This repository is a side-by-side comparison of how different AI coding models tackle the exact same task. Each branch contains a version of the game built by a different model, working from an identical prompt sequence and the same coding harness, swival.dev.
Every model received the same two prompts, in two distinct sessions:
- "In a markdown file, accurately describe the game Trackmania to someone who doesn't know the game at all, and is blind."
- "/goal Implement this in TypeScript."
That's it. No follow-ups, no nudges, no hand-holding. The first prompt produces flashmania_description.md, the design document the model writes for itself. The second asks the model to turn that document into a working TypeScript game. The harness handles tool use, file editing, and the agent loop; everything else is on the model.
The point isn't to crown a winner. It's to see how different models interpret an open-ended brief, where they invest their effort, what they leave out, and what kind of code they produce when nobody is steering.
Full agent traces for the original run are available on Hugging Face.
Each branch below is an independent attempt at the same task. Check them out and run them to see what each model produced.
main— Gemini 3.5 Flashgpt5— GPT-5ds4— DeepSeek 4 Flashglm-5.1— GLM 5.1qwen3.6-27b— Qwen 3.6 27Bminimax— MiniMax M2.7mimo— MiMo v2.5 Proqwopus— Qwopus 3.6 27B v2
These are from the main branch.