Voice prompting on Mac: dictate to Claude, ChatGPT and Cursor without sending your voice to the cloud

The quality of what an AI tool gives you depends on how much context you give it. That’s the problem with typing prompts: typing is slow, so prompts end up short. “Fix the login bug” gets a guess. Two minutes of explaining what you tried, what the logs said and what you suspect gets a fix.

Most people speak at 130 to 160 words per minute and type at around 40. Speaking a prompt is the easiest way to give a model the full picture, and it’s why so many developers have started dictating to Claude, ChatGPT, Cursor and their terminal agents.

This guide covers how to set that up on a Mac, and how to do it without routing everything you say through someone else’s servers.

Why dictation and AI prompts fit so well

• Prompts are prose, not code. You’re describing intent, constraints and context. That’s what speech is good at.
• Rambling is fine. A model handles “so, uh, the thing is the session token is set but then it’s gone after the redirect” perfectly well. You don’t need polished sentences.
• Long prompts stop feeling expensive. When a paragraph takes 20 seconds to say instead of two minutes to type, you stop trimming context out.
• Your hands stay where they were. You can keep the file open, keep the cursor in place, and talk.

The privacy catch with cloud dictation

Most popular AI dictation apps send your audio to a server to be transcribed. When you’re prompting a coding assistant, what you say usually includes internal project names, customer details, unreleased features and sometimes the shape of a security problem.

You may already accept that your prompt goes to the AI provider. A cloud dictation app adds a second company that hears it, and it hears everything you dictate in every other app too.

On-device dictation removes that second hop. The speech is turned into text on your Mac, and the only place your words go is wherever you choose to send them.

Setting up voice prompting with spellsafe

spellsafe is a Mac dictation app that recognises speech on your Mac with NVIDIA’s Parakeet model on the Neural Engine. It has no account and sends no usage data, and after a one-time model download it works with Wi-Fi off.

1. Install it and grant Microphone and Accessibility access. Accessibility is what lets it type into other apps.
2. Click into the prompt box in Claude, ChatGPT, Cursor’s chat, a VS Code extension, or your terminal.
3. Hold Right Option and talk. When you let go, the whole clip is transcribed in one pass and typed. If you prefer to see words appear as you speak, tap Right Option instead and tap again to stop.

A few details that matter for AI chat boxes specifically:

• “New line” doesn’t send your message. spellsafe types new lines as Shift-Return, so chat apps add a line break instead of submitting half a prompt.
• “Press enter” sends it. Finish with “press enter” and it types your text and presses Return, so you can send a prompt without touching the keyboard.
• It works in Electron and Chromium apps. Claude’s desktop app, Slack, VS Code and Chrome only share their text with other apps when asked in a particular way, and spellsafe handles that so spacing and capitals follow what’s already in the box.
• Teach it your project’s words. Add library names, internal service names and colleagues’ names in Vocabulary, so “post gress” lands as “Postgres” and your codename is spelled your way every time.

Cleaning up a messy prompt before you send it

Spoken prompts are often good enough to send as they are. When they’re not, spellsafe has two ways to tidy them, both using a language model (Qwen3-4B) that runs on your Mac:

• Clean up on release. Turn on “Clean up with AI on release” and hold-to-talk dictation has fillers and self-corrections removed before it’s typed. “Use the, no, the staging database” becomes “Use the staging database.”
• Command mode. Select your draft, hold Right Command and say what to change: “turn this into a numbered list”, “make this shorter”, “fix the typos”. The text is rewritten in place, and ⌘Z undoes it.

Because the cleanup model is told to edit what you said and never to answer it, a dictated question stays a question. If you dictate “what’s the fastest way to dedupe this array”, that’s what gets typed into the chat, not an answer.

What it’s not

It’s worth being clear about the limits:

• It’s for prose, not syntax. spellsafe types what you say. It isn’t designed for dictating code character by character (“open paren, i, colon”). Use it to tell the AI what to write, and let the AI write the code.
• English only, for now.
• Apple silicon Macs on macOS 14 or later. The models need the Neural Engine and a few gigabytes of disk: about 450 MB for speech recognition, and about 2.3 GB more for the cleanup model.

A simple voice-first workflow

1. Open the file or error you’re working on.
2. Click into your assistant’s prompt box and hold Right Option.
3. Explain the problem out loud as you’d explain it to a colleague: what you expected, what happened, what you’ve ruled out.
4. Let go, glance at the text, say “press enter” or hit Return.
5. When the reply comes back, talk through your follow-up the same way.

Most people find their prompts get three or four times longer within a week, and the answers get noticeably better.

Try it

spellsafe has a 7-day free trial with every feature, then costs $9 a year. Download it for your Mac.