Add stop-manual-testing: machine-checkable verification system for AI agent projects

[Kaiji-Z]

What does this skill do?

Agent developers spend ~90% of dev time manually testing their agent: clicking through the UI, watching the run step by step, judging by gut feel whether it got better or worse. This skill attacks that pain directly.

stop-manual-testing has two phases:

1. Diagnosis (read-only): a 7-step audit of the project's verifiability — can the agent system run without the UI (ACI audit)? are intermediate states logged? are interfaces programmatic? what test infra exists? Outputs a priority-sorted gap list (P0/P1/P2) and updates the project's AGENTS.md.

2. Verification system build: helps stand up a machine-checkable verification system — regression suite built from happy paths, a two-layer judge (deterministic assertions + a clean-context LLM supervisor that scores but never judges), flag-based A/B regression runs, and a GATE (Declare-Verify-Enforce) mechanism that makes agent compliance visible and checkable. Orchestrates existing eval tools (DeepEval / LangSmith / pytest) rather than reinventing.

Design notes

• Consent-first loading: the skill announces capabilities when loaded and waits for an explicit trigger — it never auto-runs a pipeline that writes files.
• Design rationale is verifiable: built on Verifier's Law (Jason Wei, 2025), the Agent-Computer Interface concept (SWE-agent, NeurIPS 2024), and the Visible Checklist Pattern (Declare-Verify-Enforce).
• Honest limits documented: it orchestrates eval tools rather than replacing them; GATE makes compliance checkable, not forced.

Skill contents

• stop-manual-testing/SKILL.md — entry point (announce + wait semantics)
• stop-manual-testing/references/VERIFICATION.md — protocol body (7-step diagnosis, GATE, two-layer judge, red lines, tool orchestration)

No scripts, no runtime dependencies — pure instruction skill, MIT-licensed upstream (https://github.com/Kaiji-Z/stop-manual-testing).

Checklist

• SKILL.md with required frontmatter (name, description, author, version, tags)
• Real use case (built to solve the 90%-manual-testing pain in AI agent development)
• Not a duplicate of existing testing-strategy (that one covers writing unit/integration tests for project code; this one builds a verification system for the AI agent system itself)
• Self-contained folder, no external dependencies