Dictating Prompts: The Fastest Way to Talk to ChatGPT and Claude

The prompt has quietly become the longest thing most knowledge workers write all day. Not the email — the prompt. A good one carries the goal, the constraints, the examples of what you don't want, the stack you're on, and the two edge cases you already know about. Three hundred words, easily.

And almost everyone types it. Which is strange, because a prompt is the ideal candidate for voice: it is long, it is conversational, it is disposable, and nobody is grading your grammar. You are talking to a machine that will read it once and throw it away.

This is about why dictating prompts makes AI tools work better — not just faster — and how to set it up on a Mac.

Typing is now the slow part of AI work

The model answers in seconds. You spend a minute and a half composing the question. On a day with thirty or forty prompts, the bottleneck is not inference — it is your hands.

Comfortable typing runs around 40–70 words per minute. Conversational speech runs around 130–160. On a 250-word prompt, that is roughly four minutes of typing against ninety seconds of talking. Do it thirty times a day and the arithmetic gets hard to ignore. We broke that maths down in AI voice to text is 4x faster than typing.

But throughput is the boring half of the argument.

The real gain: you give the model more context

Watch what typing does to a prompt. Because every word costs keystrokes, people compress:

“write a function to validate emails”

Now watch what the same person says out loud when the cost of a sentence is two seconds:

“I need an email validation function in TypeScript for a signup form. Don't use a giant regex — I want it readable and testable. It should reject disposable domains from a list I pass in, and it needs to return a reason for the failure rather than just false, because the UI shows that message to the user. We're on Node 22 and I'd rather have no dependencies.”

Same intent. The second version gets an answer you can ship, because it contains the constraints that were always in the author's head and never made it to the keyboard. Under-specified prompts are largely an input-cost problem, and dictation removes the cost.

Where dictation fits in an AI workflow

Some prompt types benefit far more than others. In practice:

The spoken context dump

You are debugging something and the model has none of the background. Rather than typing a tidy summary, talk through it the way you would to a colleague: what you expected, what happened, what you already tried, which two files are involved. It comes out messy and it works — models are excellent at extracting signal from a rambling but complete account, and terrible at inventing context you left out.

Iterating on a long answer

Refinement rounds are short, frequent, and conversational: “keep the structure but make the second paragraph less formal, and cut the last bullet entirely”. Typing those is pure friction. Speaking them keeps you in the loop instead of breaking flow to compose an instruction.

Driving a coding agent

Terminal-based agents take instructions in prose, which means they take dictation natively. “Add a regression test for the empty-array case, run the suite, and if it passes commit with a conventional message” is faster said than typed — and any dictation tool that inserts text at the system level works in a terminal window, not just in a browser.

Writing the spec before the code

The most valuable AI prompts read like small specs. Specs are exactly what people avoid writing because writing them is tedious. Talking one out takes two minutes.

Two things that make dictated prompts worse — and their fixes

1. Self-corrections that read as instructions

Real speech contains reversals. “Do it in Python — actually no, TypeScript.” Submit that raw and you have handed the model two contradictory requirements; sometimes it picks the wrong one, sometimes it hedges and gives you both. This is a real failure mode of naive dictation, and it is invisible until the answer comes back odd.

The fix is a cleanup pass between your voice and the prompt box. Air Wisper resolves self-corrections before the text is typed, so the model receives “Do it in TypeScript” — the thing you actually meant.

2. Homophones and technical vocabulary

Raw speech recognition works phonetically, so it cheerfully produces “a sink function” instead of “an async function”. In prose you would catch it; in a 300-word prompt you often won't, and the model gamely answers the wrong question. Sentence-level cleanup catches most of these because it can use surrounding context that a phonetic match cannot.

The privacy question you should actually ask

People dictate genuinely sensitive things into AI tools: production code, patient scenarios, deal terms, unreleased strategy. There are two separate questions, and they get conflated constantly.

• Where is your voice transcribed? Many dictation apps upload the audio to their servers. Air Wisper doesn't — recognition runs on-device via Apple's speech framework, so the recording never leaves your Mac. Only the resulting text is sent for the polish pass, and it isn't stored.
• Where does the prompt go? Wherever you submit it. Dictation cannot change your AI provider's terms. If the content is covered by an NDA or a data agreement, that constraint applies to the prompt regardless of how the words got typed.

Worth separating, because an app that keeps a permanent cloud copy of your voice adds a second exposure on top of the one you already accepted. More on that in Why your voice data is more sensitive than you think.

Setting it up on a Mac

1. Use a system-level dictation tool, not a per-product microphone button. You want one shortcut that works in the ChatGPT web app, Claude, Cursor, a terminal, and your notes app — not five different affordances with five different behaviours.
2. Bind it somewhere your hand already rests. Air Wisper defaults to ⌃⇧Space. If it takes thought to trigger, you will type instead.
3. Speak the whole prompt in one go. Resist the urge to stop and check after each sentence — cleanup works better with a complete thought, and stopping is what breaks flow.
4. Read before you submit. Not to fix grammar — to confirm the intent survived. This takes three seconds and catches the rare mangled technical term.
5. Keep the keyboard for edits. Voice is for producing text, hands are for rearranging it.

What changes after a week

The habit shift is not “I type less”. It is that you stop pre-editing your own thinking. When a prompt costs four minutes of typing, you unconsciously ask smaller questions — ones you can specify cheaply. When it costs ninety seconds of talking, you ask the real question, with the real constraints, and you get a first answer that is close enough to use.

That is the compounding effect: not words per minute, but the size of question you are willing to ask.

Frequently asked questions

Air Wisper is a native macOS app and requires macOS 14 or later. Free tier includes 200 AI polish requests per week — no credit card.