Home/Docs/Voice Workflows

Voice in, voices out: less friction, clearer context

Updated October 31, 2026. Practical habits for working with AI—not a requirement to install a custom voice stack.

Sometimes the bottleneck isn't the model. It's getting the thought out of your head and into the text box. Dictation lets you explain an idea at the pace you speak, instead of making your fingers keep up. If speaking works well for you, try it before buying another tool or writing another elaborate prompt template.

Adam recommends two complementary habits: use your voice to capture your thinking, and use distinct agent voices to recognize context when replies arrive. They're different jobs: speech-to-text on the way in, text-to-speech on the way out.

Start with the dictation you already have

Windows: Click in the prompt's text field, then press Windows + H to open voice typing. A working microphone and internet connection are required for the standard online voice-typing path.

Mac: Enable Dictation in System Settings → Keyboard → Dictation. Use the microphone/Dictation key if your keyboard has one, your configured Dictation shortcut, or Edit → Start Dictation in a supported app.

Mac keyboards and shortcut settings vary; don't assume one function key or double-press works everywhere. Check the shortcut shown on your own Mac. Official instructions: Windows voice typing, Mac Dictation.

Operating-system dictation enters text into a field. It isn't the same as an AI app's live voice conversation, which may submit speech or take turns differently. Use an editable text field when you want to inspect the prompt before it reaches the model.

Speak freely, then edit deliberately

  1. Explain the situation. What are you trying to do, who is it for, and what context matters?
  2. Name the outcome. A plan, a comparison, a private draft, or an implemented change are different requests.
  3. State the boundaries. Relevant sources, constraints, and what needs your approval.
  4. Stop dictation before correcting the text. Delete abandoned wording, transcription mistakes, and mid-sentence reversals manually. Don't make the model choose between conflicting instructions you no longer mean.
  5. Read the final prompt once, then submit it. Check names, dates, numbers, negations, and action permissions in particular.

For example, a spoken false start might become:

Draft an email—actually don't email anyone, make it a checklist—about getting the project ready.

Edit it to:

Prepare a private checklist for getting the project ready. Include missing information and next steps. Do not create or send an email.

You don't need a perfect script. The aim is to keep the useful context while removing the contradictory leftovers. Dictation reduces the work of capturing a thought; a short edit makes that thought easier to act on.

Let the voice tell you which context arrived

Adam describes using different text-to-speech voices for different specialist contexts. When an agent speaks through the Mac mini's speakers, the familiar voice helps him recognize the context before reading the screen. This is a reported working habit, not a new end-to-end test of every voice service.

Think of the voice as an audible label, not just decoration. Keep the mapping stable: one voice for teaching, another for infrastructure, another for personal planning, for example. You can do this with ordinary licensed or built-in voices; cloning isn't required.

  • Keep spoken updates short: identify the role, say what changed, and name the next decision.
  • Keep the written result, sources, and explicit role label available. A familiar voice is a convenience cue, not authentication or evidence that its answer is correct.
  • Queue updates rather than letting multiple agents speak over one another. Offer quiet hours, mute, and a text-only path.
  • Choose where private information may be spoken. A shared room and headphones are different output destinations.
  • Use voices you own or have permission to use. Pick local or hosted speech based on the actual tools, permissions, and data requirements.

Voice output is optional. It shouldn't hide a failed task, replace review, or make silence look like a successful check.

Try it in five minutes

Dictate a non-sensitive request into an editable text box. Stop, remove one false start, and check the final instructions before submitting. Then listen to a short response with an available read-aloud tool, if you want to. Notice whether speaking helps you supply more useful context, and whether listening helps you catch something you missed on screen.

Don't assume dictation is offline just because it is built into the OS. Windows' standard voice typing uses online recognition; Mac processing depends on the supported configuration, and its Keyboard settings explain that path. Check before dictating sensitive material. The AI provider receiving the final prompt is a separate data route.

Start with this habit, not a new server. For specialist speech output, use the Chatterbox and completion-hook guide, including its downloadable example. The voice-system design prompt helps adapt it. Pair it with model routing, onboarding, and cost controls.