How do you dictate a good prompt for ChatGPT or Claude?

Say the goal first, then the context, the constraints and the output you want, and read the transcript before you send. That order keeps the main request intact even when the recognizer gets a detail wrong. Speaking is also quick. In a controlled study on an iPhone, English speech input reached 153 words per minute, against 52 on the built-in keyboard ¹. At those rates, a 1,000-word draft takes about 6.5 minutes to speak, against 19.2 minutes to type ². The time you save is best spent on the detail a model needs.

Say the task first and the constraints last

Research on long inputs suggests where the request should sit. Language model accuracy was often highest when the relevant information came at the start or the end of the input, and dropped significantly when it sat in the middle ³. A spoken prompt that wanders for a minute before it names the task puts the request in the weakest place.

Use this order:

1. Goal: the task in one sentence, such as "I need to reply to my landlord about the boiler."
2. Context: the material and who will read the result, such as the email thread or the team.
3. Constraints: what must stay the same and what to avoid, such as "keep the two dates from her email" or "no threats."
4. Output: the format and length, such as "a subject line and an email under 150 words."

Anthropic's prompting guidance for Claude makes the same point. It asks you to think of the model as a brilliant but new employee who lacks context, and says the more precisely you explain what you want, the better the result ⁴. Each part above answers a question that new employee would otherwise have to ask.

Spoken, that prompt might sound like this: "I need to reply to my landlord about the boiler, which has been off since Monday. Keep the two dates from her email, the twelfth and the nineteenth. Firm but polite, no threats. Give me a subject line and an email under 150 words." Each sentence does one job, so a misheard word in one of them does not break the others. The dates are the words to read back, because a small error there changes the meaning.

Use the chat microphone for short prompts and a text editor for long ones

For a sentence or two, the microphone in the chat box is the simplest choice, because you can edit the text before you send it. Our overview of dictation in ChatGPT and Claude covers the controls and their limits.

For a long prompt, especially one with pasted material, dictate into a text editor or notes app, read the whole draft in one place, and paste it into the chat. Yap, which we make, is an on-device dictation app for Mac and Windows, and you can download it to try it. Built-in tools are enough for many people, so start with those.

Keep each burst short. Apple's guide for macOS Catalina 10.15 advised speaking for no more than 40 seconds at a time ⁵. That guidance is for an older version of macOS, so check the limit in yours. Short bursts also give you natural points to check the transcript as you go.

Read the transcript and edit it before you send

Editing matters. In a 2026 preprint, students who typed their prompts were more likely than students who submitted unedited voice prompts to succeed on the first attempt, on two of three problems ⁶. When students edited their transcribed voice prompts before submitting, the study found no difference in success rate ⁷. The study used introductory programming tasks, so it may not carry over to every kind of prompt, but the direction is clear: read and fix the transcript first.

Check names, numbers and code by hand

Names, numbers and code are where speech input is weakest. A 2026 study of spoken questions about code named custom identifiers, such as variable and function names, as a central difficulty ⁸. In practice:

• Spell an unusual name once in the prompt, so the model has it in writing.
• Say symbols by name, such as "underscore" or "open bracket", or type them yourself.
• Paste exact code, file paths and error messages instead of dictating them.
• Check each number against your source, not against the transcript.

If an app keeps misspelling a name, see how to make dictation spell names correctly. For speaking punctuation and symbols, see dictation commands.

It depends on what the prompt contains

Speaking is fast enough that you can afford to add the context a model needs. The trade-off is that a spoken draft needs one careful read, and a prompt sent without that read can carry a misheard detail straight into the answer.

For whether a few dictation errors still get through, see do AI chatbots understand dictation errors. For how much detail is too much, see do longer prompts get better AI answers. The wider picture is in our guide to speech to text for AI and our overview of speech to text and AI.