Bughunt Skill — Testing skill for Claude Code

bughunt — an agent skill that plays your app until it breaks

[](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview) [](LICENSE)

Unit tests and a fresh-game autoplay pass can both be green while a real player hits a bug in the first minute: the first command after "Continue" is rejected, an AI acts before its screen is visible, a finished player's save comes back to life. These bugs live on **transitions**, and only show up when someone plays the way people actually play.

`bughunt` makes an agent do that playing for you: seeded **personas** drive the shipped build through its real input paths, **invariant oracles** check every step, and every failure becomes a replayable bundle, a regression test, and a root-cause fix.

What it does

1. Build the target as shipped (exported / simulator build, not only the editor).
2. Run seeded personas: new player, button masher, save → quit → resume at every phase, background / foreground interruptions, slow player who lets timers expire, speed changer, random tapper, other languages, small and large screens, Reduce Motion.
3. Check oracles at every step: no errors or crashes, every enabled command accepted on the first try, no AI action before its screen is ready, UI inside the safe area with no overlap, sane state (no negative currency, no vanished units), deterministic replay of the same seed, no softlock, persisted state round-trips and stays eligible.
4. On failure: save seed + action log + screenshot + state dump → minimise → regression test → root-cause fix → replay the whole matrix.
5. Stop when every persona × seed passes twice in a row; report the evidence.

Adapters: **Godot** (in-game QA runner driven by command-line flags), **native iOS** (XCUITest / Maestro), **web** (Playwright). App profiles fit the personas and oracles to the product: **game**, **CRUD app** (rec