outsmart-claude-escape-room

Design escape rooms, watch Claude try to solve them in real-time. Can you outsmart the AI?

Comment

Inspiration

We were fascinated by the question: Can humans design puzzles that outsmart AI?

With AI becoming increasingly capable at problem-solving, we wanted to flip the script. Instead of AI helping humans, what if humans tried to trap AI? Escape rooms are the perfect testing ground, they require spatial reasoning, multi step planning, and the ability to avoid traps. We wanted to create an interactive experience where anyone can design devious puzzles and watch in real time as AI attempts to escape.

What it does

Outsmart Claude is a game where you design escape room puzzles and challenge AI to solve them.

  • Build custom rooms using a visual drag-and-drop editor with walls, keys, color-coded doors, clues, and traps
  • Watch AI think in real-time - see the AI's reasoning process as it analyzes the room and makes decisions
  • Multiple ways to win: Trap the AI, make it exceed 50 moves, or watch it escape
  • Fog of War mode limits AI vision to nearby cells, making puzzles significantly harder
  • Share your puzzles via URL - challenge friends to see if their rooms can trap the AI too
  • Sound effects provide satisfying audio feedback for every action

The game works with Claude API or runs completely free locally with Ollama.

How we built it

  • Frontend: React 18 with TypeScript for type safety and better developer experience
  • Styling: Tailwind CSS for rapid UI development with a dark, futuristic theme
  • Build tool: Vite for lightning-fast development and optimized production builds
  • AI Integration: Abstracted provider system supporting both Claude API and Ollama (local LLMs)
  • Audio: Web Audio API for synthesized sound effects (no external audio files needed)
  • State Management: Custom React hooks for game state, room building, and AI playback control
  • Sharing: Base64-encoded URL parameters for compact room serialization

The AI receives a text prompt describing the room state, available actions, and must respond in a structured format that we parse to execute moves.

Challenges we ran into

  1. AI prompt engineering - Getting the AI to consistently follow the action format was tricky. We had to iterate on the system prompt many times, adding explicit examples and warnings about common mistakes (like trying to walk through locked doors).

  2. Key-door color matching - Ensuring the AI understands that red keys only open red doors required careful prompt design and contextual hints.

  3. Fog of War implementation - Filtering what the AI can "see" while maintaining game integrity required changes across the prompt generator, grid renderer, and game state.

  4. Local LLM support - Smaller models struggle with multi-step reasoning. We had to tune prompts differently for Ollama vs Claude.

  5. Sound without files - We wanted zero external dependencies for audio, so we synthesized all sounds using Web Audio API oscillators.

Accomplishments that we're proud of

  • It actually works! Watching AI reason through puzzles and sometimes fail is genuinely entertaining
  • Runs locally for free with Ollama - no API costs for testing and demos
  • Fog of War creates a meaningful handicap that makes puzzles significantly harder for AI
  • URL sharing lets anyone try your exact puzzle design with one click
  • Clean architecture - provider abstraction makes it easy to add new AI backends
  • Polished UX - confetti on wins, sound effects, smooth animations, responsive grid sizes

What we learned

  • AI is surprisingly good at spatial reasoning - simple puzzles get solved quickly
  • Prompt engineering is an art - small wording changes dramatically affect AI behavior
  • Local LLMs are viable for interactive applications with good prompt design
  • Web Audio API is powerful for generating sounds programmatically
  • Base64 URL encoding is great for shareable state without a backend

What's next for Outsmart Claude

  • Leaderboard - Track which room designs have the highest AI failure rate
  • Multiple difficulty levels - Adjust AI "intelligence" by changing models or adding handicaps
  • Manual play mode - Let humans race against AI to escape the same room
  • Room templates - Themed puzzle packs (prison escape, haunted mansion, etc.)
  • AI memory - Let AI learn from failed attempts across multiple rooms
  • Multiplayer - Real-time competitions to see who can trap the AI first

Built With

Share this project:

Updates