SpeakribeDictation for Claude
Claude rewards context. The more of the situation you give it, the better the answer. That makes prompt length the bottleneck, and prompt length is exactly what dictation fixes.
Why it helps here
You know the background that would make the answer good: the constraints, what you already tried, why the obvious approach is wrong. Typing all of it is a paragraph of work before you get anything back, so most prompts get sent underspecified.
Context becomes cheap
A spoken paragraph takes about fifteen seconds. Once background costs that little, you stop rationing it, and the answers improve accordingly.
Edit before you send
Text lands in the prompt box, so you can restructure or trim it, unlike a spoken conversation, which commits you to your first phrasing.
The transcription step is local
Your prompt is turned into text on your Mac. What you send onward is your decision.
Things people say into Claude
- “Here's the situation, then I'll ask the question. We have a monorepo with two apps sharing a component library, and the build is getting slow.”
- “Review this plan and tell me what I'm missing, especially anything that would only show up at scale.”
- “Rewrite this in a plainer register, keep the structure, and don't lose the caveat in the third paragraph.”
- “I've tried two approaches and both failed for different reasons. Let me describe both, then suggest a third.”
Setting up for Claude
Claude is used from three quite different places on a Mac — the desktop app, a browser tab, and Claude Code in a terminal — and the third one is the reason this page has a setup section at all.
Use a modifier-only binding if you use Claude Code
This is the important one. macOS lets an app block other software from reading the keyboard, and terminals with Secure Keyboard Entry enabled do exactly that. Keyed shortcuts can stop firing there; modifier-only bindings are unaffected. Speakribe does shadow the binding through a fallback when it detects this, but choosing a modifier-only chord means never meeting the problem.
Focus the prompt box before you speak
Same rule in all three places: text lands in the focused field. In the desktop app and the browser that means clicking the composer; in the terminal it means the shell is waiting for input.
Say the whole situation, not the question
The prompt that gets a good answer includes what you already tried and why the obvious approach is wrong. That is a paragraph, which is why it usually gets left out. Speaking it costs about fifteen seconds.
Leave auto-submit off in the terminal
The setting that presses Enter when transcription finishes is convenient in a chat box and unwise in a shell, where Enter runs whatever is on the line. If you enable it for one, be aware it applies everywhere.
One dictation, start to finish
A debugging prompt with the context included, spoken in about twenty-five seconds. The typed version of this is the one that gets cut down to the first sentence.
I've got a race condition I can't pin down. We have a queue worker that pulls jobs and marks them as claimed in the same transaction, and about once a day two workers process the same job. I've already checked that the transaction is actually committing and it is, and I've ruled out the retry logic by turning it off for a day and still seeing it. My suspicion is the isolation level, but before I go and change it, tell me what else would produce exactly this and what I could log to distinguish between them.
I've got a race condition I can't pin down. We have a queue worker that pulls jobs and marks them as claimed in the same transaction, and about once a day two workers process the same job. I've already checked that the transaction is actually committing and it is, and I've ruled out the retry logic by turning it off for a day and still seeing it. My suspicion is the isolation level, but before I go and change it, tell me what else would produce exactly this and what I could log to distinguish between them.
- Four sentences with the ruled-out hypotheses intact. That middle section is the part that disappears when you type, and it is the part that stops the answer restating things you have already tried.
- "isolation level" and "race condition" came through cleanly. Standard technical vocabulary is handled; the words that need the custom-words list are the ones your company invented.
- No paragraph breaks — it is one block of text. For a prompt that is fine, and it is worth knowing that a longer dictation aimed at a document would need breaking up by hand.
What it won't do in Claude
It writes the prompt, nothing else
You cannot start a new conversation, attach a file, switch model or send by voice. Speakribe fills the box; every other action is unchanged.
Terminals can block keyed shortcuts
If you use Claude Code, this is the one to know about. Secure Keyboard Entry stops apps from watching the keyboard, so a keyed binding may go quiet; a modifier-only one keeps working, and there is a fallback for the rest.
No spoken structure
You cannot dictate a numbered list, a code block or a paragraph break. A long spoken prompt arrives as one paragraph, which Claude handles fine and which you would notice more in a document.
It does not tidy your thinking
What you said is what lands, hesitations and all. Some people want a model to reshape a rambling prompt on the way in; that is a real preference, and this is not that tool.
Questions people actually ask
- Does it work in the Claude desktop app and browser?
- Both. Speakribe types into the focused field, so it does not care which.
- Can it keep up with a long, rambling prompt?
- Yes. It is built for continuous natural speech and adds punctuation as it goes, so a long thought comes out as readable paragraphs.
- Does it work in Claude Code in the terminal?
- Yes. Speakribe types wherever the cursor is, terminal included.
- Does it work in Claude Code in the terminal?
- Yes, the same as any focused field. The caveat is Secure Keyboard Entry, which macOS uses to stop apps reading the keyboard and which some terminals enable — a modifier-only shortcut is unaffected by it, and there is a fallback for keyed ones.
- Is my prompt sent anywhere before it reaches Claude?
- No. Transcription happens on your Mac and the text goes into the prompt box. Whatever you then send to Anthropic is sent by you, when you press send.
Ready to stop typing?
One payment. Every future update included.