Every AI coding session ends the same way: the agent says "Done ✓" — and you spend the next 30 minutes manually clicking through everything to find out it isn't.
I got tired of being my agent's test suite. So I built stop-manual-testing — a skill that flips the loop: every task ends with machine-checkable criteria the agent must actually run and pass before it's allowed to say "done".
It's modular, checks can run in parallel, and it works across agent-driven workflows.
A skill that stops you from manually testing your AI agent. The agent reads it, builds a machine-checkable verification system, and self-converges in a closed loop — collapsing the ~90% of dev time you spend staring at runs and judging by gut feel.
[中文]· English A skill for AI coding agents (Claude Code / Codex / ZCode / Cursor). Load it once, and instead of you manually clicking through the UI and eyeballing whether the agent "got better or worse this run," the agent builds itself a verification system where correctness is machine-checkable — then iterates inside a closed loop until it converges.
If you develop AI agents, you are likely stuck here:
If you're spending more time reviewing your agent's work than actually creating, give it a try — feedback welcome.