AI-First
In this article:
- 🔑 Claude Code generated a complete Pokémon-like game autonomously, with 25 unique playable animals, from a single voice instruction.
- 🎯 All assets are produced by code: sprites via Canvas API, music and effects via Web Audio API, zero external files.
- 💡 The self-test clause in the prompt forces Claude to validate end-to-end before returning, ensuring a game that actually works.
- 🚀 A video game orchestrates rendering, game loop, audio, AI, states: it's a major stress test for an AI agent's coherence.
- ⚠️ Observed limitations: overflowing text, unclear capture mechanics, missing polish, but the structure and mechanics hold up solidly.
I asked Claude Code to create a Pokémon-like game with a single instruction. A complete browser game, with 25 playable animals, all assets AI-generated, music included. No files provided, no template. Zero. And it did it.
The result isn't perfect. Text sometimes overflows the dialog boxes, instructions are a bit lacking. But the game runs. You pick your animal, you fight, you catch creatures, you visit the healing center. Everything works. For a first attempt at this scale, it's genuinely impressive.
The instruction that started it all
Here's the exact instruction, dictated by voice via Typist:
Claude, generate a basic Pokémon game where all the Pokémon are animals. At least 25 unique animals. No assets will be provided, you must generate everything on the fly: music, sound effects, illustrations, sprites, everything. The game must run in the browser. Host it locally so I can play it. Test it yourself, don't come back until you've verified it works completely, with no bugs.
That's it. One instruction. Claude Code then worked autonomously: generating sprites with procedural algorithms, composing music in JavaScript, creating sound effects via Web Audio API, building combat mechanics, handling captures, healing centers, stats for each animal.
What's remarkable is the quality clause built into the prompt: "test it yourself before coming back." Claude Code honored it. It launched the local server, simulated play sessions, verified interactions. Not a rushed, sloppy delivery, a genuine end-to-end validation.
What Claude actually generated
No imported assets. Everything is produced by the code itself. Here's what that means in practice:
Element | How it's generated | Technology used |
|---|---|---|
Animal sprites | Procedural drawing pixel by pixel | HTML5 Canvas API |
Background music | Algorithmic audio synthesis | Web Audio API |
Sound effects | Generated on the fly per action | Web Audio API (oscillators) |
Stats | Calculated based on animal archetype | Pure JavaScript |
Attacks | Named after real characteristics | Parametric logic |
World map | Rendered with generated tiles | Canvas API |
25 distinct animals, each with a unique sprite, its own attacks, its base stats. The wolf is the starting choice in the demo. You then encounter a fox, a bear, more exotic creatures. Every encounter is different.
Why this demo changes things
There's a lot of talk about "vibe coding" right now, but most examples stay basic: a UI component, an automation script, a small tool. This is a full video game, with all its interlocking systems.
A game is a complex technical stack: graphics rendering, game loop, state management, data persistence, multi-layered audio, rudimentary enemy AI, responsive user interface. Claude Code orchestrated all of that simultaneously, with no manual guidance between steps.
This is exactly the same principle I had explored with AutoDream and persistent memory: Claude Code can maintain complex context and produce a coherent system, not just an isolated function.
A video game is a good stress test for an AI agent: it demands coherence across dozens of interdependent systems. If it holds, that's a strong signal.
The limitations I observed
Let's be honest. A few friction points in the demo:
Attack text sometimes overflows outside dialog boxes
No clear instructions at startup, the interface is left to guesswork
The capture mechanic isn't intuitive
Visual quality of sprites is minimalist, very basic pixel art
These are polish issues, not structural ones. The distinction matters. The engine works. The mechanics are there. What's missing is polish, UX details, help text. That's a second pass, not a full refactoring.
In traditional development, a prototype like this would take 2 to 3 weeks for a single person. Here, we're talking a few dozen minutes of unsupervised machine time. The next version with a more precise UX prompt would probably be playable by anyone.
What this means for indie game dev
Indie game dev is changing profoundly. Historically, a solo developer needed to master: code, pixel art, music composition, level design, sound design. Five distinct skills to make a game.
With an agent like Claude Code, the constraint shifts. It's less "can you do this?" and more "can you ask for it correctly?". As I had explored with cloning a 100,000-euro website: creative direction takes priority over technical execution.
Tomorrow's developer may not code 25 sprites. They write the brief, validate the prototype, iterate on gameplay, refine the experience. The agent handles the rest. It's a redistribution of tasks, not a replacement.
The real skill in 2026 is knowing how to write a brief as precise as a technical spec. Not coding it yourself.
My verdict
What I saw in this demo is a clear signal: AI agents can now tackle complex, multi-system projects without constant supervision. The game isn't a masterpiece, but it's a game that runs, with content, in a matter of minutes.
The next step is obvious: refine the prompts, iterate with Claude on visual bugs, add a save system, enrich the world. Each improvement is an instruction, not a day of coding. That's the new reality of assisted development.
If you've had a game project sitting in a drawer for years because you can't code, now you're out of excuses.