human-action-neededlisted

# Human action needed The point of this skill is to be thorough on your own behalf, not to save yourself the trouble of asking. A step that *might* need a human is not a reason to ask early — it's a reason to get as close to it as you possibly can first, so that when you do ask, there is almost nothing left to do. ## Rules 1. **Run this on the main agent, never a dispatched subagent.** A subagent's browser tab and task state disappear when it finishes — only the main agent can leave a tab open, pause, and resume around a human's action in the same session. 2. **Try it yourself until you are fully, genuinely stuck — not before.** "This step looks like it needs a human" is not stuck. Actually attempt the click/navigation/command first. Stop only on a real, confirmed block: an explicit tool-safety classifier denial, or a documented structural human-only boundary (Shopify has no API to grant an app scopes on a merchant's behalf; a payment confirmation; a 2FA code sent to a personal device; a CAPTCHA). Don't guess that something is blocked — try it and see. 3. **Always navigate there yourself first, via Chrome MCP.** By the time you ask, the browser must already be sitting on the exact page the human needs — never make them find, search for, or type a URL themselves. Do every other safe step too (fill in what you can, select what you can, get the page into the state where literally one click/action is left). 4. **When you ask, be thorou