Talk to your agent: voice in, spoken pings out

In the last session you installed your agent. Now fix the two slowest parts of using it. Typing long, detailed prompts is slow, and most people then sit and watch the terminal scroll while the agent works. That is wasted time.

Two upgrades fix both: speak your prompts instead of typing them, and make the machine speak (or ping) when the agent finishes. By the end of this lesson you can start a job by talking, walk away, and hear "Your agent is done" from across the room.

What you will use

• The agent you installed in session 1 (Codex, Claude Code, or Grok Build)
• A Mac with built-in Dictation, or a Windows 10/11 PC with voice typing
• A microphone (the one built into your laptop is fine)
• 10 minutes to set two things and test them

Step 1: Turn on Dictation on your Mac

macOS has Dictation built in, and it works inside Terminal. Open System Settings and look under Keyboard for a Dictation section. Apple moves these menus around between macOS versions, so if it is not exactly there, type "Dictation" into the search box at the top of System Settings.

Turn Dictation on, then set the shortcut. The options include pressing a single key twice (Control twice is a common one). Pick something you can hit without looking.

Now test it. Start your agent in Terminal (codex, claude, or grok), hit your Dictation shortcut, and talk your prompt instead of typing it. Your words appear at the prompt. Stop dictating (usually the same shortcut or the Escape key, depending on your macOS version), then press Enter to send.

Do not try to speak like a programmer. Long, rambling spoken prompts are fine; agents handle them well. A shop owner can say: "Look at the sales CSV in this folder and give me the top ten products by revenue this month, put it in a file called top products, and flag anything that dropped more than twenty percent." A nonprofit director can say: "Read the volunteer signup spreadsheet and draft a short welcome email for each person who joined this week." Both of those would take two minutes to type and fifteen seconds to say.

Step 2: Turn on voice typing on Windows

Windows has the same thing under a different name. Click into the PowerShell window where your agent is running, press Win + H, and voice typing starts listening. Speak your prompt, and it types into PowerShell wherever your cursor is. Press Win + H again (or click the microphone) to stop, then Enter to send.

Same rule applies: talk naturally, in full sentences, with all the detail you would put in a good typed prompt.

Step 3: The quick "tell me when you are done" trick

macOS ships with a say command that speaks any text out loud. That gives you a done-ping with zero setup, two ways:

1. Put it in the prompt. End your request with: "When you finish everything, run this command: say 'All done'." The agent runs it as its last step.
2. Chain it yourself. If you run a long command directly, append it: python import_donors.py && say "import finished".

Example: "Build me a one-page price list from prices.csv, and when you finish, run: say 'Price list is ready'." Then go help a customer or return a call. The computer will tell you when to come back.

On Windows there is no say, but PowerShell can beep. This one-liner works in the prompt trick above:

[console]::beep(800,400)

Step 4: The proper way in Claude Code: a Stop hook

Asking for the ping in every prompt gets old. Claude Code supports hooks: commands that run automatically on certain events. A Stop hook runs every time the agent finishes responding. Put this in ~/.claude/settings.json (create the file if it does not exist):

{
  "hooks": {
    "Stop": [
      {
        "hooks": [
          { "type": "command", "command": "say 'Your agent is done'" }
        ]
      }
    ]
  }
}

Now every job, big or small, announces itself when it lands. On Windows, replace the say command with a beep:

{ "type": "command", "command": "powershell -c [console]::beep(800,400)" }

Do not want to edit JSON by hand? Ask the agent itself: "Add a Stop hook to my ~/.claude/settings.json that runs say 'Your agent is done'." Then restart Claude Code and test it.

If a spoken announcement is too much (shared office, sleeping baby), Claude Code also has a quieter terminal bell option. Type /config inside Claude Code and look for the notification preferences; the exact wording shifts between versions, but a bell or notification setting is what you are looking for.

Step 5: Codex and Grok Build

Codex and Grok Build each have their own settings and add notification features quickly, so check their docs or settings screens for hooks or notifications. Either way, the Step 3 trick works everywhere: end your prompt with "when you finish, run: say 'done'" (Mac) or the PowerShell beep (Windows), and any agent will ping you.

You are done when...

You can start your agent, hit one shortcut, speak a full detailed prompt without touching the keyboard, send it, walk away, and hear the machine announce that it finished. Test it with a real job: have it summarize a folder of documents or draft five emails, and leave the room while it works.

This matters in the next sessions, when you will hand your agent jobs that run for many minutes at a time. The habit you just built (speak the job, leave, come back when called) is how you run those without babysitting the terminal.

Template included: Members get a printable cheat sheet with the Mac Dictation and Windows voice typing setup steps, the say and beep one-liners, the ready-to-paste Stop hook JSON for Claude Code, and five example spoken prompts to try. Find it in your member dashboard under Templates.