How to give tasks to an AI coding agent by voice

Why tasks come out better by voice

When typing, people save words: “fix the form”. When speaking, they explain like to a colleague: what broke, where, what was tried, how to check. The agent needs exactly that. Voice nudges you toward a complete brief.

A few habits that help:

• start with the goal, then the details: “the email should go out after payment, it does not, here is what I saw…”;
• say file and command names the way they are written;
• end with what counts as done;
• do not worry about slips: the agent gets the meaning, and you see the text before sending and can fix it.

Dictate straight into the session field

In Hedoq dictation is built into every session’s input field:

1. Press ⌘. or the mic in the field.
2. Speak. While you talk, the field shows a “speak” prompt and three dots.
3. Press ⌘. again or Enter. Hedoq transcribes the whole recording at once.
4. The text lands at the cursor, not sent. Read it, fix it, send it.

Esc during recording cancels it entirely. If voice input is not installed yet, the mic button offers to install it with one button.

Recognition runs on your Mac with a Whisper model, and the audio is never uploaded. You do not set a language in advance: mixed speech like “run npm run build and look at the errors” in another language is handled.

Teach recognition your words

The weak spot of any dictation: library names, service names, in-house terms. A misheard “Kubernetes” breaks a code search.

Hedoq has “Dictation vocabulary” for this, in the settings under Voice. List the words you say often, separated by commas: project names, libraries, services, colleagues’ names. Up to 600 characters. The vocabulary tells the recognition model how these words are spelled.

What to add first:

• the names of your projects and services;
• libraries and tools with unusual spelling;
• foreign terms you pronounce your own way.

Dictate beyond Hedoq

A task for an agent often starts somewhere else: in a tracker ticket, an email, a note. Turn on “Dictation in every app” in the settings under Voice.

Then ⌘. works in any app while the Hedoq window is not in focus. The same engine on your Mac records and transcribes, and the text is pasted where the cursor is. Pasting needs macOS Accessibility access. Without it the text stays on the clipboard, and Hedoq tells you so.

Hear answers without looking at the screen

Voice is handy not only for tasks but for answers: while you walk around the room, sketch on a whiteboard or hold something else.

In Hedoq every Claude or Codex session has a voice mode:

• turn it on with the speaker button in the field or ⌘/;
• when the agent ends its turn, the result is read aloud, with a chime before and after;
• the first message in this mode asks the agent to answer briefly: a long answer does not work by ear;
• to stop it: Esc, a new recording or sending a message;
• answers from several sessions are read one after another.

Pick the voice and speed in the settings, “Answer voice”. There are four Russian voices and English voices too; the language is detected from the text. Speed goes from 100 to 150 %, 130 % by default. Speech also runs on the Mac, offline.

You see the mode is on at a glance: the field outline shimmers and the session name in the sidebar is tinted.

Voice from the phone

Away from the computer, the Hedoq app for iPhone also has a mic and read-aloud in the input field. There dictation goes through iOS speech recognition, and the task goes to the session on your Mac. How to run agents from the phone: a separate guide.

A habit for the week: dictate every task longer than one line. Your briefs get fuller, and the agent asks fewer follow-up questions.