Context
Plutus licensed games from third parties, creating strategic dependency risk. We explored whether AI could autonomously generate complete, playable games—enabling scalable, defensible content creation.
My Role: Sole architect and implementer for feasibility study and PoC.
Phase 1: Feasibility Validation
Models Evaluated: Claude 4, ChatGPT-4.5, Gemini
Results:
- Gemini and Claude generated playable browser games in single prompt
- ChatGPT required 2-3 iterations
- Output quality sufficient for JavaScript games
Conclusion: End-to-end game generation was feasible.
Phase 2: System Design
Core Principle: Reliable autonomous generation requires structured intent, not free-form prompting.
Architecture
1. Idea → Structured Game Brief
- Users submit rough idea
- System converts to JSON game brief (mechanics, style, controls, difficulty)
- User explicitly reviews and approves
- Eliminated ambiguity before execution
2. Autonomous Generation Pipeline
- Expansion Agent: Detailed technical description
- Coding Agent: Full game code in sandboxed environment
- RAG-Backed Memory: All files indexed for context-aware iteration
- Multi-step reasoning across growing codebase
3. Compile, Host, Serve
- Automatic compilation
- Deployed to unique URL
- Time-to-play: minutes (vs hours/days)
4. Iteration Loop
- Natural language change requests
- Agent navigates codebase via RAG search
- Targeted code changes applied
- Recompiled and re-served
Platform Capabilities
Beyond generation:
- User authentication and storage
- Credit-based model usage
- Payments and usage tracking
- Integration hooks for Plutus hosting
Production-adjacent, not throwaway demo.
Outcome
- Validated autonomous game generation as feasible product direction
- Delivered working PoC without allocating full engineering team
- De-risked new product line before major investment
- Established architectural patterns reusable across AI products
Legal Disclosure
This documents work limited to feasibility validation and PoC that I personally led. Pixelsurf evolved significantly after my departure (health-forced sabbatical). Specific fine-tuning, orchestration, and optimization details intentionally excluded.
Consulting Takeaway
This demonstrates how I help startups de-risk technically ambitious ideas early, avoid premature scaling, and identify real constraints before committing roadmaps.
Completed during pre-seed stage at Plutus