An AI Game Director That Grades Its Own Homework

What it does

This repository is essentially a management layer and scaffolding kit for AI agents that want to build Three.js browser games. It bundles SKILL.md instruction files, checklists, prompt templates, and a Vite + TypeScript + Three.js scaffold into a set of installable skills for Codex and Claude Code. The threejs-game-director skill acts as a project manager, automatically routing tasks to sibling specialists—gameplay systems, AAA graphics, UI design, audio generation, debugging, and QA—so you can ask for a game in plain English instead of hand-picking tools.

The interesting bit

The system treats the agent as a full studio rather than a code generator. It forces the agent to gather evidence before declaring success: builds must pass, Playwright screenshots must show nonblank canvases, mobile viewports must check out, and performance budgets must be met. Optional API keys for Tripo (3D models), Gemini (images), and ElevenLabs (audio) let the agent generate external assets when available, but the director is explicitly instructed to fall back to procedural or local assets and report credential gaps rather than silently failing.

Key highlights

• threejs-game-director auto-routes to nine specialist skills covering gameplay, graphics, UI, debug, QA, and asset generation
• Bundled scaffold includes deterministic test hooks, a seeded RNG, and Playwright templates for smoke tests, visual regression, and bot playtests
• Agents must produce tangible evidence—build passes, canvas pixel checks, performance snapshots, and visual scorecards—before claiming completion
• Optional integrations with Tripo, Gemini, and ElevenLabs for AI-generated 3D, image, and audio assets, with automatic fallback when keys are missing
• Same SKILL.md files work across both Codex and Claude Code runners

Caveats

• The standard installer CLI may not support all agent targets, forcing a fallback to manual shell installation from a cloned checkout
• The README repeatedly references “remaining risks” and iteration cycles, suggesting agent output often needs multiple passes before it is release-ready
• Verification is entirely agent-mediated: the bundled QA scripts are designed to be run by the agent, not the user, so a drifting agent can undermine its own checks

Verdict

Worth exploring if you want to prototype or polish Three.js games through high-level prompts rather than boilerplate. Skip it if you prefer direct control over your build pipeline or view agent-managed QA as a liability rather than a convenience.