#Three.js Game Skills: One Prompt, Nine Skills, a Full Browser Game

An open-source repo tries to turn an AI coding agent into a complete Three.js game studio, complete with a director, specialized roles, and built-in QA.

#What it is

Three.js Game Skills is an MIT-licensed repository that packages a set of skills for AI coding agents like Codex and Claude Code. Instead of prompting an agent to build a game in one shot, you install the skill pack, call a single director command, and describe the game you want. The director routes the work across nine specialized skills: gameplay, AAA-style graphics, user interface, debugging, release QA, and optional branches for 3D model, image, and audio generation.

The defining choice is that the workflow starts with a playable loop, then layers on visuals, HUD, effects, and audio. The demos are browser games you can actually play, not static 3D scenes built for screenshots.

#Why it matters

• It shows a real multi-agent routing pattern: one director picks the right specialist so you do not have to manage each skill manually.
• The playable-first approach is harder and more valuable than generating a decorative 3D scene.
• Generated games include seeded RNG, deterministic test hooks, and Playwright templates for smoke tests, visual regression, and bot playtests.
• The agent must self-verify before declaring completion, which adds discipline that most prompt-and-pray workflows lack.
• The core runs without paid APIs, using local or procedural fallbacks, with optional paid integrations for Tripo, Gemini, and ElevenLabs.

#How it works

After installing the pack, you call threejs-game-director and describe your game. The director selects the appropriate skill from the nine available. Gameplay comes first, then graphics, UI, effects, and audio are upgraded in sequence. When API keys are available, Tripo handles 3D models, Gemini generates images, and ElevenLabs generates audio. Without keys, the system falls back to local or procedural assets so the pipeline still produces something runnable.

The QA layer is the standout feature. Seeded randomness makes runs reproducible. Fixed test hooks give deterministic behavior. Playwright templates let the agent run smoke tests, visual regression checks, and bot playtests, so the agent checks its own work before reporting done.

#Caveats

This is a skill pack on top of AI coding agents, not a game engine. Output quality depends on the model, the prompt, and how clearly you scope the game. The optional asset generation branches need paid API keys, and the fallback procedural assets will not match the polish of the demo games. The repo has multiple working browser demos, but it is still experimental tooling for developers comfortable with agents and Three.js.

#Who it's for

This is for developers who want to experiment with structured, QA-backed AI game generation and who already know their way around Three.js and coding agents.

If you are curious about what a disciplined, role-based approach to AI game dev looks like, Three.js Game Skills is worth trying, and the built-in testing is the part most competitors skip.