Dictating Claude prompts: say the whole situation

Last updated 11 September 2026

Why length matters more here

All models do better with more context, but the effect is especially visible with a model that handles long prompts comfortably. There is no penalty for explaining the whole situation, and a large penalty for explaining a third of it.

The problem is purely mechanical: typing two hundred words of background is a chore, so people type twenty. Speaking two hundred words takes ninety seconds and feels like explaining something to a colleague — which is exactly what it is.

The shape of a good spoken prompt

1. Who you are and what the situation is "I run a two-person consultancy. A client has asked for something outside the scope we agreed."
2. What you want The actual request, stated once and plainly.
3. What you have already tried or considered Stops it proposing the option you rejected for good reasons.
4. The constraints Length, tone, format, jurisdiction, things that must not appear.
5. What a bad answer looks like The most under-used sentence in prompting. "Don't give me a list of generic options; I want you to pick one and defend it."

That is around a hundred and fifty words. Typed, it is a three-minute job most people skip. Spoken, it is under a minute and you would have said all of it to a colleague anyway.

Standing context as a snippet

Most people have a paragraph they should attach to every prompt and never do: their role, their jurisdiction, their house style, the things they never want in an answer.

Put it in a spoken snippet. Say three words at the start of a prompt and the whole paragraph appears. Then dictate the specific request after it. See voice snippets.

The privacy point, which is sharper for Claude users

Claude is used heavily for work that carries real confidentiality: legal drafting, medical questions, security analysis, commercial strategy, code. Prompts in those domains contain names, figures and systems you would not put in a public document.

The prompt goes to Anthropic, because that is what you chose. With a cloud dictation app, the audio also goes to a transcription vendor — a second organisation, probably not in anyone's assessment. On-device recognition removes that second transfer entirely.

Where dictation fits in a Claude session

• The opening prompt. Highest value by a distance. Everything downstream depends on it.
• The refinement. Where people go shortest — "make it shorter" rather than explaining what was wrong with it. Speak the whole objection instead.
• Pasting a document and explaining it. Paste by keyboard, dictate the framing around it.
• Project instructions. Long prose documents that set standing behaviour. Ideal dictation material.
• Not the code or the data. Paste those. Speech is for the explanation around them.

Habits that compound

After a fortnight, most people who dictate prompts report the same two changes. They ask for things they would not have bothered typing, and their prompts contain the reasoning rather than just the request — which means the answers argue with them rather than agreeing.

Neither is a feature of any AI product. Both come from removing the cost of explaining yourself.