Getting LLMs to reliably generate functional games required solving three specific engineering bottlenecks:

1. The Training Data Scarcity: LLMs barely know GDScript. It has ~850 classes and a Python-like syntax that will happily let a model hallucinate Python idioms that fail to compile. To fix this, I built a custom reference system: a hand-written language spec, full API docs converted from Godot's XML source, and a quirks database for engine behaviors you can't learn from docs alone. Because 850 classes blow up the context window, the agent lazy-loads only the specific APIs it needs at runtime.

2. The Build-Time vs. Runtime State: Scenes are generated by headless scripts that build the node graph in memory and serialize it to .tscn files. This avoids the fragility of hand-editing Godot's serialization format. But it means certain engine features (like `@onready` or signal connections) aren't available at build time—they only exist when the game actually runs. Teaching the model which APIs are available at which phase — and that every node needs its owner set correctly or it silently vanishes on save — took careful prompting but paid off.

3. The Evaluation Loop: A coding agent is inherently biased toward its own output. To stop it from cheating, a separate Gemini Flash agent acts as visual QA. It sees only the rendered screenshots from the running engine—no code—and compares them against a generated reference image. It catches the visual bugs text analysis misses: z-fighting, floating objects, physics explosions, and grid-like placements that should be organic.

Architecturally, it runs as two Claude Code skills: an orchestrator that plans the pipeline, and a task executor that implements each piece in a `context: fork` window so mistakes and state don't accumulate.

Everything is open source: https://github.com/htdt/godogen

Demo video (real games, not cherry-picked screenshots): https://youtu.be/eUz19GROIpY

Blog post with the full story (all the wrong turns) coming soon. Happy to answer questions.