What's inside

One SKILL.md (the playbook) + one starter template. A preview:

Agent-Driven QA Loop

When to use

• Your build needs playtesting and your QA team is you (or is asleep).
• You want regression coverage on every nightly build without writing a full test suite.
• Bugs keep shipping that "looked fine on my machine" — visual, timing, and state bugs.
• You're already using an AI agent in development and want it pulling QA shifts too.

Workflow

1. Define the QA contract (input: build → output: test charter)
   Write what the agent is allowed to do: launch the build via the exact command, play for N minutes per session, capture screenshots at fixed intervals and on errors, read logs, and file bug reports — but never modify source, never touch networked services, never spend money or create accounts. A QA agent with write access to your repo is not QA, it's a liability.
2. Script the play session (input: charter → output: session script)
   Give the agent a concrete session: start new game, play the core loop for 5 minutes pursuing the win condition, then deliberately do the wrong things (mash buttons, pause mid-animation, alt-tab, resize the window, idle for 2 minutes). Wrong-thing testing finds the crashes; happy-path testing finds nothing. Include the oracle: what "correct" looks like (target frame rate, no error dialogs, state transitions listed).
3. Set up screenshot inspection (input: session script → output: inspection protocol)
   The agent compares screenshots against expectations: HUD elements present, no rendering artifacts (black boxes, stretched textures, z-fighting), text readable and unclipped, menus navigable. Teach it your game's specific failure signatures with 3–5 example screenshots of past bugs ("this is what the inventory soft-lock looks like"). Vision-based QA beats log-based QA for the bugs players actually see.
4. Run the overnight loop (input: protocol → output: bug inbox)
   Schedule: build at midnight, agent plays three 10-minute sessions with different seeds/characters, files one report per distinct bug with screenshot + repro steps + log excerpt. Cap the run (timebox, max sessions) so a crash loop doesn't burn your API budget by morning. In the morning you triage, not debug — the agent already attached the evidence.
5. Triage with severity math (input: bug inbox → output: fix queue)
   Score each bug: crashes and soft-locks first, then progression blockers, then visual glitches, then polish. For each: can a player hit it in normal play (multiply severity), and did it appear in the last build (regression = fix today)? Anything the agent found that you can't reproduce in 5 minutes goes back to the agent with "reproduce this deterministically" before it earns a place in the queue.
6. Close the loop into the build (input: fix queue → output: regression set)
   Every fixed bug becomes a permanent check: add its screenshot signature and repro to the agent's protocol. Over weeks this compounds into a regression suite you never had to hand-write. Review the protocol monthly — delete checks for systems you removed, or the agent will cry wolf about ghosts.

Template preview

Agent-Driven QA Loop Worksheet

Set this up once; it compounds into a regression suite.

1. Test charter

• Launch command:
• Session length × count:
• Agent may: play, screenshot, read logs, file reports
• Agent may NEVER:
• Timebox / budget cap:

2. Session script

Happy path (core loop, N minutes, pursuing the win):

Wrong-thing list (button mash, pause mid-animation, alt-tab, resize, idle 2 min):

Correctness oracle (frame rate, no error dialogs, state transitions):

3. Inspection protocol

Your game's failure signatures (with 3–5 example screenshots of past bugs):

1. 2. 3.

4. Overnight run

• Build time:
• Sessions (count × length, seeds/characters):

Buy the full skill pack — $4.99 Pro — all 511 skills