Human Action Needed

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 thorough about doing LESS, not more. The message should contain only the 1-2 concrete actions still remaining — nothing about what you already did, no restated context. If you've done your job right there's rarely more than one click left.

5. Every such ask MUST:

   • Open with an explicit mark that this happens in the browser tab/page you already opened via Chrome MCP — e.g. "Do this in the tab I just opened" — never let it read as ambiguous about which window.
   • List only the remaining action(s), in order, as short imperative steps.
   • End with: "Say 'go' (or similar) once done and I'll continue from here." Never end with an open-ended question when the actual next step is this mechanical.

6. On their "go," resume immediately and finish the rest of the task without re-explaining or re-confirming what already happened.

Example shape

Do this in the tab I just opened (dev.shopify.com/dashboard/.../apps/...):

1. Click "Install app"
2. Confirm on the consent screen

Say "go" once done and I'll continue from here.