AI Compass / Knowledge cards
"Done" only with evidence
a check command instead of self-reporting
An acceptance criterion settles before the work how “done” is recognised. It becomes checkable when each criterion comes with a command and the expected output. The open-source skill Unlazy writes this into a gates file. You can also do it by hand with a simple list.
What it is for
You do not rely on the agent’s self-report. You read an output that anyone can verify.
Steps
- Before the task, write the acceptance criteria into a file. Each criterion lists: result, the command that proves it, expected output, evidence line.
- A criterion marked “pending” or without evidence counts as not met.
- Choose at least one check that can fail, for example “nobody gets in with a wrong password”.
- Have the tests written by a separate agent, not by the one that builds.
- After the work, the main agent or a reviewer runs the commands itself, not the builder.
- A criterion that cannot be met may be dropped with a justification. It is then listed in the final report.
- For large projects, have the open decisions clarified beforehand and check the result against this answer key, not against what the agent thinks is right. In the video, the skill Wayfinder does this with a round of questions (34 questions in the example).
Limits
- A check only proves what it checks. In another AI LABS video, all 11 checks were green, yet the order form had not been tested at all.
- Unlazy and Wayfinder are third-party skills. According to the README, the Unlazy gates template can also be copied without installing anything, and inspected with the included check script before running it. According to the README, Wayfinder now creates decision tickets in the issue tracker. That may differ from the video.
- The check script runs commands that an agent wrote. So use it in a test project only.
- Time and cost figures in the video are the channel’s claims: Unlazy took 3–4 hours for a login page, and a Wayfinder run is said to correspond to about $116 in API costs.
Links
- Checkpoints and gates: the full version with four checkpoints. This card is the light introduction.
- A yes/no guard for Claude Code: Belay automatically checks whether evidence exists. The gates file defines beforehand what it looks like.
- Let Claude work until the goal is met – with /goal: The checklist becomes the stop condition.
- Checking your own website from the outside: The labels seen / assumed / unknown separate evidence from assumption in the report.
- Self-improving loop with locked checks: Where the example with the 11 green checks comes from and how the checks are protected from the agent.
Sources
- AI LABS: GitHub’s #1 Trending Author’s New Claude Skill Is Insane, 20 August 2026, 6:27, 8:04–8:37, 9:08, 11:16–11:51. Sponsored by DataForSEO and SerpApi, plus a paid community
- AI LABS: This Claude Skill Just Fixed Loop Engineering, 14 August 2026, 11:48–13:57. Self-promotion for the paid community
- AI LABS: Claude Code Setup Guide By The Creator Himself, 31 July 2026, 5:22–6:10. Sponsored by Buzzy
- AI LABS: How To Use Claude Fable 5 Better Than 90% Of The People, 6 July 2026, 7:00–7:55. Sponsored by VEED (approx. 5:53–6:40); the finished test-author agent is in the paid community
- AI LABS: He Finally 10x Claude Code With This Method, 2 October 2026, 8:09. Sponsored by Upstash Box
- GitHub: Leonxlnx/unlazy and mattpocock/skills (Wayfinder), retrieved 11 October 2026
Status
Checked 11 October 2026, extended 16 October 2026 · Maturity new (the method is usable immediately without any tool; the skills are available, not tested by me)
Want to try this with your team? That’s what the AI Lab is for. See the AI Lab