Talk to Claude Code instead of typing.

Why prompts are the worst thing to type

The difference between a prompt that works and one that wastes a minute is usually context: which file, which constraint, what you already tried, what the failure looked like. You know all of it. Writing it down is pure transcription, and it is the kind of transcription that gets cut short because typing is slow. So you send three words instead of three sentences, get back something generic, and type again.

Said out loud, the same context takes about fifteen seconds. That is the whole argument for dictating into Claude Code, and it does not need Claude Code to add anything.

How it works

Claude Code runs in your terminal. Zumbo types into whatever has the cursor, so the terminal is just another place with a cursor. There is no extension, no integration and nothing to configure per tool.

your shortcut

Listening… 0:04

Hold it to talk, tap it to toggle, Escape cancels. Pick the key in Settings.

1. Put the cursor in the Claude Code prompt.
2. Hold your shortcut. You choose which key that is in Settings during setup, and you can change it whenever you like.
3. Talk. Say the whole thing, including the parts you would normally abbreviate.
4. Let go. The text lands at the prompt in about a fifth of a second, punctuated.
5. Read it, then press return.

Because there is no session limit, a long prompt is one press. You can pause mid sentence to look at a stack trace and carry on.

The part that decides whether you keep doing it

A prompt full of misspelled package names is worse than no prompt, because the model will cheerfully reason about the wrong thing. This is where a general speech model falls down: it spells what it thinks it heard, and what it thinks it heard is ordinary English.

Turn on the developer tools vocabulary pack during setup and the everyday names come out right the first time: the runtimes, the build tools, the package managers, the common libraries. For anything it still gets wrong, including your own internal service names, select the word, type what you meant, press Teach. It is heard your way from then on and the correction stays on your Mac.

What it is good for beyond prompts

Once the shortcut is in your hand you stop reserving it for one thing. Commit messages get written properly instead of being a verb and a noun. Pull request descriptions get the reasoning in them. The comment explaining the workaround actually explains it. All of those are in the same terminal or the same editor, and all of them are things you have already thought through and are only slow at recording.

Privacy, since this is a terminal

Zumbo has no server for your voice. There is no account, no API key and no network request needed to transcribe: the speech engine ships inside the app and runs on your Mac. Whatever you then choose to send to a model is your decision, taken with the text in front of you, rather than something that happened to your audio on the way.

Related: dictation for developers, dictating into Cursor, and dictation with the Wi-Fi off.