~/blog/ai · June 14, 2026 · 5 min read
I Told an AI to Build Me a Pokémon Game. It Did.
#ai#gaming#claudecode#tutorial#agentic-ai
I’ve been a casual gamer my whole life. I’ve also been a dad who talks about tech for a living. So when I heard people say AI coding assistants could “build anything,” I figured I’d test that claim in the most ridiculous way I could think of.
I told it to build me a Pokémon game.
Not a demo. Not a proof-of-concept clone. A real, playable RPG — with its own region, gym leaders, wild encounters, a starter selection screen, a working economy, and a battle engine. The kind of game that takes a team of developers months to build.
I wrote zero lines of code myself. Here’s what happened.
The Setup
The tool I used is called Claude Code — Anthropic’s agentic coding assistant. It’s not a chatbot in the traditional sense. It doesn’t just answer questions. It reads your files, writes code, runs commands, debugs errors, and iterates — autonomously — based on what you tell it in plain English.
I gave it a one-liner: “make a Pokémon-style browser RPG.”
Then I just kept talking.
What It Built
The final game — called The Mosaic — is a fully playable browser-based Pokémon RPG. Here’s what that actually means in practice:
A real battle engine. Under the hood, every fight runs on Pokémon Showdown — the same battle simulator used by millions of competitive players. Accurate damage calculations, type matchups, status effects, switching mechanics, all 1025+ Pokémon with their actual moves and stats. I didn’t implement any of that. The AI figured out how to wrap it.
A full 8-gym region. Fourteen interconnected maps stretching from Verdant Hollow all the way to Shadowmere — towns, routes, biome-themed wild encounters, Pokémon Centers, shops, and NPC trainers on every route. The maps are big. Route 1 alone has tall grass patches spanning the whole thing.
Randomized starter selection. Every new game is different. The server picks three random Pokémon from a pool of 37 three-stage evolution lines and guarantees they don’t share a type. You get a slick animated UI to choose between them. This was a one-sentence request: “every game should be unique, have any 3 Pokémon with 3 evolution stages.”
An adaptive AI opponent. Gym leaders and trainers use a custom decision brain that scales to your progress. Early gyms make cautious plays. Later gym leaders are aggressive, draft type-counter teams, and punish you for bad switches.
A full save system. PostgreSQL-backed saves with Docker Compose deployment. You can continue from the title screen, save anywhere, and it all persists.
A title screen, Pokédex, Trainer Card, Region Map, Shop UI. Premium-looking screens, dark glassmorphism aesthetic, animated particle effects. These were all built in a single session from nothing.
How the Conversation Actually Went
This is the part that I think is important to understand. I wasn’t typing code prompts. I was having a conversation.
Some of my actual inputs:
- “sprites dead on docker”
- “i notice when entering maps, player is entering random places”
- “some are not visible, also ensure not same type in 3” (with a screenshot)
- “starter comes at level 5, route 1 is really hard”
- “create full game and story and maps.. routes.. make big routes”
That’s it. That’s the prompting. The AI read the error logs, found the root cause (Nginx wasn’t serving the sprites-master folder), fixed the Docker config, and rebuilt the container — without me asking it to do any of those individual steps.
The type overlap fix — where it checked for duplicate types before showing starters — it did by querying the Pokémon Showdown data layer directly. I said “no same type” and it figured out how to enforce that from the existing codebase.
What This Actually Means
I want to be honest about something: this isn’t the future of gaming. Professional game developers have nothing to fear from a single dad prompting an AI on a Saturday.
But here’s what it does mean.
The gap between “I have an idea” and “there’s a working thing on my screen” has collapsed in a way I didn’t fully believe until I lived it. The conversation I had with Claude Code over a single session produced something I genuinely play. My kids play it. It has bugs. Some screens have rough edges. The overworld movement could be smoother.
But it works. And I didn’t know TypeScript three months ago.
The skill that matters now isn’t knowing how to write a database query or configure an Nginx reverse proxy. It’s knowing what you want clearly enough to describe it. That’s a different kind of literacy — and it’s one that anyone can build.
The Stack (Since You’ll Ask)
For the curious:
| Layer | Tech |
|---|---|
| Battle engine | Pokémon Showdown (via Node.js) |
| Backend | Node.js + Express + TypeScript |
| Frontend | Vanilla TypeScript + Vite (no framework) |
| Database | PostgreSQL 15 |
| Deployment | Docker Compose + Nginx |
| Sprites | PokeAPI sprite repository |
| 3D assets | KayKit Adventurers + Kenney Fantasy Town Kit (CC0) |
The whole thing runs on docker-compose up --build and opens in a browser on localhost:8080. No installation, no setup, no accounts.
Try It Yourself
The source code is on GitHub: github.com/dadwritestech/the-mosaic
Clone it, run Docker, and pick your starter. The three you get will be completely different from mine.
This is v1. There’s a lot more I want to add — proper sound effects, move animations, a full Elite Four sequence, mobile controls. But that’s the thing about building this way: the gap between “I want that” and “it exists” keeps getting shorter.
Pokémon and all related names are trademarks of Nintendo, Game Freak, and The Pokémon Company. This is a non-commercial fan project built for educational purposes.