plan-verification-splitlisted

# plan-verification-split A lot of implementation plans end with a list of acceptance criteria that the agent is supposed to "verify" before closing the task. The default reflex is to paste the whole list into chat and ask the user to click through it. That's lazy and it wastes the user's time, because most items can be proven by evidence the agent can collect itself — a DB row, a response body, a grep hit, a file existing — while only a smaller set genuinely requires human eyes (visual rendering, real user input, external hardware). This skill exists so you split the list once, do the auto work yourself, and only pull the user in for the parts that actually need them. ## Core idea Every acceptance criterion gets one tag. Tags are of two families: - `[auto:<method>]` — the agent can verify this directly, right now, with tools it already has. - `[human:<reason>]` — the agent cannot verify this; a human (or a physical environment) must. The method or reason goes after the colon so future readers can see **how** the item was checked, not just that it was. This matters because "auto" without a method is a lie — six months later nobody remembers whether "item 4 PASS" meant "I saw the DB row" or "the toast popped up in my head." ## When to use this skill Invoke it the moment you encounter a verification checklist in a plan or spec — before you type "please check items 1 through 8" into the chat. Also invoke it when: - A user asks you to run a plan's Task 5 / Task N "manual