How to dictate prompts to Claude, ChatGPT and Cursor

Why dictate prompts at all?

What an AI assistant gives back depends mostly on what you tell it. A good prompt says what you are trying to do, what the code or the document looks like now, what must stay as it is, and what a good answer would be. That is a lot of words, and typing them is tiring enough that many prompts end up two lines long.

People talk at roughly 150 words a minute and the average typist manages about 40, so a 200 word prompt takes well under two minutes to say and about five to type. You also tend to include more. When detail costs little, the assistant has the background from the first message and needs fewer follow-up questions.

Short commands are different. “Fix the typo in line 12” is as quick to type as to say. Dictation earns its place on the prompts you find yourself shortening because writing them out is tiresome.

What should you say, and in what order?

A four-part order works for almost any task, and it is easy to follow out loud:

1. The goal, in one sentence. “I want the checkout page to retry a failed payment request twice before it shows an error.”
2. What exists now. “The request lives in the checkout hook, and right now it gives up on the first failure.”
3. The limits. “Keep the public signature the same and do not add a dependency.”
4. What a good answer looks like. “Give me the changed code and one test for the timeout case.”

Speak in whole sentences rather than fragments. Speech recognition does best on ordinary spoken English, and a sentence gives the assistant the same structure a typed paragraph would. If you lose your thread, stop for a moment and say “new paragraph” before the next part.

Check that your tool understands spoken punctuation. Dictate Safe, the dictation app we make, does: “full stop”, “question mark”, “new line” and “new paragraph” all work, and filler words such as “um” and “uh” are removed automatically, with no AI involved.

What does a dictated prompt look like?

This is the shape of it, not a recording. Said at a normal pace, the four parts above become one short, clean prompt:

It takes about 25 seconds to say. Typed at 40 words a minute, the same prompt takes about a minute and a half.

How do you get names, jargon and code right?

Speech recognition is trained mostly on everyday language, so it stumbles on the words an assistant needs: a library name, a teammate’s surname, an internal service with an unusual name.

• Keep a vocabulary list. Add the names, product terms and acronyms you use every week, and add any name that keeps coming out wrong instead of correcting it by hand each time. Dictate Safe uses the list to spell them correctly, and its AI cleanup, which is off by default, also fixes near-misses against it.
• Say symbols by name. In Dictate Safe’s Code Mode, “open paren” types an opening bracket, “equals equals equals” types ===, “arrow” types =>, and “camel case get user name” types getUserName. About 110 spoken forms are covered. Code Mode is off by default, so ordinary prose never turns into symbols.

Which kind of dictation tool should you use?

The most useful way to sort dictation tools is by where the speech recognition happens.

Dictate Safe is in the last group. It runs speech recognition on your own Mac or Windows PC, needs no account to dictate, and is a single payment rather than a subscription. It is English only. For a sense of speed: on a 2018 Intel MacBook Pro, a deliberately old machine, its default speech model runs about 12 times faster than the audio plays, turning almost six minutes of speech into text in about 30 seconds, entirely on the device and with AI cleanup off.

What if your prompts fall under an NDA or a company policy?

Many companies restrict which AI tools staff may use, and some extend that to dictation tools that send audio to a server. Dictating on your own computer removes one question from that conversation: your voice is not uploaded.

The words you dictate are still pasted into whichever assistant you are prompting, and your company’s rules for that assistant still apply. What stays local is the recording and the transcription.

The longer version of this, including how to check what any tool really does, is in Does dictation software send your voice to the cloud?