Typing makes you cut things out, often the bits AI needed
Type the same job into Claude or ChatGPT and you write a fraction of it. Not because you forgot the rest, but because you edit yourself as you go, and three things go first.
- What you're trying to achieveThe objective, and why you're asking at all. Easy to leave out when you're typing, even though it can change the answer substantially.
- The background around itWhat came before, who it's for, what's happened already. Feels like preamble when you're the one typing it.
- The bits you notice as you goEdge cases, exceptions, the "oh, and watch out for". Saying something out loud makes you think about it, and you catch things you'd never have typed.
Speaking reduces that pressure. It’s much easier to add another sentence of context out loud than to type another paragraph.
Those are often the missing pieces. Giving the model your objective, background and exceptions gives it much more to work with.
It’s the microphone in the prompt box
On Claude Cowork and ChatGPT Work it’s the microphone icon, sitting in or beside the box where you type. Click it, talk, and what you said appears as text where you’d have typed it. You may need to allow microphone access the first time.
On ChatGPT Work there’s a second button next to it, a little waveform. That one starts a spoken conversation where it talks back to you, which is a different thing and not what this page is about. You want the plain microphone.
Ramble at it, and don’t tidy it up before you send
Talk for as long as you’ve got something to say. Two minutes, five. Stop, change your mind, go back on yourself.
Don’t feel you need to turn it into a neat written prompt. Fix anything materially wrong, then send the rest as it came out.
You usually don’t need to organise it nearly as much as you might think. Current models are good at pulling the useful parts out of fairly messy speech. Andrej Karpathy, who co-founded OpenAI, says he works this way on purpose: leaning back and rambling for ten minutes, “total mess, anything goes, full stream of consciousness”.
The easiest time to start is the next time an answer comes back wrong. That’s the moment you’ve got a lot to say and you’re least inclined to type it.
Dictation will sometimes mishear names. If Claude or ChatGPT already has relevant reference material, it can often infer what you meant, but check names, figures and other details that matter.
Find a meeting room, or get used to it
The thing that actually stops people is feeling daft doing it with colleagues in earshot. It’s a real objection and not a silly one.
The practical answer is a meeting booth, a spare room, or doing it at home.
It can feel awkward at first, but in many offices it sounds much like being on a call. Use the same judgement you would on a call about confidential or sensitive information.
You don’t need a separate dictation tool
There are dedicated ones, and people will tell you they’re better.
For simply getting a long prompt into Claude or ChatGPT, the built-in dictation is now good enough that most people should start there.
They do still do one thing the built-in microphone can’t, which is dictate into anything on your laptop: an email, a document, a spreadsheet, not just the prompt box. If you want that, it’s worth paying for. If you only ever wanted to talk to AI, use the microphone that’s already there.