dayan-agent-skills

Release review loop

Use this before publishing an AI-assisted artifact, Skill, documentation change, or verifier.

Builder/checker split

Stage	Builder owns	Checker owns
Frame	goal, audience, artifact	missing authority, wrong audience, unsafe scope
Produce	implementation or content	no direct edits unless the boundary allows it
Verify	local tests and examples	independent failure-mode review
Release	package notes	evidence, limitations, human approval

Review checklist

• What exact claim is the release making?
• Which test proves that claim?
• Which important claim is still unproven?
• Is any customer, credential, private path, or internal strategy visible?
• Can a user install or inspect without touching their live configuration?
• Is there a feedback path for failures?

Minimum release note

What changed:
Who should try it:
What is verified:
What is not claimed:
Known blockers:
How to report a failure:

Result labels

• BLOCK: do not publish until the named issue is fixed.
• CONCERNS: publish only if the limitation is explicit and acceptable.
• CLEAN: no blocking issue found within the review scope.

CLEAN never means universally safe. It only means the reviewed evidence supports the release boundary.

This site is open source. Improve this page.