I Stopped Typing for a Week. Here's What Happened.

A few weeks ago, I wrote about talking instead of typing as one of the essential habits of AI power users. I called it the single biggest speed multiplier most people aren't using and estimated I was roughly 3x faster speaking than typing.

Then I realised I was still mostly typing. So I decided to practice what I preach: seven days, a target of 95% voice input, across everything. Phone, laptop, emails, AI prompting, coding tools. The real number turned out to be closer to 4x faster using voice. Here's what happened.

The Numbers Don't Lie

I ran a simple test. I took a 200-word passage and timed myself typing it, then speaking it. Typing: 40 words per minute, with plenty of errors. Speaking: 190 words per minute, cleanly transcribed. That's not a marginal improvement. That's a completely different operating speed.

Over a full week of real usage, Wispr Flow measured my sustained average at around 160 WPM. Your pace will naturally vary depending on what you're saying, but even on a slow day, that's 4x faster than typing.

But speed alone isn't the story. When I typed that passage, the output was riddled with errors: missed letters, transposed words, the kind of mess that comes from fingers trying to keep up with thoughts. When I spoke it, the transcription came back clean. Faster and more accurate. That combination is what makes voice-first viable as a daily operating mode, not just a party trick.

The Tool That Makes It Work

The key enabler for me was Wispr Flow, though there are several tools in this space with similar capabilities. I'm not sponsored. I just liked it enough to move to the paid plan (£15/month) after a week of heavy use. It strips out the ums (of which there were many), the false starts, the verbal stumbles, and gives you clean output at speaking speed. Without that kind of cleanup layer, you'd spend half the time you saved on editing, and voice-first would stay a novelty.

It does smart things you don't expect. I dictated an email and it automatically formatted it: greeting at the top, line break, body paragraphs, bullet points where they made sense. I didn't say "new line" or "bullet point." It inferred the structure from how I spoke. That's not just transcription. That's a lightweight writing assistant baked into the tool.

One thing to watch: proper nouns and homophones. I tried to reference an app I’m building called "Encore" and the transcription rendered it as "on call." A perfectly valid phrase, just completely wrong. These errors are harder to catch than typos because they read as normal text. You need to actually review what was transcribed, especially for names and technical terms. Most of these tools have custom dictionary features where you can teach them your frequently used words. I hadn't set mine up yet, which is a lesson in itself: spend ten minutes configuring the tool before you go all-in.

Where Voice Changes the Game

Three modes of voice-to-create emerged during the week, each serving a different purpose.

1. Real-time dictation for when you need to get something written fast. Emails, messages, prompts, any task where you roughly know what you want to say. Wispr Flow or similar tools handle this. It's where the 4x speed gain shows up.

2. Conversational AI for when you need to think something through. ChatGPT's Advanced Voice Mode is superb for this. Instead of composing a linear prompt, you have a back-and-forth conversation where ideas develop naturally. Use this when you want to explore a problem space, not just execute a task. Claude's voice mode works here too, though I'd recommend changing the default voice setting, as the standard one feels quite robotic.

3. Voice notes for when ideas hit you and you're not ready to sit down and work yet. Record voice notes throughout the day when fragments of ideas hit you, then upload the transcripts to an AI with a prompt like "extract the most useful ideas from these notes." This meets you where you are. Not everyone wants a live conversation with an AI, but everyone can press record while walking the dog.

The Unexpected Upside

Here's something I didn't anticipate: voice makes you give AI more context, and that produces better output. When I type a prompt, I naturally (attempt to) write something tight and brief because every keystroke costs effort. When I speak a prompt, I give more background, more direction, more connection between ideas. Yes, it's messier. But the AI doesn't care about mess. It can extract the signal from a rambling spoken prompt far more easily than it can guess your intent from a brief typed one.

My ability to waffle is well-documented. Turns out, when you're talking to AI, that's actually a feature, not a bug.

Voice-first isn't perfect.

If you give this a proper go, here's what I learned to watch for.

Voice is looser than typing, and that's context-dependent. When you're prompting AI, the extra context is gold. When you're drafting an email or writing something structured, you'll want to tighten up. Knowing which mode you're in matters.

Social spaces get awkward. My wife thought I was talking to her when I was dictating in the living room. That's a real friction point in shared spaces. The workaround: these tools work at whisper volume too, when you need it. Feels odd at first. Works brilliantly.

Public feels weird, then doesn't. I used voice in a coffee shop and nobody batted an eye. For all anyone knew, I was on a call. The barrier is entirely in your own head, and it fades within a couple of days.

Some things stay keyboard-only. Passwords, sensitive information in public, anything requiring exact formatting or code syntax. That's the 5% that stays typed, and that's fine.

Your Voice-First Starter Kit

You don't need to go all-in like I did. Start here.

1. This week: Pick one task you do daily, like email drafting, and switch it to voice for five days. Use Wispr Flow, or the built-in dictation on your device. Notice the speed difference and whether you adjust your editing habits.

2. Next week: Try voice-prompting your AI tool of choice. Speak your next ChatGPT or Claude prompt instead of typing it. Give it the full context you'd normally trim. See if the output improves.

3. When you're ready: Try ChatGPT's Advanced Voice Mode for a thinking session. Don't type a prompt. Just talk through a problem and let the AI push back in real time. It's a different experience entirely.

The Atomic Builder is written by Faisal Shariff and powered by Atomic Theory Consulting Ltd — helping organisations put AI transformation into practice.