Collaborative 3D browser game development with AI
While recovering from illness, the author built the browser-based off-road racing game Slipface using Claude Code. By writing documentation, defining skills, providing reproduction links, and describing game feel, they achieved high-quality AI frontend and game development collaboration.
Reddit r/vibecoding/u/Slipfacegame
Claude Code (Opus / Sonnet)Three.jsWebGPUMeshySunoCloudflare Workers
Added 2026-10-31 · originally published 2026-10-29
Key points
- Write system docs for the AI: Every system should have documentation explaining the design rationale and be updated with every code commit to prevent the AI from mistakenly 'fixing' intentionally designed features in new sessions.
- Turn tedious tasks into skills: Encapsulate multi-step workflows (like adding a vehicle) into single-request commands, and include a manual confirmation (OK) step for actions that consume credits (like generating 3D models).
- Let the AI reproduce and inspect bugs itself: Provide direct debugging links to jump into levels, play solo, or spectate AI vehicles so the AI can observe and troubleshoot issues on its own.
- Describe game feel instead of prescribing fixes: Describe the actual experience to the AI (e.g., 'feeling floaty and sliding sideways when crossing the ridge') and let it find the root cause.
- Commit frequently in small steps: Physics engines can easily break after a few modifications; frequent commits and checkpoints turn issues into simple rollbacks rather than painful debugging sessions.
Pitfalls
- The AI in new sessions may attempt to 'fix' system logic that was intentionally designed that way.
- Directly letting the AI auto-generate 3D models and other assets can waste a large number of credits unnecessarily.
- Physics effects are prone to breaking after multiple modifications, and lacking frequent commits and checkpoints makes troubleshooting difficult.
.cursorrules
- Write docs for every system explaining why it works, and update docs in the same commit as the code. - Turn multi-step chores into single skills with a human confirmation step if resources/credits are consumed. - Describe the feel of the issue rather than guessing the fix, and let the AI investigate the cause. - Commit small and often to make physics or logic breaks easy to revert.
Copy-ready prompt
Please help me implement [specific functional module]. Before writing any code, please consult and follow the relevant system documentation. Do not modify underlying logic that was intentionally designed this way. If you encounter issues while tweaking [e.g., physics feel / vehicle controls], please first analyze its behavior and feel (e.g., [describe specific feel experience]), identify the root cause, and then make modifications. Ensure small, incremental commits with each change so we can easily roll back at any time.
#AI Frontend#Claude Code#Dev Workflow#3D Game
Related techniques
Summarized by AI from the original source - refer to the original post for details. Copyright belongs to the original author.