Guide

Voice coding: dictating to coding agents

How to speak prompts to Claude Code, Codex and Cursor with blabe: what to say, how to get names right, reusable prompts, safe sending and limits.

Voice coding, as we mean it here, isn't reading code out loud. It's telling a coding agent what you want, the way you'd explain it to a teammate, and letting the agent write the code. Claude Code, Codex and Cursor already turn plain English into edits. The slow part left is getting a good prompt out of your head, and for most people talking is quicker than typing.

This guide covers how we do it with blabe: setup, the shape of a prompt that works, getting project names right, reusing prompts, sending safely, and what blabe won't do.

What you need

  • blabe on macOS 14 or later (Apple silicon or Intel), or on Windows 11 on an x64 PC. On a Mac you can install it from the terminal with brew install --cask telenel/blabe/blabe, or from the download page.
  • An internet connection. Every dictation is transcribed by blabe's service, with xAI or Microsoft's MAI-Transcribe 2 depending on your Transcription setting.
  • The 2-day trial: no sign-up, no card, up to 30 minutes of audio.

On a Mac, blabe asks for the microphone and Accessibility on first launch. Accessibility lets it hear the hotkey in any app and paste for you. Permissions explains each one. On Windows the microphone is the only permission.

The loop

  1. Click into your agent's prompt: Claude Code or Codex in a terminal, the Claude or Codex app, or the terminal inside your editor.
  2. Press your hotkey and say the prompt.
  3. The text lands at the prompt. On a Mac it took 1.0 s at the median from the end of speech to pasted text, across 43 real dictations with AI cleanup on (how we measured).
  4. Read it, fix a word if you need to, and send.

blabe is small enough to leave running all day: on a Mac it's a 4.6 MB download and uses about 75 MB of memory while it waits, in one process (memory measured with blabe 1.0.21).

Say the goal, not the syntax

blabe pastes the words you say. It doesn't turn "open paren" into a bracket or "camel case user ID" into userId, and that's on purpose: the agent is better at writing code than speech rules are. So talk in outcomes and let the agent handle syntax.

A prompt that works well out loud usually has five parts:

  • The goal. "The coupon field clears when you go back to the cart."
  • Where to look. "It's in the checkout package, the cart form component."
  • The change you want. "Keep the code in the URL so it survives navigation."
  • What not to touch. "Don't change the payment step."
  • How to check it. "Add a test that goes back and forth and checks the field."

Spoken, that's about fifteen seconds. Typed, it's the prompt most of us would cut down to "fix coupon field clearing", and the agent's first attempt is worse for it.

For long prompts, say the whole thing first, then fix the two or three words that matter before sending. Editing a pasted paragraph is quicker than typing one.

Get your project's names right

Speech models know common words well. Your project's names are another matter. Three settings help:

  • Words. In Settings → Words, add the names you use: libraries, services, internal tools, teammates. blabe sends your top 50 with each dictation as hints to the model.
  • Fixes. A fix tells blabe what to paste when it hears something else, like "Supabase" when it hears "super base". Fix Last Word in the blabe menu corrects what blabe just pasted, and blabe asks before it learns a fix.
  • Spelled acronyms. Say "E-N-G-I" and blabe pastes "ENGI".

When a reference has to be exact, such as one index.ts among several, dictate the prompt and then type the path or the @ reference yourself. In Claude Code, typing @ opens a file menu. A few seconds of typing saves the agent a guess.

blabe stores fixes and snippets locally and applies them after transcription. It sends your Words as hints with each dictation.

Reuse the prompts you repeat

Most of us send the same few instructions every day: a review checklist, "run the tests and fix what fails", a house style for commit messages. Save each one in Settings → Snippets under a short phrase. Say the phrase on its own and blabe pastes the saved text exactly; say it inside a sentence and blabe swaps it in place.

Sending: when blabe presses Return

By default blabe only pastes, and you press Return. On a Mac there are two other ways:

  • Automatically. Turn on Send automatically to coding agents in Settings → Dictation, and pick which agents get it. blabe presses Return after each paste into one of them.
  • Once. In the default Hold mode, press Return before you let go of the hotkey, and blabe pastes and sends that one prompt.

In a terminal, Return runs whatever is typed, so blabe is strict. It sends only when the tab that got the paste is running a coding agent it recognizes (claude, codex, cursor-agent, gemini, aider and others), the terminal shows the paste arrived, and only one agent could have received it. Otherwise it pastes and stops. It never presses Return at a shell prompt. Coding agents lists the terminals and agents it checks.

On Windows, blabe presses Enter only in the Codex and Claude desktop apps' message boxes, only if you turn sending on, and never in a terminal or an editor.

Our own rule: questions can go straight through, edit requests get one read first. Asking a question doesn't make an agent read-only, so keep an eye on what it does with either.

When your speech is messy

Explaining a bug out loud comes with false starts and the same point said three times. AI cleanup, in Settings → Dictation, sends the transcript to a language model that removes fillers, repeats and false starts and says each point once. It aims to keep every fact, number, name and "not". Pick Aggressive, Standard or Off. Aggressive is the default for new installs. Updating keeps the setting you had. If it fails, blabe pastes the original.

We use Off for short prompts full of file names, where we want exactly what was transcribed, and cleanup for long "here's what's going on" explanations.

Limits

  • blabe needs an internet connection; there's no local model.
  • It doesn't format code, add brackets or turn speech into syntax.
  • On Windows, apps running as administrator get your text on the clipboard instead of a paste. See admin windows.
  • If a paste can't happen, your text stays on the clipboard and the pill says why.

Which speech model does blabe use?

blabe's default sends most dictation in English or with no pinned language to xAI's Grok Voice Transcribe 2.0, and other pinned languages to Microsoft's MAI-Transcribe 2. Add your stack to Words and expect to fix the odd name at first.

Next

  • Coding agents: sending rules, supported terminals and agents, and adding your own.
  • Hotkey: choosing a key, and Hold or press-to-start.
  • Speed: how the latency and footprint numbers were measured.
  • Download blabe and try it on the longest prompt you'd write today.