Letting AI play my game – building an agentic test harness to help play-testing

· ai-agents · Source ↗

TLDR

  • Developer built a text-based HTTP test harness so Claude can play-test Crossword Dungeon autonomously, cutting manual QA from hours to under 30 minutes per milestone.

Key Takeaways

  • The harness wraps the browser game with a Node.js text renderer and synthetic input events without touching any game source files, making it completely non-invasive.
  • Stateless HTTP calls expose game state as text snapshots; Claude navigates by sending key events and reading panel output, no screenshots required.
  • AI play-testing caught and fixed real bugs in the same session – a cage trap state bug (state.caged never cleared after solving) was found, diagnosed, and patched during live play.
  • Claude ran milestone 13 (5 new special encounters: Thieving Goblin, Cage Trap, Slime Blob, Dormant Sentinel, Thief) in 12 minutes and ~120k tokens using Sonnet 4.6 on Claude Pro.
  • Custom fixtures let the AI recreate specific bug states on demand rather than navigating to them manually each time.

Hacker News Comment Review

  • No substantive HN discussion yet.

Original | Discuss on HN