Lead with voice: My Setup for AI Agents

Thomas Prommer Technology Executive & AI Strategy Consultant Connect on LinkedIn

Published: October 19, 2026

Updated: October 27, 2026

Observation

Why will voice become our primary input?

We will use voice as our primary way to work with computers, ahead of keyboard and mouse. That is my conviction, based on how I work with AI agents. When I type a prompt, I cut it short. When I speak, I give the agent far more context: the background, the constraints and what I actually want.

Why is Apple’s built-in Dictation not enough?

AI voice recognition has improved a lot in my experience, but Apple’s built-in Dictation on unpersonalized defaults still gets my names and terms wrong. Every miss sends me back to the keyboard, and after a few of those I stop dictating. A tool that knows my vocabulary changes how often I reach for the keyboard.

My setup

What do I actually run?

I use Wispr Flow with a Shokz headset across my devices. A personal dictionary and a set of replacements cover the names and terms I say every day, and the microphone is set up for the room I work in. Long prompts to a coding agent go into the terminal by voice. Short commands and edits stay on the keyboard.

How do I hold myself accountable?

I chart my voice share. The measured window is narrower than the setup I run today: one Mac, the Ghostty terminal, and completed transcriptions from two apps, Wispr Flow and DictationPal, that are associated with Ghostty. Typed characters come from the hardware keyboard. Pasted text is excluded, because pasting a log file into a prompt says nothing about how I produced input. The daily share is voice characters divided by voice plus typed characters.

Typing data starts on September 21, 2026. Across the 14 complete days from September 21 to October 4, voice made up 122,821 of 515,393 counted characters, or 23.8%. Single days ranged from 1.0% to 55.3%. The best day, October 4, had 40,161 voice characters out of 72,618, with 32,457 typed. October 5 shows 69.6%, but it was a partial day when I pulled the data, so I leave it out.

The spread is the main finding. My voice share is not a stable habit yet, and the data does not tell me why it swings.

What does the count get wrong?

It counts transcriptions, and none of them is confirmed to have landed in the terminal. A completed DictationPal record in particular does not prove the text was pasted anywhere, so the share is more likely too high than too low. The 90-day chart on the practice page has voice data from July 8, but those earlier days have no typed counterpart, so no share exists for them.

Characters also measure input volume only. They say nothing about time saved, whether spoken prompts get better results than typed ones, or how much output the walks below add. A weekly loop and an email report now run on these numbers, so I can compare weeks, but that comparison will not answer the speed question.

Your next steps

Step 1: Which tool should you start with?

Sign up for a specialized dictation solution and use it across your devices, rather than the dictation built into one operating system. If you choose Wispr Flow, this is my referral link: wisprflow.ai/r?THOMAS22399, with my referral code THOMAS22399. I make no promise about trials or discounts.

Step 2: How do you teach it your vocabulary?

Build a personal dictionary with the names and terms you say every week: people, products, tools and project names. Then keep maintaining it. When the tool gets a word wrong, fix it in the dictionary or add a replacement right away, while you still remember what you said.

Step 3: How do you send a prompt without the keyboard?

Set up voice or keyboard commands, such as “press enter” to send. In Wispr Flow this is an experimental setting. Say the phrase, then stop recording explicitly, because the phrase alone does not stop it. Leave the command out when you want to review the text before it goes. Another setup page covers the spoken command to press Enter in Wispr Flow.

Step 4: How do you push your voice share up?

Vibe-code a small tool that tracks your voice share, or build a daily accountability habit, and push the share up over time. A tally at the end of each day works: roughly how many prompts you spoke and how many you typed. Write down the date you start, because a share without a coverage date is hard to read later. A prompt count and my character share measure different things, so don’t compare them. After two weeks, look at what your high days had in common before you try to lift the low ones.

Bonus: What happens when you take it for a walk?

Go for a walk and talk to ChatGPT or Claude as a conversational agent, with the screen off where the app supports it. In ChatGPT the option is under Settings > Voice > Background conversations. Use the walk to test whether your personal context engine and AI executive assistant can do actual tasks by voice. That is where I see the impact compound into more output.

The practice card for this setup is Let voice carry more of the input, where the voice-input chart sits with the other measured entries in the Compounding CEO playbook.