A fully playable browser game — clone it, run it, and extend it with your AI agent.
You get a working dodge game out of the box. Then paste a prompt and your agent builds the next feature. No setup. No fighting your AI to stay on track — the architecture rules and skill files are already wired in.
🚀 Pro Kit — coming soon React Three Fiber architecture, real project structure, case study branches, and a Vercel-ready deploy. Subscribe on Gumroad to get notified when it drops + lock in launch pricing. → Subscribe for early access
| 🎮 | Cube Runner — a fully playable dodge game, ready to extend |
| 🧠 | 11 Three.js skill files your agent loads automatically — no explaining Three.js every session |
| 📋 | 4 prompt files to extend the game feature by feature |
| 📐 | AGENTS.md — architecture rules that keep your agent on track |
| 🌿 | new-game branch — blank canvas + /plan prompt to design your own game from scratch |
git clone https://github.com/heagandev/threejs-agent-starter
cd threejs-agent-starter
npm install
npm run devOpen http://localhost:5173 — the game runs immediately.
This branch was built from new-game using the /plan prompt and a single agent session. That's the whole workflow:
- Start from
new-game— clean canvas, no code to fight - Use
/planto design your game with your agent - Build it out, one feature branch at a time
Open source (MIT) — build with it, extend it, share what you made with this.
This kit works with any AI coding agent — Claude Code, Codex, Cursor, Windsurf, Copilot, or any tool that reads files from your repo.
Paste any prompt file into your agent to add a feature:
| Prompt | What it builds |
|---|---|
/power-ups |
Speed boost, shield, and slow-mo power-up system |
/game-juice |
Screen shake, squash/stretch, score milestones |
/level-two |
A second level with transitions |
/plan |
Plan an entirely new game from scratch (git checkout new-game first) |
Recommended workflow:
# Start a new feature on its own branch
git checkout -b feat/power-ups
# Paste prompts/power-ups.md into your agent
# Review the output, run it, commit when it works
git add -A && git commit -m "feat: add power-up system"Your agent reads AGENTS.md automatically and follows the architecture rules — no extra setup needed.
The skills/ folder contains Three.js reference sheets. Your agent loads them automatically via AGENTS.md — you don't need to explain Three.js in every prompt.
| Skill | Covers |
|---|---|
threejs-fundamentals.md |
Scene, camera, renderer, transforms |
threejs-lighting.md |
Lights, shadows |
threejs-geometry.md |
Shapes, BufferGeometry |
threejs-interaction.md |
Input, raycasting |
threejs-animation.md |
Keyframes, procedural motion, spring physics |
threejs-materials.md |
Materials, PBR |
threejs-textures.md |
Texture loading, mapping |
threejs-shaders.md |
GLSL, ShaderMaterial |
threejs-postprocessing.md |
Bloom, effects |
threejs-loaders.md |
GLTF, asset loading |
threejs-game.md |
Game loop, input, collision, HUD, audio, game-feel patterns |
Once you've run through the included prompts, here are directions worth exploring:
- Make your own game —
git checkout new-game, use/plan, start fresh - Mobile controls — touch/swipe input for mobile players
- Leaderboard — score submission with a simple backend
- Custom shaders — visual effects via
skills/threejs-shaders.md - Spatial audio — sound design with the Web Audio API
- Procedural generation — infinite obstacle layouts from a seed
Each one makes a great focused agent session. Open a branch, paste a prompt, review and commit.
MIT — free to use, extend, and ship.
This is the free starter. Here's what the paid kits add:
| Kit | What you get | Price |
|---|---|---|
| Pro Kit | React Three Fiber architecture, real project structure, case study branches, Vercel deploy | $49 at launch → $79 |
| Pro Plus | Everything in Pro + ESLint/Husky guardrails, AI memory system, advanced prompts | $79 at launch → $99 |