Dictating to AI Assistants
Talking to an AI assistant is the single most common thing people do with Voibe, and it's easy to see why. A good prompt is long, specific, and full of context you already have in your head. Typing all of that is the bottleneck. Speaking it isn't.
This page covers how to voice-prompt Claude, ChatGPT, and anything else with a chat box — which mode to use, the habit worth unlearning, and what actually happens to your words.
On this page
There's nothing to set up
Voibe types into whatever text field is focused, anywhere on your system. That means AI assistants need no integration, no extension, and no connected account. Click into the chat box, hold your hotkey, speak, and the prompt appears.
This works the same way in every one of them — the desktop apps, the web versions in Chrome or Safari, and the assistants built into other tools. If it has a text box, Voibe types into it.
Not to be confused with Developer Mode
Developer Mode is a separate feature for code editors — Cursor, VS Code, and Windsurf — where Voibe reads your open workspace so spoken file names and identifiers resolve correctly. Chat apps don't need it and don't use it. If you're prompting an agent inside your editor, read that page too.
Which assistants work
All of them, in practice. Because Voibe types into whatever field has your cursor, there's no list of approved apps and nothing to check before you start — if an assistant has a message box, you can talk into it.
That said, people like seeing names. Here are the ones we know are in daily use. If yours isn't here, it almost certainly still works.
Desktop apps
- Claude Desktop
- ChatGPT
Installed assistants with their own window.
In a browser tab
- Claude
- ChatGPT
- Gemini
- Copilot
- Perplexity
- Mistral
- DeepSeek
- Grok
The web version of any assistant — a chat box in Chrome, Safari, Firefox, or Edge is just a text field.
Coding agents in your editor
- Cursor
- VS Code
- Windsurf
- Antigravity
Developer Mode adds workspace awareness in the first three.
Command-line agents
- Claude Code
- Codex CLI
- Gemini CLI
- Aider
CLI agents read an ordinary text prompt, so dictation works the same as anywhere else.
Terminals
- Terminal
- iTerm2
- Warp
- Wave
- Ghostty
- cmux
- Alacritty
The terminal you run those agents in doesn't matter — Voibe types at the prompt in any of them.
Using one we haven't listed?
Tell us and we'll add it. We keep this list from what people actually report using, so it's genuinely useful rather than a scrape of every AI product that exists. Email hi@getvoibe.com with the name — and if it's not working properly, tell us that too, because that's a bug we want.
Which mode for prompts
Prompts are longer than most dictation. That changes which mode is comfortable.
| Mode | Use it for |
|---|---|
| Hands-Free | The default for prompting. Double-tap to start, talk for as long as you need, tap again to stop. Nothing to hold. |
| Push-to-Talk | Short follow-ups — "make it shorter", "now in Python", "try again but formal". Holding a key through a 200-word prompt gets old fast. |
| Live Dictation (Mac) | Long prompts you want to watch build, so you can fix a wrong turn before it lands in the chat box. |
If you haven't set a Hands-Free hotkey yet, it's worth doing before you read further — Dictation Modes covers how.
Stop editing while you speak
This is the habit worth unlearning, and almost everyone arrives with it. People dictate a prompt the way they'd type one — carefully, restarting sentences, trimming as they go. The result is a short, vague prompt that took longer to produce than typing would have.
Two things make the careful version pointless:
- Voibe cleans it up for you. Smart Formatting removes the ums, the false starts, and the "actually, wait" — and adds punctuation and capitalisation. What lands in the box is already tidier than what you said.
- The assistant doesn't need polish, it needs detail. A rambling prompt with all the context beats a crisp prompt missing half of it. Models handle messy phrasing well; they cannot handle information you never gave them.
So say the whole thing. Explain the background, what you've already tried, what you want back, and what to avoid — the way you'd brief a colleague who just walked in. Thirty seconds of talking produces a prompt most people would never type.
A useful habit
End your prompt by saying what you want the shape of the answer to be — "give me three options with trade-offs", "just the code, no explanation", "keep it under 200 words". It's one extra sentence out loud and it saves a round trip.
Reusable prompt scaffolds
Most people reuse the same framing over and over — the same reviewer persona, the same output format, the same list of things to avoid. Retyping it every session is wasted effort, and re-dictating it is only slightly better.
Memory turns any block of text into a single spoken trigger. Say the keyword mid-prompt and the whole block expands inline, then keep talking.
| You say | Voibe types |
|---|---|
| review rules | Be direct and specific. Point out anything that would break in production. Don't compliment the code. |
| house style | British spelling, plain language, no marketing adjectives, short paragraphs. |
| our stack | We use Astro, Tailwind, and TypeScript. No React. Static output only. |
A spoken prompt then sounds like: "Have a look at the checkout flow. review rules. Focus on the error handling."
Technical terms and code
Prompts often carry things that don't survive plain dictation — symbols, file paths, library names, error strings. Two features handle this.
Spoken punctuation
Say symbols by name and Voibe types the symbol — brackets, underscores, slashes, @, currency, and more. See the full list.
Memory for jargon
If Voibe consistently mishears a product name, a library, or a colleague's name, teach it the correction once and it sticks. How corrections work.
For prompting an agent inside Cursor, VS Code, or Windsurf, Developer Mode goes further: it reads your open workspace, so spoken file names and identifiers resolve to the real thing and mentioned files get attached as context. Developer Mode →
Agents in the terminal
Command-line coding agents read from an ordinary text prompt, so Voibe types into them exactly as it does anywhere else — no configuration needed. The terminal makes no difference: macOS Terminal, iTerm2, Warp, Wave, Ghostty, cmux, Alacritty, or whatever you've settled on. Multiplexers are fine too — whether that's tmux, or a tool like cmux that runs several agents side by side in their own panes. Each pane is still just a prompt, so click into the one you want and talk.
Hands-Free is the mode to use here for the same reason as elsewhere: agent instructions tend to be a paragraph, not a line. One thing to watch — in a terminal your prompt is usually sent the moment you press Enter, so read it before you commit rather than after.
Where your words actually go
Worth being precise about this, because it's two separate steps with two different owners.
- 1. Voice to text — that's Voibe. In on-device mode nothing leaves your Mac. In zero-retention cloud mode your audio goes to open-source models, is transcribed, and is deleted the moment transcription completes. Never stored, never sold, never used to train AI. Full privacy overview.
- 2. Text to assistant — that's them. Once the prompt lands in the chat box, it's an ordinary message to that assistant, governed by that provider's terms, exactly as if you'd typed it. Voibe doesn't change what they retain or train on.
In other words: dictating a prompt is no less private than typing the same prompt, and the transcription step adds nothing that outlives it. But it also doesn't make an AI assistant private that wasn't already. If you're dictating something you wouldn't paste into that chat box, the assistant is the thing to check, not the microphone.
For confidential work, our guidance on legal and healthcare use applies to what you put into the assistant, not just how you dictate it.
Questions? Email hi@getvoibe.com and we'll help you out.