Developer Alex Ermolov released Godogen on March 16, 2026, an AI pipeline that generates complete Godot 4 games from text descriptions. The project, which gained 150 points and 90 comments on Hacker News, represents a year of development through four major rewrites to solve fundamental challenges in autonomous game development.
Two-Skill Pipeline Generates Architecture, Art, and Code
Godogen orchestrates a two-skill pipeline where users describe desired games and the system handles everything else. According to the documentation, "An AI pipeline designs the architecture, generates the art, writes every line of code, captures screenshots from the running engine, and fixes what doesn't look right." The orchestrator plans project structure while the task executor implements each piece in a 'context: fork' window to prevent error accumulation.
The technical stack combines Claude Code with Opus 4.6 or Sonnet 4.6, Python 3, Gemini API for visual QA and 2D art generation, and Tripo3D for 3D model generation. Asset generation uses Gemini API for 2D art and textures, while Tripo3D converts images to 3D models with budget-aware optimization.
Three Critical Engineering Breakthroughs
Ermolov solved three fundamental problems that prevented autonomous game generation. First, GDScript's ~850 classes use Python-like syntax that causes LLMs to hallucinate invalid Python idioms. He built a custom reference system with hand-written language specs, full API docs converted from Godot's XML source, and a quirks database for undocumented engine behaviors. The agent lazy-loads only needed APIs to manage context windows.
Second, build-time versus runtime state creates fragility. Ermolov explained that "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." The system teaches the model which APIs are available at each phase and ensures every node has its owner set correctly or it "silently vanishes on save."
Visual QA Agent Prevents Coding Agent Bias
The third breakthrough addresses evaluation loop bias. Ermolov noted, "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."
The system produces properly organized Godot 4 projects with correct scene trees, readable scripts, and structured asset organization. Generating a single game typically requires several hours, making cloud VM deployment practical for extended sessions. The project is available under MIT license and has accumulated 407 GitHub stars. A demo video is available at https://youtu.be/eUz19GROIpY.
Key Takeaways
- Godogen autonomously generates complete Godot 4 games from text descriptions, handling architecture design, asset generation, code writing, and visual testing
- The system uses a custom reference system with hand-written language specs and full API docs converted from Godot's XML source to prevent LLM hallucination of invalid GDScript
- A separate Gemini Flash agent acts as visual QA, evaluating only rendered screenshots to catch visual bugs like z-fighting and physics explosions that text analysis misses
- Scenes are generated by headless scripts that build node graphs in memory and serialize to .tscn files, avoiding fragility of hand-editing Godot's serialization format
- The project represents a year of development through four major rewrites and is available under MIT license with 407 GitHub stars