Here is a spoken prompt and what lands in the chat box with Polished:
"so I need you to rewrite this email to sound more polite but uh keep it short, like three sentences max"
"I need you to rewrite this email to sound more polite. Keep it short, like three sentences max."
Notice what made it work: a goal (rewrite, more polite) and a constraint (three sentences max), said in one breath. This guide is about that part, the technique of speaking prompts that get good answers. If you want the tooling side first (how dictation types into a chat box, and when to use Plain or Polished), read voice to AI prompt: dictate prompts to ChatGPT, Claude and Gemini.
The short version, with real before and after examples, is on the voice to prompt page.
What voice prompting actually is
Voice prompting means speaking your prompt instead of typing it. You click into the message box of ChatGPT, Claude, Gemini or any AI tool, hold a hotkey, talk, and let go. A speech model turns your voice into text and a dictation app types it into the box. You read it and press Enter. The AI tool answers in its own window, exactly as it would for a typed prompt.
Nothing about the AI side changes. You keep your account, your model choice and your chat history. What changes is the input: how fast you can produce a prompt, and how much context you are willing to put in it.
Why people switch to voice prompting
- Speed. Most people speak much faster than they type. Commonly cited figures put comfortable speech around 130 to 160 words per minute and typing around 40 to 70.
- Better prompts. Typed prompts get terse to save keystrokes, which gives the model less to work with. Spoken prompts tend to include the reasons and constraints, because one more sentence costs a few seconds.
- Less strain. Long prompts are a lot of typing. If your wrists complain by late afternoon, moving prompts to voice is an easy win. See dictation for RSI and wrist pain.
The trap is that speaking also makes it easy to ramble. A spoken prompt without structure is long but vague. The next two sections fix that.
The structure of a spoken prompt
Good prompts, typed or spoken, carry the same four things. When you speak, order matters more than it does on a keyboard, because you cannot easily go back and insert the goal at the top. So say them in this order:
| Part | What to say | Spoken example |
|---|---|---|
| 1. Goal | The one thing you want, as a verb | "Rewrite this product description..." |
| 2. Context | Who it is for, what you already tried, what it is part of | "...it's for a landing page aimed at small accounting firms..." |
| 3. Constraints | What must stay, what must not happen, limits | "...keep the price and the feature names exactly, no buzzwords..." |
| 4. Format | Length, shape, how many options | "...give me three versions under fifty words each." |
Spoken in one take, that is a twelve-second prompt that a model can act on without asking a single question. Typed, most people would have sent "make this better".
Why goal first
When you start with backstory, you tend to wander and forget the actual ask until the end, if at all. Starting with the verb ("summarize", "rewrite", "compare", "explain", "find the bug") commits you and the model to a task. The context that follows is then interpreted in light of it.
Why constraints out loud
Constraints are the part people skip when typing, and the part that most often saves you a second round. "Keep the numbers as they are." "Do not change the API." "British spelling." "No bullet points." Each takes two seconds to say.
Optional fifth part: done when
For bigger tasks, end with how you will judge the result: "done when it runs without the warning and the tests still pass", or "good enough if a non-technical manager would understand it". It gives the model a target and gives you a checklist.
How to speak it: five habits
- Think one sentence ahead. You do not need the whole prompt planned, just the next sentence. This becomes automatic after a few days of dictating.
- Do not correct mid-sentence. If you misspeak, keep going and fix it when you read it back, or say the corrected sentence again. Stopping and restarting is what makes dictation feel slow.
- Use hold for short prompts, toggle for long ones. In AirTypes, Press to Toggle lets you talk through a long brief without holding three keys.
- Leave exact strings to the keyboard. Say "the error from the checkout page", then paste the actual error text. Say "the function that parses dates", then add the file reference by hand.
- Read before you press Enter. Two seconds catches a misheard name. It is cheaper than a wasted reply.
Voice prompting ChatGPT
ChatGPT has voice input inside its own app, which is fine while you are in that app. A system-wide dictation hotkey gives you the same habit in ChatGPT and everywhere else, including the email or document you paste the answer into. Click into the ChatGPT message box (browser or desktop app), hold the hotkey, speak, let go, read, send.
ChatGPT is a common pick for quick explanations and drafts. A spoken prompt that follows the structure:
Say this
"Explain the difference between a Roth and a traditional retirement account. I'm thirty, self-employed, and I've never invested before. Don't recommend specific funds. Use a short table and then two sentences on which questions I should ask an adviser."
Goal (explain the difference), context (age, self-employed, beginner), constraint (no specific funds), format (table plus two sentences). Under fifteen seconds to say.
Voice prompting Claude
Same method: click into the Claude message box in the browser or the desktop app and dictate there. Claude is a common choice for long-form writing and for working through long documents, which suits voice well, because spoken prompts tend to be longer and more detailed anyway.
A pattern that works: paste or attach the document by keyboard, then speak the instruction.
Say this (after attaching the document)
"Read this contract and list every clause that limits what I can do after the project ends. I'm the freelancer, not the client. Quote the clause number for each one, and flag anything unusual compared with a standard services agreement. Plain English, no legal advice."
Notice the context sentence: "I'm the freelancer, not the client." That one line changes the whole reading, and it is exactly the kind of thing people leave out when typing.
Voice prompting Gemini
Click into the Gemini prompt box and dictate. Gemini is often used for research-style questions and for summarizing long material, so spoken prompts that state the audience and the shape of the answer do well.
Say this
"Summarize the main arguments for and against a four-day work week for a team of about twenty people. I'm presenting this to my manager on Friday. Keep each side to four bullet points, and finish with one question I should expect her to ask."
The same structure carries across every tool. The model names change; the four parts do not. It also works in any other AI tool you can type into, including AI panels inside editors and office apps, and coding agents in the terminal (see voice coding with AI agents).
Follow-ups by voice
The first prompt is rarely the last. Follow-ups are short, specific and frequent, which makes them the single best place to use voice. Instead of typing "shorter", say what you mean:
- "Good, but cut the second option and make the third one more formal."
- "Keep the table, drop the intro paragraph, and add a column for cost."
- "That's wrong for my case, because we already use a managed database. Redo it with that in mind."
- "Explain the second step again like I've never used a terminal."
Each of those is two to five seconds of speech and gives the model a clear direction. Typed, most people would send one word.
Spoken prompt patterns that work
You do not need saved templates. These are openings you can say from memory, then fill in with the specifics:
- Rewrite: "Rewrite this [email / paragraph] to sound [more direct / warmer]. Keep [the dates / the numbers]. Under [N] sentences."
- Reply: "Draft a reply to this message. I want to say [yes / no / not yet] because [reason]. Friendly, short."
- Summary: "Summarize this for [audience]. Keep every number. [N] bullet points, then one sentence on what to do next."
- Compare: "Compare [A] and [B] for [my situation]. Table with [criteria]. Say which you'd pick and why in one line."
- Explain: "Explain [thing] to someone who [knows X but not Y]. One analogy, no jargon, under [N] words."
- Code task: "Write a [language] function that [does X]. Use [library], not [other library]. Include tests."
- Commit message: "Write a commit message for this diff. Conventional Commits, subject under seventy-two characters."
- Translate: "Translate this into [language]. Keep it informal. Return only the translation."
Every pattern starts with the verb and includes at least one constraint. That is the whole trick.
Accuracy tips
Push-to-talk beats always listening
Recording only while you hold a key gives cleaner audio than always-on listening: no stray room noise, no half-sentences picked up while you think. AirTypes records only while the hotkey is held (or between two presses in toggle mode).
Pick the right model tier
In AirTypes Plain, Whisper runs on your CPU and you choose between two tiers. Swift is the fastest, and Crisp is the tier we recommend. Both are small models, so they make more mistakes on names and technical terms; when a prompt is full of jargon, use Polished, which transcribes with a large speech model on the server. The Whisper model sizes guide covers the trade-offs.
Fix the microphone before the model
A headset or a quiet room improves accuracy more than any setting. A laptop mic next to a fan makes every speech model look bad.
Let Polished handle the mess, not the meaning
If you think out loud a lot, Polished removes filler words and false starts and fixes punctuation before the prompt lands in the box. It aims to keep your meaning, names and numbers, and a guard keeps your numbers and catches rewrites that drift too far from what you said; when it trips, it types your own words instead. For prompts that ask for code, that guard sometimes leaves your words unchanged; the prompt still works. More examples in the voice to AI prompt post.
Prompt in your own language
Plain is English only. Polished works in the language you speak (English, Hindi, Bangla, Japanese, German and most other spoken languages; try yours), and its Always translate to English setting lets you speak your language and get an English prompt typed. On translations the guard checks numbers only, so check names before you send.
Privacy model: what goes where
A voice prompt passes through two separate steps, and each has its own privacy story:
- Speech to text. In AirTypes Plain, Whisper runs on your own CPU and 0 bytes of audio are uploaded. In Polished, the recording is sent to AirTypes to transcribe and clean it up.
- The AI tool. When you press Enter, ChatGPT, Claude or Gemini receives the prompt text under that provider's own policies, exactly as if you had typed it. If you run a model locally with its own chat window, Plain plus that local model keeps the whole round trip on your machine.
For confidential prompts, use Plain. For the regulated-industry side, see enterprise offline voice recognition, and for a broader look, is voice dictation private?
A 5-minute setup
-
Install AirTypes
macOS and Linux today; the Windows desktop app is in development (join the waitlist). Get the build for your OS and create a free account. On Windows today, Dictate in your browser instead (new in November 2026): 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.
-
Pick a Plain model
Crisp is the tier we recommend; Swift is smaller and faster. On macOS, allow Microphone and Accessibility.
-
Choose Plain or Polished on Home
Plain is free and offline. Polished is prepaid: $5 buys 600 minutes of speech.
-
Check your hotkey
The default is Ctrl + Shift + Space. Change it in Settings if an app already uses it.
-
Speak a structured prompt
Click into the chat box, hold the hotkey, say goal, context, constraints, format. Let go, read, press Enter.
FAQ
Can I dictate prompts directly to ChatGPT?
Yes. ChatGPT has voice input inside its own app. To prompt by voice from one hotkey that also works in every other app, use a desktop dictation tool like AirTypes: click into the ChatGPT message box, hold the hotkey, speak, let go, read the prompt and press Enter.
Can I voice-prompt Claude?
Yes. Click into the Claude message box in the browser or the desktop app and dictate your prompt there. AirTypes types at the cursor, so it needs no integration with Claude.
Does Google Gemini work with voice prompting?
Yes. The same method works in the Gemini prompt box, or in any other AI tool you can type into.
What is the difference between voice dictation and voice prompting?
Voice dictation turns your speech into text. Voice prompting is voice dictation aimed at an AI tool: you speak a prompt into the chat box, read it, and send it. The skill is in how you structure what you say.
How should I structure a spoken prompt?
Say the goal first, then the context, then the constraints, then the format you want back. Saying the ask up front keeps you on track, and constraints are the part people skip when typing.
Is a spoken prompt better than a typed one?
It is usually faster to produce and tends to carry more context, because one more spoken sentence costs only a few seconds. It is not automatically more accurate, so read it before you send it.
Can voice prompting work offline?
The speech-to-text step can. AirTypes Plain runs Whisper on your own CPU with no internet. Cloud AI tools like ChatGPT, Claude and Gemini still need a connection to answer.
Get AirTypes, free
Plain types your prompts free and offline, in any app. Polished is optional prepaid credit that cleans up rambling speech.
Download for macOS and Linux