I made LLM-generated games actually playable.

Prompt in, complete Godot project out — architecture, 2D/3D assets, all the code, visual QA. No human in the loop. Before these pieces came together, the output was consistently unplayable. The whole system is two @claudeai Code skills — an orchestrator and a task executor.

LLMs barely know GDScript. 850 engine classes, Python-like syntax that isn't Python. Without docs in context, you get hallucinated methods and confident code that doesn't compile. Fix: hand-written language spec + API docs + engine quirks DB. Two-tier lazy loading.

The key insight: the coding agent is biased toward its own output. So a separate vision agent (Gemini) sees only the rendered pixels and compares to a reference image. It catches what the coder can't judge about itself: z-fighting, floating objects, physics explosions.

Open source: Every prompt, instruction, and engine quirk is in the repo. Blog post with the full story (all the wrong turns, not just what works) is coming.

Your godogen is GREAT! Making some cool stuff already! Please keeep up the good work! BTW: I see you are studying in an Italian University, we can talk in Italian :-)

The Alpine simulator looks good. IMO, that's the best result out of those

Hola, ¿podrías escribirme por mensaje privado? Espero que podamos lograr una colaboración de patrocinio.

Can I use this on OpenCode?

Have you a little game that can I test?

Not yet — next step in the roadmap is to implement game export rules (focus on Android for now) and prepare real game

No way! Do you have YouTube to watch it how you do it from scratch??

You realize almost every image featured on their site was stolen from actual artists, right?

This is great