A developer built an agentic test harness so an AI could play their game and help with play-testing.
A developer built an agentic test harness so an AI could play their game and help with play-testing.
The game is a browser-based, turn-based crossword dungeon. Claude wrote a node.js wrapper that leaves the game files alone. It adds a text renderer and synthetic key events, served over HTTP. The AI played via curl calls, built fixtures to reproduce bugs, and was told to playtest each milestone feature before moving on.
Milestone 13, with five new special encounters, took 12 minutes and about 120k tokens. The AI caught and fixed a cage bug on its own. Manual playtesting was still needed and found more bugs, but it took under 30 minutes.
It ran on Claude Code with Sonnet 4.6 on the Pro plan; the approach suits text-heavy, turn-based games better than screenshot-driven browser automation.
Look at my game in [path] and build a test harness that lets an AI agent play it: expose the game state as text or JSON and accept actions like [moves or inputs]. Then have the agent play [number] sessions and report bugs, balance problems, and confusing moments. Finished means a written play-test report with steps to reproduce each issue.
Claude Code got an old DirectX game running on modern hardware, decompiled much of it, built a headless parallel runner, and now writes full level guides with screenshots.
An agent runs overnight, collects the author's AI newsletter subscriptions and produces a daily podcast to listen to on walks.
Commenters use agents for weekly meal planning, daily briefs, errand tracking, email sorting, home repair load and household task organization.