AI Workflow

Voice to AI Prompt: Dictate Prompts to ChatGPT, Claude and Gemini

To dictate a prompt to ChatGPT, Claude or Gemini, click into the chat box, hold Ctrl + Shift + Space, say what you want, and let go. AirTypes types your spoken prompt into the box and you press Enter. Plain mode is free and runs offline on your own computer; Polished turns rambling speech into a tidy prompt at 120 minutes of talking for every dollar.

This is what it looks like in practice. You are stuck on a React bug, the ChatGPT tab is open, and you think out loud:

You say

"um can you explain why my react component re-renders every time the parent updates, I'm passing an object as a prop and uh I already tried memo"

→
Polished types into the chat box

"Why does my React component re-render every time the parent updates? I'm passing an object as a prop, and I already tried memo."

Same question, same details, none of the "um". You read it, press Enter, and the model gets a clear prompt that took four seconds to say. This post covers the tooling: how it works, when to use Plain or Polished, and how to set it up. For the prompting technique itself (how to structure a spoken prompt, with examples per tool), see the companion voice prompting guide for ChatGPT, Claude and Gemini.

The short version, with real before and after examples, is on the voice to prompt page.

The problem with typing AI prompts

Everyone who uses ChatGPT, Claude or Gemini regularly runs into the same thing: a good prompt takes effort to type. It needs context, the goal, the format you want, and the constraints that matter. Typing all of that is slow, so most people cut corners. The prompt shrinks to "fix this" or "make it shorter", and the model spends its first reply guessing.

Speaking removes most of that friction. You explain the problem the way you would explain it to a colleague, reasons and constraints included, because one more sentence costs a few seconds instead of half a minute of typing. You end up with longer, more specific prompts for less effort.

The catch is that spoken language is messy. People say "um", restart sentences, and repeat themselves. Raw speech pasted into a chat box works, since models are forgiving, but it is harder for you to read back and check before you send. That is the gap the two AirTypes modes cover in different ways.

How dictating a prompt works

AirTypes is a desktop dictation app for macOS and Linux, plus Dictate in your browser (new in November 2026), which copies the Polished text for you to paste. The desktop app does not plug into ChatGPT or any other AI tool. It types at your cursor, wherever that is, using native keyboard injection. That is why it works everywhere you can type a prompt:

  • ChatGPT, Claude and Gemini in the browser or in their desktop apps
  • Other chat tools and AI search boxes
  • AI panels inside your editor, notes app or office suite
  • Coding agents in the terminal (see voice coding with AI agents)
  1. Click into the prompt box.
  2. Hold the hotkey (default Ctrl + Shift + Space) and speak.
  3. Let go. In Plain, Whisper transcribes on your CPU. In Polished, the recording goes to AirTypes for transcription and cleanup.
  4. The text is typed into the box. AirTypes stops there.
  5. You read it and press Enter. Nothing reaches the AI tool until you do.

Because AirTypes only types, the AI side is unchanged. You keep using the model, account and chat history you already have, and the AI tool handles your prompt exactly as if you had typed it. AirTypes is not affiliated with OpenAI, Anthropic or Google.

Plain: private prompts, offline

Plain is the free mode. Whisper runs offline on your own CPU and 0 bytes of audio are uploaded, so the recording of you thinking out loud stays on your computer. It works in airplane mode and has no time limit.

Plain does light cleanup on your machine: it drops a leading "um" or "uh", capitalizes the first letter and adds a period. Everything else is your words as you said them. For many prompts that is exactly right, because the model does not care about a stray "like", and you get the most literal version of what you asked.

Plain is English only. Pick one of its two model tiers in Settings: Swift (~78 MB) is the fastest, and Crisp (~148 MB) is the tier we recommend. When a prompt is full of technical terms and product names and you need more accuracy than Crisp, use Polished for that prompt. The Whisper model sizes guide explains the trade-offs.

One honest note: Plain keeps your audio private, but the prompt text still goes to whichever AI tool you send it to, under that tool's own policies. Plain removes AirTypes from the picture; it does not change what ChatGPT, Claude or Gemini receive.

Speak your next prompt instead of typing it

Plain is free and offline on macOS and Linux. Add Polished credit when you want rambling speech cleaned up.

Download AirTypes

Polished: rambling speech in, tidy prompt out

Polished sends the recording to AirTypes to transcribe and clean it up. A large speech model transcribes it on our servers, then an AI removes filler words and false starts and fixes grammar and punctuation, keeping your meaning, numbers and names. It does not answer your prompt or add instructions you did not say. It makes what you said readable. These are real Polished outputs:

You say

"so I need you to rewrite this email to sound more polite but uh keep it short, like three sentences max"

→
Polished types

"I need you to rewrite this email to sound more polite. Keep it short, like three sentences max."

You say

"refactor the payment service so it uh retries failed webhooks three times with backoff and like logs each attempt"

→
Polished types

"Refactor the payment service so it retries failed webhooks three times with backoff and logs each attempt."

You say

"ok so write me a um python function that takes a list of dates and uh returns the ones that fall on a weekend, and add tests, use pytest not unittest"

→
Polished types

Your own words, unchanged. The guard kept the original because the rewrite drifted.

The guard, honestly

That last example matters. Polished has a guard: it checks that the rewrite keeps every number you said and stays close to your length. If not, it catches the drift and you get your own words (the raw transcript) instead, and the Home screen notes that it kept them as you said them. For prompts that ask for code, the guard sometimes returns your words unchanged rather than risk altering the request. That is a deliberate trade. A prompt with an "um" in it still works; a prompt where "pytest" quietly became "unittest" does not.

Polished can also drop a greeting such as "hi" or "hey" at the start. In a prompt that rarely matters.

What it costs

Polished is prepaid credit: $5 buys 600 minutes of speech, charged by the second. A 30-second prompt costs under 1¢. Credit never expires and there is no subscription. When credit runs out, Polished pauses and Plain keeps typing your prompts for free. See pricing.

Prompts in your own language

Polished works in the language you speak, while Plain is English only. That includes English, Hindi, Bangla, Japanese, German and most other spoken languages, so try yours. If you prefer to prompt in English but think faster in your own language, turn on Always translate to English in Settings: speak your language, and English is typed into the chat box. On translations the guard checks numbers only, since the wording cannot be compared word for word, so check names and numbers before you press Enter.

Plain or Polished: which to use for prompts

SituationUseWhy
Confidential project, client names, unreleased workPlainAudio never leaves your computer in Plain
Offline, on a plane, flaky Wi-FiPlainWorks with no connection
Short, clear prompts you say in one goPlainLittle to clean up, and it is free
Long, thinking-out-loud prompts you want to read backPolishedFiller and false starts removed, punctuation fixed
Prompting in a language other than EnglishPolishedPlain is English only
Older or slower laptopPolishedNothing heavy runs locally

You switch modes on the Home screen in one click, so you can use Plain for the private prompt and Polished for the next one.

Why not the chat app's own voice button?

Many AI chat apps now include some form of voice input inside their own window, and inside that window it works fine. The difference is scope. A built-in button works in one app. AirTypes is one hotkey that works in every app: the ChatGPT tab, the Claude desktop app, the Gemini page, the AI panel in your editor, your terminal, and the email you paste the answer into afterwards. Your habit, your hotkey and your model choice stay the same across all of them.

You also get the choice of where your audio goes. In Plain, it is transcribed on your own CPU. In Polished, it goes to AirTypes. Either way you see the text before anything is sent.

Who benefits most

Developers

Bug descriptions, refactor requests and review questions are prose, and prose is faster to say than type. Speak the intent, type exact identifiers. More in voice coding with Claude Code, Cursor and Copilot Chat.

Writers and marketers

Describe audience, length and angle out loud. Spoken briefs tend to include the context that typed ones leave out.

Non-native English speakers

Polished fixes grammar while keeping your meaning, or lets you speak your own language and get English with Always translate to English.

Researchers and students

Long questions about a paper or a dataset are easier to say than to type. See voice dictation for academics.

People with wrist or hand pain

Every prompt you speak is one you do not type. The dictation for RSI guide covers setups that reduce keyboard time.

Set it up in five minutes

  1. Install AirTypes

    Download for macOS or Linux and create a free account at first launch. The Windows desktop app is in development; join the waitlist. On Windows today, Dictate in your browser instead: click the mic, speak, and paste the Polished prompt into the chat box. AirTypes for Chrome, an extension that types at the cursor in text boxes in Chrome, is coming soon.

  2. Pick a Plain model and allow permissions

    Choose a tier (Crisp is the one we recommend). On macOS, allow Microphone and Accessibility when asked.

  3. Choose your mode on Home

    Plain to start. For Polished, add credit from the Home screen; a payment page opens in your browser.

  4. Pick Hold or Toggle

    Hold suits short prompts. Press to Toggle is easier for a long brief, since you do not keep the keys pressed.

  5. Prompt by voice

    Click into the chat box, hold the hotkey, speak, let go. Read it, then press Enter.

Full setup details are in how to use AirTypes. To get better answers from what you say, read the voice prompting technique guide.

FAQ

Can I dictate prompts to ChatGPT, Claude and Gemini with AirTypes?

Yes. AirTypes types at your cursor in any desktop app, so you click into the chat box of ChatGPT, Claude, Gemini or any other AI tool, hold the hotkey, speak your prompt and let go. The prompt is typed into the box and you press Enter yourself. There is no plugin, and AirTypes is not affiliated with OpenAI, Anthropic or Google.

Does AirTypes send my prompt to the AI for me?

No. AirTypes only types the text. Nothing is sent to the AI tool until you press Enter, and then the AI tool handles it exactly as if you had typed it.

Can I dictate private prompts without uploading audio?

Yes, in Plain mode. Whisper runs on your own CPU, no audio is uploaded and it works offline. Plain is free and English only. Keep in mind that the AI tool still receives the prompt text once you send it.

What does Polished do to a spoken prompt?

Polished sends the recording to AirTypes to transcribe and clean it up. It removes filler words and false starts and fixes grammar and punctuation while keeping your meaning. It does not write the prompt for you. A guard keeps your numbers and catches rewrites that drift too far from what you said; when it trips, you get your own words instead.

Get AirTypes, free

Plain types your prompts free and offline, with no time limit. Polished is optional prepaid credit when you want rambling speech cleaned up.

Download for macOS and Linux