AI Automation Resource

Delegating work to AI agents by voice — Checklist

Open this before you speak: the workspace that has to exist first, the phrases that put work in the right project, and what you still open and check yourself.

Keep this open while you work, not after. The sections run in the order a voice session runs: what has to exist before the first sentence, how to phrase the request, what to watch while threads run, and what you open yourself at the end. Every item is an action with a result you can check.

0. Before you turn voice mode on

Voice reaches the same agent you would otherwise type to, so it can do what that agent can do and nothing past it. The quality comes from the workspace it reads.

  • The project exists as a named workspace, not a chat window you keep reopening.
  • A written description of the project sits inside it: what the work is, who it is for, which files a fresh session reads first.
  • Prior work the agent is allowed to read lives in the project, not scattered across your disk.
  • Your calendar is connected, so anything time-bound resolves to real entries instead of a plausible guess.
  • The channel your team actually talks in is connected.
  • The task tracker is connected and shows current state, not last month's export.
  • Meeting notes land somewhere the agent can reach without you pasting them.
  • Repeatable jobs are written down as skills, so you brief by name instead of respelling the standard every time.
  • Run the test before anything else: take the task you delegate most often and ask whether an agent could find everything it needs without asking you. Whatever it could not find is the next thing to build.

1. Phrasing the request

Speaking is fast, and speed is where the scaffolding gets dropped. Name the context first, then the task.

  • The project name is in the first sentence, ahead of the work itself.
  • One task per thread. A sentence carrying two jobs produces one thread doing both of them thinly.
  • Dependencies are spoken out loud: say which thread waits on which approval, or the agent will run them as independent work.
  • Say where the output goes in the same breath as the request.
  • Say what finished looks like, in terms you can count or open.

Phrases to say

  • "In project X, start a separate thread that drafts the article."
  • "Work inside project X for all of today's tasks."
  • "Stop this thread, hand over what you have verified, and restart it inside project X."
  • "That cover is approved. Pass it to the article thread and carry on there."
  • "Save the result in project X under this campaign name, then tell me where it landed."
  • "Start this one now. It does not wait for the other thread."

2. While it runs

  • Check where each thread actually landed: inside the project, or in a loose chat outside it. A thread running in the wrong context looks perfectly healthy from the outside.
  • Landed outside? Stop it, ask it to hand over whatever it has verified, restart it in the project. Patching mid-flight leaves the gap inside the output.
  • Check again every time you dispatch something new. Threads opened later drift the same way the first ones did.
  • Pause the voice chat when you have nothing to say. The work keeps running and you come back when there is something to look at.
  • Pick the same conversation up on your phone when you leave the desk. The conversation is the persistent object, and voice is only how you reach it at that moment.
  • Keep a running list of what you started. Voice makes it easy to have several things finishing at once and lose track of which ones you actually looked at.

3. Reviewing the result

Voice removes the typing. It removes nothing downstream of "the agent says it's done."

  • Open every output. A message reporting that a file was created is not the file.
  • Screenshots: check placement paragraph by paragraph, not just that images turned up somewhere.
  • Screenshots: check the crop shows what the sentence beside it claims it shows.
  • Screenshots: look at every blurred region yourself before publishing. Automatic redaction and a near-miss look identical in a thumbnail, and a half-covered key or private address is an uncovered one.
  • Dashboards and summaries: spot-check rows against the source. Open the calendar entry, open the meeting note, confirm the figure came from there.
  • Start that spot-check with the entries that look most plausible. Data that is obviously wrong costs you nothing; data that looks right costs you a decision.
  • Pages: click through every link before you send the address to anyone.
  • Pages: open the page yourself, at full size, on the kind of device your reader will use.

4. One sentence to keep

A control surface only amplifies what it is pointed at. Voice pointed at an empty workspace gives you generic output faster than typing would. Voice pointed at a workspace holding your context and your written-down standards gives you work you would have done yourself, minus the keyboard.

  • Next session, start small: one request, one project, one task. Watch where it lands, then open what comes back.