How to talk to Claude Code: /voice, dictation apps, and tasks it gets right

The short answer

Talking to Claude Code means you say your task out loud instead of typing it, and the words land in Claude Code’s prompt, ready for you to check and send. There are two ways to do it. Claude Code has its own voice dictation: run /voice, then hold Space while you talk. Anthropic lists 20 languages for it, Arabic is not one of them, and it needs a claude.ai sign-in and a microphone on the same computer. The other way is a dictation app that types into any window, including the terminal where Claude Code runs. Whichever you pick, say the task in four parts: the goal, where in the code, the rules, and how to check it is done. That matches Anthropic’s own advice for Claude Code: name the files, state the limits, and give Claude a check it can run, like a test. Talking makes this easier, because you say the details you skip when typing.

Every fact about Claude Code on this page comes from Anthropic’s documentation, read on 19 October 2026 and linked at the end. We make DanDin, one of the dictation apps this page mentions, and we say so where it comes up.

Why give Claude Code tasks by voice?

Because the quality of Claude Code’s work depends on how much you tell it, and people type less than they know. Anthropic’s best practices put it simply: "The more precise your instructions, the fewer corrections you'll need." The same guide asks you to reference specific files, mention constraints, point to patterns in your code, and give Claude a check it can run.

All of that is easy to say and tiring to type. When you type, you cut corners: "fix the login bug". When you talk, you naturally add the story: when it fails, what you already tried, which file you suspect, what "fixed" looks like. That extra context is what Anthropic’s "after" examples look like.

It is also faster for long tasks. One DanDin user’s 605-word brief to an AI tool took 5.6 minutes to say. At a typing speed of 40 words a minute, the same brief would take about 15 minutes.

Voice is not better for everything. Exact symbols, long file paths and code you want copied character for character are still easier to type or to point at with @. A good habit is to talk the task and type the precise bits.

How should you structure a spoken task?

Say it in four parts, in this order. Each one maps to a rule in Anthropic’s best practices.

A four-part spoken task, and the Claude Code advice behind each part (Anthropic best practices, read 19 October 2026).

Part	What you say	Example	Why it helps
1. Goal	What should change, in one sentence	"Lock an account for 15 minutes after three wrong passwords."	Claude knows what done means before it starts
2. Where	The files or folder, and a pattern to copy	"The login code is in src/auth. Follow how the session timeout works."	Anthropic: reference specific files and existing patterns
3. Rules	What must not change, and any limits	"No new libraries. Do not change the database tables."	Anthropic: mention constraints
4. Done when	A check Claude can run itself	"Write a test for the fourth try, then run the whole suite."	Anthropic: give Claude a check it can run

Swipe the table to see every column

Then a few habits that make spoken tasks land:

1. Ask for a plan first on bigger tasks

   Press Shift+Tab until Claude Code shows plan mode, then say the task. Claude reads the code and proposes a plan without changing anything. Anthropic suggests skipping this only when you could describe the change in one sentence.

2. Say names slowly, then check them

   File names, function names and product names are where any speech engine slips. Read them before you send. Claude Code’s own dictation adds your project name and git branch as hints for this reason.

3. Type the exact bits

   After talking, type @ and pick a file, or paste the error message. Mixing voice and typing in one message is normal.

4. Read, then press Enter

   Fixing a wrong word before you send costs seconds. Fixing the wrong change after costs much more.

For a large feature, Anthropic suggests a different start: ask Claude to interview you, then write a spec. Answering questions out loud is where voice shines most.

What does Claude Code offer for voice?

Claude Code has built-in voice dictation, turned on with the /voice command. This is how Anthropic documents it.

• Hold mode (the default). Hold Space to record and let go to stop. There is a short warmup while Claude Code checks that you are holding the key, not tapping it. Your words appear in the prompt as you talk.
• Tap mode. Run /voice tap. Tap Space once to start and again to stop. It sends the prompt by itself when the transcript is at least three words. Recording stops after 15 seconds of silence or two minutes in all.
• Sending. In hold mode, Claude Code waits for you to press Enter, unless you set autoSubmit to true in the voice settings.
• Changing the key. Bind voice:pushToTalk to another key in ~/.claude/keybindings.json. A combination like meta+k skips the warmup.
• Coding words. Anthropic says it is tuned for terms like regex, OAuth, JSON and localhost.
• Where the audio goes. It streams to Anthropic’s servers for transcription. It is not processed on your computer.
• Cost. Transcription does not use Claude messages or tokens, and does not count toward the limits in /usage.
• What it needs. A claude.ai sign-in. It is not available when Claude Code runs on an Anthropic API key, Amazon Bedrock, Google Cloud or Microsoft Foundry. It needs a microphone on the same computer, so it does not work in cloud sessions or over SSH. On WSL, it needs WSLg.

Which languages does Claude Code voice dictation support?

Twenty, set by the same language setting that controls the language of Claude’s replies. If that setting is empty, dictation is English.

Claude Code dictation languages as Anthropic lists them (voice dictation docs, read 19 October 2026).

Listed	Languages
Listed (20)	Czech, Danish, Dutch, English, French, German, Greek, Hindi, Indonesian, Italian, Japanese, Korean, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Turkish, Ukrainian
Not listed	Arabic, Hebrew, Urdu, Persian and every language not named above

Swipe the table to see every column

If your setting is a language that is not on the list, Anthropic says /voice warns you and falls back to English for dictation. Dictation follows that one setting, so there is no way to set it for a sentence that mixes Arabic and English.

Arabic speakers have asked. A request on Claude Code’s GitHub, "Add Arabic language support to /voice dictation", was opened on 3 July 2026 and closed on 26 August 2026 as not planned, after it was marked stale.

Built-in /voice or a dictation app?

Here is the honest comparison. DanDin is our product, so check the facts on its side against our download page too.

Claude Code's built-in dictation next to a system-wide dictation app. Claude Code facts from Anthropic's docs, read 19 October 2026.

	Claude Code /voice	A dictation app (DanDin)
Where it types	The Claude Code prompt, and the Claude Code extension in VS Code	Any app with a text box: terminals, editors, browsers, chat apps
Languages	20 listed, one at a time; no Arabic	Arabic and English, alone or mixed in one dictation
What you get	What you said, tuned for coding words	Cleaned text: no "ums", your corrections applied, punctuation; English from Arabic if you want
Cost	No extra cost; not counted in /usage	Free for 2,000 words a week; Pro is USD 16.99 a month
What you need	A claude.ai sign-in and a local microphone	A DanDin account, the internet, and a Mac or a Windows PC
Limits	Not in cloud or SSH sessions; tap mode stops at 2 minutes	Up to 30 minutes per dictation; no Linux, no phone app
Start key	Space, hold or tap; can be changed	Fn on a Mac, Ctrl and the Windows key on a PC; can be changed

Swipe the table to see every column

Just use /voice if you speak English or another listed language, you mostly dictate into Claude Code, and you sign in with claude.ai. It costs nothing extra, there is nothing to install, and it is tuned for coding words.

A dictation app makes more sense if you speak Arabic or mix two languages, you also want voice in the Claude app, Cursor, Slack or your browser, your Claude Code runs on an API key or a cloud provider, or you want the text cleaned up before Claude sees it.

How does DanDin type into Claude Code on Mac and Windows?

Click into the terminal where Claude Code is running, hold the start key, talk, and let go. DanDin puts the text on the clipboard, presses paste for you, and then puts back whatever you had copied before. It never presses Enter, so you read the task first.

On a Mac, terminals like Terminal and iTerm2 take the paste. We built this path against a real Claude Code session in the Mac Terminal, and found that a busy terminal app can take longer than usual to read a paste. So DanDin waits half a second before it restores your old clipboard, and your words are not swapped out before Claude Code reads them.

On Windows, DanDin pastes with Ctrl+V. Microsoft lists Ctrl+V as a default paste key in Windows Terminal, along with Ctrl+Shift+V and Shift+Insert, so Claude Code in Windows Terminal or PowerShell gets the text. Windows has no check like Secure Keyboard Entry, so never dictate a password into a terminal on a PC.

The one paste rule to know

Claude Code’s docs say that when you paste more than 800 characters or more than three lines, the prompt shows a placeholder such as [Pasted text #1 +12 lines] and still sends the full text. A dictation app delivers its words as a paste, so a long spoken task will show up this way. That is fine.

The same page adds one detail that matters: Claude is told that pasted text can contain instructions you did not write, and to follow instructions inside it only where the message you typed asks it to. So for a long dictated task, also type a few words of your own, like "Do the task below." Then your typed line clearly hands the job to the pasted text.

For a very long brief, Anthropic suggests a file instead: dictate into a file in your editor, then ask Claude to read it. Its docs also warn that the VS Code built-in terminal can drop characters from very large pastes.

A worked example: Arabic in, English out

This is a real request in Saudi Arabic, written as said and run through DanDin’s engine on 11 October 2026, with Translate to English on and Claude as the app in front. It is the same example as on our homepage.

In Saudi Arabic, as said: «أبي صفحة تسجيل دخول فيها إيميل وكلمة مرور، ولو غلط المستخدم ثلاث مرات قفّل الحساب ربع ساعة.»

DanDin’s real output: I want a login page with an email and password. If the user enters the wrong credentials three times, lock the account for fifteen minutes.

Claude Code’s /voice could not take this sentence, because Arabic is not one of its languages. Through a dictation app, Claude Code gets clear English and never sees the Arabic.

The output covers the goal. It does not yet say where, the rules, or how to check. Here is how the speaker might finish the task in a second breath. This part is an example we wrote, not engine output:

Where: the login code is in src/auth, follow how the session timeout works there. Rules: no new libraries, do not change the database tables. Done when: write a test that fails on the fourth wrong password, then run the whole test suite and fix anything that breaks.

Put a typed "Do the task below." in front, read it once, and press Enter. That is a task Claude Code can finish and check on its own.

Getting DanDin to work with Claude Code

Download DanDin for macOS 12 or newer, or for 64-bit Windows 10 and 11. On a Mac, allow the microphone, Accessibility, which lets DanDin paste into other apps, and Input Monitoring, which lets Fn start a dictation. To get English from Arabic, open the dictation options from the gear by the small DanDin bar on your screen and turn on Translate to English. The platform pages cover the rest: DanDin on Mac and DanDin on Windows.

If your tasks are getting long and repeat the same rules, turn those rules into a skill so you stop saying them every time. Our guide to Claude Skills shows how, with a complete example.