Productivity
Your prompts are short because typing is slow
Nobody decided that a nine-word prompt was the right amount of context. It is what you write when writing sixty words costs a minute you would rather spend elsewhere.
Last updated
The mechanism
Models work with what you give them. They do not know your situation, your constraints, what you already tried, or what a useless answer would look like. If you do not say it, it is not in the answer.
You know all of it. You simply do not type it, because typing a paragraph of background takes a minute and you would rather see what happens. So you get a generic answer, refine it twice, and spend four minutes getting to where a longer prompt would have started.
That is not a prompting skill problem. It is an input cost problem, and it has an input solution.
The same request, two ways
Typed, eleven words
help me write a job ad for a senior backend engineer
Spoken, fourteen seconds
Help me write a job ad for a senior backend engineer. We are eight people, profitable, no VC, and the role reports to me. The honest pitch is autonomy and no on-call rota; the honest downside is that we cannot match London salaries and the stack is boring on purpose — Postgres and Go, no Kubernetes. I want to filter out people who want a big-company career ladder, because we do not have one, and I would rather lose ten applicants than interview them. No "rockstar", no "fast-paced environment", no list of twenty required technologies. Two hundred words maximum.
Fourteen seconds, and the first response is usable. The typed version needs three rounds of "no, more like this" to get to the same place — and each of those rounds is also typed, and also too short.
The five things worth saying
What a spoken prompt has room for
- What you want. The part everyone types.
- The context. Who it is for, what happened before, what the situation actually is.
- The constraints. Length, tone, format, things that must not appear.
- What you already tried. Stops it suggesting the thing that failed.
- What a bad answer looks like. Astonishingly effective, almost never included, and the single highest-value sentence you can add.
That is about a hundred and twenty words — two minutes typed, twenty seconds spoken. The asymmetry is the whole argument.
Why not voice mode?
Most AI products now have a conversational voice mode. It is a different thing and good for different work.
| Built-in voice mode | Dictating the prompt | |
|---|---|---|
| What you get | A spoken conversation | Editable text before sending |
| Can you fix a misheard name? | No | Yes, before the model sees it |
| Where it works | Products that have the feature | Every prompt box on your Mac |
| Audio goes to | The AI provider | Nowhere, with an on-device app |
| Good for | Thinking aloud, hands-free | Prompts you want to get right |
| Leaves a written record | A transcript, sometimes | The prompt itself |
For anything where the prompt matters, dictation wins on one point above all: you can read it before sending. A misheard file name or client name in a voice conversation sends the model somewhere useless and you find out from the answer.
The privacy point, which is specific here
Prompts contain more than people notice: customer names, internal systems, salary figures, medical questions, unannounced plans, code. The prompt goes to the AI provider — that is what you asked for and you presumably decided it was acceptable.
With a cloud dictation app, the audio also goes to a second company that nobody assessed. On-device recognition removes that hop and leaves only the transfer you intended.
Habits that compound
- Make your standing context a snippet. Your role, your jurisdiction, your house style, the things you never want. Three spoken words, attached to every prompt. Highest-leverage single change.
- Dictate the refinement too. This is where people go shortest — "make it shorter" instead of explaining what was wrong. Speak the whole objection.
- Teach it your vocabulary. A prompt with three wrong proper nouns produces an answer about something else.
- Keep going past the point it feels complete. Spoken prompts feel long at forty words and are usually better at a hundred.
- Type the identifiers. File paths, exact quotations, anything case-sensitive. A misheard identifier produces a confident answer about the wrong thing.
Questions
Do longer prompts really produce better answers?
For anything non-trivial, yes, when the length is relevant context rather than padding. Models work with what you gave them. The reason prompts are short is that typing is slow, not that brevity helps.
Is this better than the AI's own voice mode?
For prompts that matter, yes, because you get editable text and can fix a misheard name before sending. Voice mode is better for hands-free conversation. They are different tools for different moments.
Does it work with Claude, ChatGPT, Gemini and Copilot?
With all of them, and in any browser or desktop app, because a system-wide dictation app types into whatever field your cursor is in. Nothing is provider-specific.
Does dictating send my prompt anywhere extra?
With a cloud dictation app, the audio goes to that vendor before the text reaches the AI provider. With an on-device app it does not — only the prompt reaches the AI.