Dictating into ChatGPT and Claude on a Mac

This is the distinction worth getting right before anything else, because the two are easy to conflate and they solve different problems.

Dictating a typed prompt turns your speech into text that appears in the input box. You can see exactly what the model will receive, fix a misheard word, add a sentence, delete a tangent, and press send only when the prompt says what you meant. The output is text, and text is reviewable.

A chat app’s built-in voice mode is a live spoken conversation. You talk, the model talks back, and the loop runs continuously. There is no prompt text sitting in a box for you to read before the model acts on it. It is excellent for hands-free, exploratory back-and-forth and poor for the moment you want to compose a precise, multi-paragraph instruction and get it exactly right.

A practical wrinkle made the distinction sharper in early 2026: OpenAI retired the built-in Voice feature in the ChatGPT macOS app on 15 January 2026, keeping it on the iOS, Android and web versions. Dictating a typed prompt was never affected by that, because it does not depend on any feature inside ChatGPT. The text is produced by a separate tool and pasted into the box.

Both the ChatGPT desktop app and the Claude desktop app for Mac are text-first: a prompt box at the bottom, a conversation above it. Parakeety pastes at the cursor in any macOS app that accepts text, so the loop is the same wherever you are:

• Click into the prompt box first. The text lands wherever the cursor is, so put it where you want the prompt before you hold the key.
• Hold the section key (§), talk, release. Speak the prompt as a full thought. The transcript appears in the box with auto-punctuation already applied.
• Read it over, then send. Spoken prompts occasionally pick up a word wrong. A two-second read-through catches the obvious cases, and editing is the whole point of dictating into a text box rather than talking to a voice mode.

Because the text goes in through the system clipboard, the same three steps work in a browser tab open to ChatGPT, Claude, Gemini or Perplexity, in a custom GPT, in the API playground, or in any other chat surface that takes typed input.

The prompts that get the best results are the long ones: full context, the constraints that matter, examples of what good looks like, what to avoid. Those are exactly the prompts people do not type, because typing four paragraphs of careful instruction is slow and the effort pushes you toward a clipped one-liner instead.

Spoken prompts tend to be more complete than typed ones. You say the context out loud because saying it is cheap, and the model rewards that context with a better answer. A two-paragraph specification that takes ninety seconds to type can be dictated in fifteen, which changes the calculus: when a thorough prompt costs ten seconds instead of a minute, you write the thorough prompt. The same effect shows up wherever the input is prose rather than syntax, which is why it carries over from chat tools to the editor; the engineer-side version of this argument is in the guides on dictating into Cursor and dictating into VS Code, where Composer and Copilot Chat prompts are the same shape of extended English.

There are two separate network legs, and it is worth being precise about which is which. The audio leg is handled entirely on-device: Parakeety captures your speech to memory, transcribes it on the Apple Neural Engine and discards it. Nothing is uploaded, and the only outbound traffic the app ever makes is a periodic license check and the one-time model download, never your audio.

The text leg is a separate decision that belongs to the chat app, not to Parakeety. ChatGPT and Claude send your prompt text to their own servers because that is how a hosted model works. Parakeety does not change that and does not pretend to. What it changes is that your spoken words do not need a second network path, through a cloud dictation vendor, just to become text in the first place. For prompts that touch client names, internal product names or unreleased work, that is the leg you get to remove.

• Think in full sentences. Dictation rewards speaking the way you would explain something to a colleague, not the way you would type bullet fragments.
• Say the structure out loud. Phrases like “first, second, finally” help the model parse a long prompt, and they read naturally when spoken.
• Type the awkward tokens. Exact identifiers, file paths and code snippets are faster typed; dictate the prose around them. This is the same hybrid pattern engineers use in the editor.
• Edit before sending, not after. The advantage of a typed prompt over a voice conversation is the read-over. Use it.
• Use push-to-talk discipline. The model only hears you while the key is held, which keeps stray room noise and side conversations out of the prompt.

Once dictating prompts becomes the default, it tends to spread to everything adjacent: system prompts you are tuning, the long issue you are about to file, the Slack message explaining a decision, the design doc you are drafting before you hand it to the model. It is all prose, and it all takes the same hold-key-and-talk loop. The audience piece on Parakeety for engineers and developers walks through where this pays off across a working day and how the privacy framing fits NDA-bound work.

Parakeety is a Mac menu-bar app for Apple Silicon. Hold the section key, talk, release; your words paste at the cursor in whichever app you were typing into, ChatGPT and Claude included. Audio never leaves the machine. There is a free 7-day trial with no card required. After that it is $30 once.

Try Parakeety free →