A cozy mystery game where every case is an AI puzzle. Set in Maple Hollow, a small town that’s gotten enthusiastic about AI without quite understanding how it works.
The problem
Most “learn AI” products are either dry online courses or shallow chatbot demos. There’s a real gap: nobody is teaching the craft of working with modern AI, the small careful skills that make AI useful in practice. This project is an attempt to fill that gap, framed as something people actually want to play.
What I’m building
A browser-based puzzle game with 15 levels, each one a self-contained “case” where the player investigates why a small-town AI system is misbehaving. The mechanics teach real concepts like system prompts, retrieval-augmented generation, tool use, agents, evaluation, bias detection, all without ever feeling like coursework.
How it works
The engine is a deterministic simulated-LLM written in Python. No real model inference happens. Each level’s behavior is defined by a rules-based pattern matcher that produces predictable, authorable AI behavior. The engine runs in the browser via Pyodide, so the entire game is a static site with no backend.
The frontend is React, with a three-panel workshop interface — configuration on the left, conversation in the middle, results on the right. A “trace inspector” lets curious players see exactly why the simulated AI did what it did, turning every interaction into a potential teaching moment.
Status
In active development. The engine and the first cases are playable now and you can try the demo above, or open it full screen.