Talk to Your Claude Code

The most effective way I’ve found to work with Claude Code is to stop typing at it and start talking to it — literally, out loud. Voice carries the context that typing strips away, and context is what separates slop from good work. This guide is the whole setup: Wispr Flow, a dictation app that types wherever your cursor is, plus what I’ve learned about running it across Claude Code, Gmail, and Messages — and how to dictate so a machine can actually build from it.

Most mornings I’m at Mocama, the coffee shop in Fernandina Beach — eggs, a French press, a corner table — talking to my machine while the tables around me pretend not to notice. This page, like most of what I make now, started as speech. Here’s how to set it up yourself, cold, in about ten minutes.

Why voice: spoken context is what separates slop from good work the why

When you type a prompt, you compress. Typing is slower than thinking, so you trim the backstory, skip the doubts, drop the “here’s why this matters” — and the machine gets an abbreviation of your intent. Then you wonder why the output feels generic. It feels generic because the input was.

When you speak, the whole thought comes through at once — the goal, the constraints, the client’s last round of feedback, the thing you’re worried about, the example you half-remember. That surrounding material is context, and context is the difference between AI output you delete and AI output you ship. The major AI companies now say the same thing: talk to your coding agent instead of typing at it, because speech moves the conscious flow of thought into the machine at something close to full resolution.

A useful way to hold it: the machine is a very fast colleague with no memory of your week. A colleague who heard the full story gives you better work than one who got the text-message version. Voice is how you tell the full story without it costing you twenty minutes of typing.

The people who build these tools say the same thing receipts

This isn’t one man’s coffee-shop habit. The people who build these tools say the same thing, in public. Boris Cherny — the creator of Claude Code at Anthropic — put voice dictation in his own list of Claude Code tips:

“Use voice dictation. You speak 3x faster than you type, and your prompts get way more detailed as a result.”

— Boris Cherny, creator of Claude Code, Anthropic

Read that second clause again — it’s the whole argument. Not just faster: your prompts get way more detailed as a result. Speed is what gets measured; detail — context — is what changes the work. Andrej Karpathy said the quiet part in 2023 — “The hottest new programming language is English” — and then showed his hand in the post that coined vibe coding:

“I just talk to Composer with SuperWhisper so I barely even touch the keyboard.”

— Andrej Karpathy, OpenAI co-founder

And this is now product, not prediction: in March 2026 Anthropic shipped built-in voice mode in Claude Code — transcription tuned for coding vocabulary, live into the prompt input. The company that makes the tool built a microphone into it. That’s as clear as a signal gets.

Set up Wispr Flow setup

My dictation layer is Wispr Flow — a small app that listens while you hold a hotkey and then types clean text into whatever field your cursor is in. Any app, any text box. It handles natural speech well: it drops the ums, follows you through mid-sentence U-turns, and punctuates. Setup, start to finish:

1. Install it. Download the app from wisprflow.ai, open it, and create an account when it asks.
2. Grant the two permissions. It needs your microphone (to hear you) and, on a Mac, an accessibility permission (so it can type into other apps). macOS will prompt you for both; if you miss one, they live under System Settings → Privacy & Security.
3. Learn the hold-to-talk key. Flow works push-to-talk: hold the hotkey, speak, release, and the text lands where your cursor is. On my Mac the trigger is the fn key — the app shows you yours during onboarding and lets you change it.
4. Test somewhere low-stakes. Open Notes, click into a blank page, hold the key, and say a full sentence with a turn in it — “actually, scratch that, start with the budget” — and watch what it does. That test convinced me faster than any feature list.

That’s it. There is no step five. From here on, every text field on your computer is a place you can talk.

Where I use it: Claude Code, Gmail, Messages everywhere

Because Flow types wherever the cursor is, the workflow is the same everywhere — what changes is what I say.

Claude Code is where voice earns its keep. My typed prompts used to be two lines; my spoken ones are two paragraphs, and the work comes back better in direct proportion. I put the cursor in the Claude Code prompt, hold the key, and brief it like a colleague:

A spoken prompt, verbatim — goal, backstory, constraint, and where the human stays in the loop. Try typing all that eight times a day.

Gmail gets the same treatment. I dictate replies the way I’d return a phone call — and they come out sounding like me on a good day, because they started as my actual voice. The emails I dictate are warmer than the ones I type; typing makes me terse, talking doesn’t.

Messages is the sleeper. Real texts to real people, at conversation speed, with both hands free for the coffee. Once the hotkey is muscle memory you stop noticing which app you’re in at all — that’s the point where voice stops being a feature and becomes how you work.

How to dictate well craft

The tool takes ten minutes. The craft took me longer, mostly because dictating well means unlearning typing habits. Four rules cover most of it:

Talk in intentions, not commands aim

Don’t dictate the keystrokes you would have typed — say what you’re trying to achieve and why. “Make this page easier to skim for someone deciding whether to hire us” beats “shorten the paragraphs.” The model is good at finding the how when you give it a real what.

Give the backstory context

Everything you’d tell a new collaborator in the hallway — who it’s for, what happened last round, what failed before, what you’re afraid of. Out loud this costs fifteen seconds. It’s the highest-value fifteen seconds in the whole workflow.

Don’t self-edit mid-stream flow

When you fumble a sentence, keep talking — say “actually, what I mean is” and continue. Fix nothing. Flow cleans up the stumbles, and the model reads corrected intent just fine. The moment you stop to repair a sentence you’ve started typing with your mouth, and you lose the whole reason to talk.

Trust the model to structure it trust

Your spoken paragraph will be a ramble — that’s fine. Structure is the machine’s cheapest trick. Deliver the thinking in whatever order it arrives; ask for the shape you want at the end (“turn all that into a plan before you touch anything”) and let it do the sorting.

About talking in public get over it

Yes — you will sit in a coffee shop and talk to a laptop, and someone will glance over. Here’s the perspective that cured me: the man two tables over is on speakerphone with his contractor and nobody has looked at him once. We long ago agreed that talking to a person through a machine is normal. Talking to a machine through a machine is one small step down the same road, and the work on the other side of that step is better. Order the eggs, hold the key, and say the whole thought.

The longer story of why this works — Homer’s spoken epics, Cicero and the secretary who invented shorthand to keep up with him, Milton dictating Paradise Lost blind — is this week’s letter in my newsletter, The Human in the Loop. The short version: writing through a medium is ancient. The medium just got faster.

— Joe