You can hand HQ a screening job that would take you twenty minutes of clicking — "screen this week's applicants, flag the gaps, move the strong ones forward" — and get it done in one instruction. The reason this is safe rather than reckless is plan-gating: HQ proposes the whole change before it touches anything, and you approve it. This recipe is about giving HQ a big request on purpose.

The skill here is phrasing one instruction well, then reading the plan carefully. Do both and HQ is a coordinator you can delegate to.

What you'll need

Applicants already in your pipeline — through the careers page or imported.
A clear screening rule in your head: which role, what disqualifies, where the strong matches go.
A role that can do the changes you're asking for. HQ acts as you — it can't do what your permissions can't.

The chain

Give HQ one specific instruction. In the HQ chat, ask for the whole job at once — for example: "Screen this week's applicants for the Backend Engineer role, flag anyone whose resume doesn't mention Go, and move the strong matches to phone screen." Name the role, the criteria, and the destination stage. Specific prompts get the plan right the first time.
HQ reads, then writes a plan. The lookups — finding the applicants, reading resumes, checking match ratings — run immediately, because reads change nothing. For the changes, HQ doesn't act: it writes a plan with a rationale and an ordered list of concrete steps, then pauses.
Review the plan in full. Read the rationale — if the reasoning is off, the steps will be. Then read the steps: each one names the tool and the exact action ("move Priya Nair to the phone screen stage on the Backend Engineer role"). Five changes means five steps, all visible before any of them run.
Approve, or reject and rephrase. Approve if the steps are right — HQ executes them in order and reports what it did, with links to every change. Reject if a candidate or a stage is wrong; nothing happens, and you clarify what you meant. Rejecting costs nothing.

Where it can break

The plan proposes more than you asked for. That usually means your prompt was broader than you meant. Reject it and narrow the request — naming the role and the exact criteria tightens the plan.
The right action, the wrong record. The failure to watch for is a plan that looks reasonable but aims a correct step at the wrong candidate or stage. The plan is shown in full precisely so you catch this — skimming it defeats the safeguard.
You approved, then wanted it back. An approved plan runs as written, and Chosen has no one-click undo. Some steps reverse by hand; a sent email cannot be unsent. Read before you approve, not after.
HQ says it can't. HQ runs with your role. If your permissions can't delete a job or manage the org, neither can HQ for you — it tells you rather than failing quietly.
A step fails mid-run. If a record was deleted out from under a step, HQ stops and reports the error instead of plowing ahead.

Variations

Make it a question first. If you're not sure what the screening pass will catch, ask HQ a read-only version — "how many of this week's applicants don't mention Go?" Reads run immediately and change nothing, so you can scope the job before you ask for the changes.
Smaller batches, more plans. A giant instruction produces a giant plan that's harder to verify. Splitting the screen by role or by day gives you shorter plans you can actually read — and a chat does one plan at a time anyway.
Hand HQ a spreadsheet. Attach a CSV of names and HQ can bulk-create candidates from the rows as part of the plan.
Let HQ start the scheduling. A screening pass can end with scheduling — ask HQ to create a scheduling request for the strong matches, and that becomes a step in the plan. See the hands-off interview loop.