Talk instead of type
Dictating gives far more context than typing – and gets better answers. The difference between dictation and voice mode, and where your voice goes along the way.
Boaz LichtensteinFacts checked on 10 November 2026
Most prompts are too short – and the reason is the keyboard. People who type save words. People who speak tell the story: the occasion, the background, the exception. Exactly the details a model otherwise lacks.
Two tools that look the same
Behind the microphone icon there are two different functions, and both OpenAI and Anthropic separate them explicitly.
Dictation converts speech into text. The text lands in the input field, you read it, correct it and send it. The model receives a written brief.
Voice mode is a conversation in both directions: you talk, the assistant answers aloud, you interrupt. OpenAI points out that the transcript of such a conversation is not a verbatim record. ChatGPT Voice · Voice mode in Claude
How to dictate a good brief
Speak in three sections with a short pause in between:
- What it is about and what you need the result for.
- What you have: documents, figures, earlier attempts.
- What the result should look like and what may be missing.
Speak for at least a minute. You do not need to announce punctuation. Read the text through once afterwards and correct names, figures and technical terms – that is where recognition errs most often, and where a mistake is most expensive.
Agents can be spoken to as well
Dictation is not limited to chat. Claude Code, for example, offers push-to-talk recording in the terminal: hold a key, speak, release. For long briefs to an agent that is the fastest route, because you can tell starting point, goal and limits in one go. Voice dictation in Claude Code
Where your voice goes
The conversion usually does not happen on your device. For Claude Code the documentation states it outright: the recording is streamed to Anthropic’s servers for transcription and is not processed locally. The same holds in principle for other assistants’ voice features unless the provider promises otherwise.
For confidential material – staff conversations, customer data, anything under a duty of confidentiality – there are two ways out. The operating system’s dictation works locally, depending on device and settings. And tools for local models ship their own transcription: LM Studio Bionic can, according to the vendor, convert speech on the device. What exactly stays local there is covered in LM Studio Bionic.
And regardless of the technology: anyone dictating a brief with customer names in an open-plan office has the privacy problem before the microphone even comes into it.