and mechanics without losing context. (huggingface.co)

¿Funciona en la práctica?

Yes, with caveats. The author built an incremental grass-collection game that demonstrates that for basic physics and rendering mechanics, the AI can deliver useful results with minimal technical experience. But when you ask for more complex systems — advanced interaction, inventory, multiplayer — the engine still falls short. In other words: it works for what’s already defined in the engine, and it requires the creator to know the limits. (huggingface.co)

Pruébalo y aprende jugando

If you want to experiment, there’s a demo on Hugging Face that runs in the browser and you can create a local project with the recommended command:

npm create vibegame@latest my-game
cd my-game
npm run dev

The author suggests pasting the contents of llms.txt into CLAUDE.md or into your context management system so the AI has good documentation from the start. That improves collaboration between human and model. (huggingface.co)

¿Y qué sigue para VibeGame?

The roadmap is clear: add more ready-made mechanics (interaction, inventory, multiplayer, animations, audio) and improve onboarding for novice users with templates and examples. There’s also an interesting idea: create high-level editors on top of established engines to leverage their power without losing the ease of use the AI needs. (huggingface.co)

Reflexión final

Is VibeGame the revolution that will let anyone create a game with AI in a couple commands? Not yet. But it’s an honest step toward a workflow where AI acts as a high-level language, and where the key is not to replace the developer but to elevate them: less dealing with routine details and more designing experience.

If you’re curious to try, check the GitHub repo and the Hugging Face demo to see how abstraction, context, and models mix in an experiment you can already touch. (huggingface.co)