Developers
Dictate the prompt.Type the code.
The keyboard did not stop mattering. What changed is what goes through it. A developer working with a coding agent now types more English than code in a day, and English is the part a voice does faster.
The short answer
Yes, dictation is worth it for developers, and not for the code. It is for the prose around the code, which is now most of the day: the prompt to a coding agent, the pull request description, the commit message, the issue comment, the Slack reply. Dictate anything longer than one line of English and keep typing identifiers, paths and commands. The function names you say out loud come back as ordinary words, and the agent reading them usually works out which function you meant anyway.
Where a voice wins
Five things you write every day, and which side of the line each one sits on.
Typing is cheap for one line and expensive for a paragraph, so a typed message is short whether or not short was the right length. Speaking is cheap for a paragraph. That one difference decides every row below.
| The text | Typed, usually | Spoken |
|---|---|---|
| A prompt to a coding agent | Three terse lines, because typing a paragraph is work. The agent fills the gaps with guesses, and you spend the next three turns correcting the guesses. | Forty seconds of context: what you tried, what you saw, what must not change. A better prompt at no cost, because saying it is easier than typing it. |
| A pull request description | The title, and "see commits". | Why the change exists, what moved, and where to look first. Said once, in less time than the reviewer spends on the diff. |
| A commit message | "fix stuff", "wip", "address comments". | The sentence you would say to the colleague next to you, which is what a commit message was always supposed to be. |
| A bug report or an issue comment | Shorter than the problem. The steps to reproduce are the part that gets cut. | Steps, expected, actual, while it is still fresh. The part that gets cut is the part worth having. |
| The code itself | Yes. | No. Syntax, identifiers and paths are typed. Dictation tools are prose tools, and this page is honest about that because the tool it sells is one of them. |
The line is not between work and chat. It is between text that has a syntax and text that does not. A prompt has none, so dictate it. A path has one, so type it.
The real failure list
What actually goes wrong when a developer dictates.
Identifiers come back as words
You say getUserById and the transcript reads "get user by ID". Snake case loses its underscores, a file path loses its slashes, and an acronym nobody outside your team has published loses to whichever ordinary word sounds closest. Recognition is prediction, and a name the model has never seen has nothing holding it up.
Symbols are said, not typed
Spoken punctuation works: period, comma, new paragraph. A spoken command flag does not. Say "dash dash force" and you get those three words, because there is no dictation grammar for code and nobody small is going to invent one. Anything with a syntax belongs to the keyboard.
The tool pins one language and your sentence has two
If you work in Portuguese, Spanish or German, your technical sentence is already bilingual: the verbs are yours and the nouns are English. A tool that decides the language per dictation forces the minority half into the majority one, so "fazer deploy" comes back as a Portuguese word that sounds a bit like deploy. The accent was never the problem. The pinning is.
The cleanup pass tidies the wrong things
Anything that turns a raw transcript into clean text is a language model, and a language model left free to rewrite will expand "repo" to "repository", formalise the verb you invented, and turn a quick note to an agent into a memo. Technical prose survives only a cleanup pass that is forbidden from rewording you.
The first two belong to the speech model and to the absence of a code grammar. No small app changes either. The last two are product decisions, and the ones worth choosing a tool on.
The fix
How to dictate to a coding agent.
Five habits, five minutes, on any dictation tool. None of them asks you to speak code. They draw the line in the right place and let the agent do the part a speech model cannot.
- 01
Draw the line at the prompt box
Dictate into the agent's input, the pull request form, the commit message, the issue, the Slack thread. Type in the editor pane. Your hands are on the keyboard either way, so the identifier costs nothing to type and a lot to say.
- 02
Describe the symbol instead of spelling it
Say "the function that loads the user from the session", not "get user by ID". A coding agent maps a description to the right symbol far more reliably than any speech model maps a sound to camel case. The thing dictation is worst at is the thing the agent is best at.
- 03
Say the shape out loud
"New paragraph", then "first, second, third". A good cleanup pass turns a spoken enumeration into a list and a topic shift into a paragraph, and a prompt with structure gets a better answer than the same words in one block.
- 04
Correct yourself out loud and keep going
"The red one, I mean the blue one." "Scratch that." A cleanup pass that resolves spoken corrections keeps only the version you meant. Stopping to delete is the habit that makes dictation feel slow, and it is a typing habit, not a speaking one.
- 05
Fix the three names at the end, then send
Let the identifiers be wrong while you speak. When the transcript lands, the hesitation and the false starts are already gone, so the only edits left are two or three names. Thirty seconds, and the prompt is still three times longer than the one you would have typed.
If the result still reads like a transcript after step five, the tool's cleanup pass is the problem, and that is the one thing in this list you cannot fix with a habit.
What we built
Hold a key, talk to the agent.
Sonira pastes where the cursor is. That is the whole integration: the terminal running Claude Code, the chat panel in Cursor, the pull request box on GitHub, the commit editor, Slack. There is no list of supported apps and no plugin to install, because a paste is the one thing every app already accepts.
The pass that cleans the transcript is written to do a short list of things and then stop. Punctuation. Filler and false starts out. "Scratch that" and "I mean" resolved to the version you meant. "New paragraph" and a spoken enumeration turned into a paragraph and a list. Nothing translated, nothing reworded, nothing formalised, so "repo" stays "repo" and a sentence with two languages in it keeps both.
What it does not do: there is no custom vocabulary yet, so an identifier comes out as words the first time and you fix it by hand. There is no code mode and there will not be one. And read-aloud reads a real text selection, so a terminal pane that draws its own selection, Claude Code inside tmux for instance, gives it nothing to read.
Honest note: the build on this site is the current one. Read-aloud is in it.
Hold a key. Speak. The words land where the cursor is.
The agent is the spell-checker
Dictation's oldest failure met the one reader that does not mind.
Speech to text has always lost on names, and for twenty years the fix was to stop and type them. A coding agent changes the arithmetic. It reads "get user by ID" against the codebase and finds getUserById, because resolving a loose description to a precise symbol is its job. The wrong word stops being an error and becomes a lookup.
What the agent cannot repair is a thin prompt. It can only work with the context you gave it, and typed context is thin because typing it is work. Dictation makes the prompt thick for free. That is the whole case for dictating to an agent, and it has nothing to do with the code.
Review by ear
A long answer from an agent is a wall of plausible prose, and plausible is exactly what the eye skates over. Select it in the browser or the editor and have it read aloud. The wrong step you would have missed while reading tends to sound wrong, which is one of the reasons we built voice out as well as voice in.
Good to know
Questions, answered.
- 01Can I write code by voice?
- Not with a dictation app, and not well with anything else. Dictation is for prose: the prompt, the pull request, the commit message, the issue, the message to a colleague. Code has a syntax and a speech model has no grammar for it. Draw the line there and dictation stops disappointing.
- 02Does it work inside the terminal?
- Yes, if the tool pastes at the cursor rather than typing into a specific app. Sonira pastes, so Terminal, iTerm and the prompt of a coding agent running inside them all take the text. A long dictation arrives as a single paste, which some terminal tools show as a pasted block until you send it.
- 03Will it get my function names right?
- Not the first time. A name the model has never seen comes back as the nearest ordinary words, and no custom vocabulary exists in Sonira yet. Describe the function instead of naming it, and let the agent resolve it, or fix the two or three names after the transcript lands.
- 04I speak Portuguese and my sentences are full of English terms. Does that break it?
- Only on a tool that pins one language per dictation. Sonira pins nothing and leaves each part of the sentence in the language it was spoken in, so "fazer deploy" and "o merge" stay exactly that. There is a separate guide on sentences that carry two languages.
- 05Does it rewrite my prompt on the way through?
- No. The cleanup pass is forbidden in writing from translating, rewording or formalising. It removes hesitation and false starts, resolves spoken corrections, applies spoken formatting, and stops. A prompt reads the way you said it, punctuated.
- 06Why not the free dictation built into the Mac?
- Try it first. It is free, and for a short message it is fine. Raw transcription prints your hesitation, and a forty-second prompt has a lot of it. A cleanup pass that removes the filler without touching the words is the difference between a transcript and a message, and it is the main thing a paid tool adds.
Dictate the prompt. Type the code.
Free while we're in beta.
When the sentence itself holds two languages
English not your first language? Read the accent guide.
New to dictation? Start with the Mac guide.