The tip for working faster with any AI is to give it more context
Whether you use ChatGPT, Claude or Gemini, the core of using AI well is the same: how deep the answer goes depends on how much context you give it.
The AI starts out knowing nothing about your situation. If what you send is short, it can only fill the gaps with general advice. If you also give it the following, the answer moves closer to your actual work.
- Purpose: what you're trying to achieve
- Background: how things got here, who's involved, what's already decided
- Constraints: budget, deadline, things that can't change
- Doubts: where you're stuck, and what you're unsure about
People who move fast with AI care less about which model they pick and more about how much material they hand over, and what's in it. How to structure a request is covered in “Voice prompts for ChatGPT and Claude”.
Read nextVoice prompts for ChatGPT and Claude: how to structure themDictate prompts for ChatGPT and Claude by stating the task, context and constraints. Typing all of that is tedious; speaking it is quicker. Here's how to structure spoken prompts, and how to check that your dictation tool doesn't change your instructions.
The biggest bottleneck is getting your own ideas and first-hand information into the AI
AI knows a great deal. But there are things it cannot know:
- What your client actually said in the meeting
- The mood on your team, and what's going on inside your company
- The feel of the ground that hasn't shown up in the numbers yet
- Your hunch that "this option just seems better"
This first-hand information, and your own thinking, exists only in your head. However capable AI becomes, this is the part a person has to hand over.
So when you work with AI, the bottleneck is usually not on the AI's side. It's moving what's in your head into the AI. If that channel is narrow, even the best AI will lean toward generic answers.
When typing, you leave out background, doubts and hunches without noticing
Yet when you type, this is exactly the part that gets cut. We, the team that makes SayonaraType, did it too when we typed:
- Squeezing the background into one line
- Not writing things that only "feel a bit off"
- Skipping an extra point because going back to type it in was a hassle
What gets cut is precisely what the AI doesn't know. Give it only the main points, and the AI either asks you to confirm its assumptions or answers in general terms. Then you type again. That back-and-forth is what slows down working with AI.
Speaking passes on the parts you'd skip when typing; the difference is quantity, not quality
The way we recommend to clear this bottleneck is dictation. Speaking takes little effort, so the parts you'd skip when typing reach the AI as well.
Our team, who use dictation in Japanese, spend most of their dictation time brainstorming with AI, typically speaking for one to three minutes and sending it all in one go. In our experience, the quality of our thinking is no different between typing and speaking. What changes is the amount:
- Typing: you send the main points → the AI checks its assumptions → you type again
- Speaking: you send the main points, your doubts and passing thoughts in one go → the AI answers with all of that in mind
With more material, each exchange goes further. As a result, you reach the same depth sooner, and your ideas develop much faster.
When ideas develop quickly, brainstorming itself gets fun, and it's easy to keep going
The time spent on a conversation may not be much different from when you typed. What changes is how far you get and how much you exchange in that time.
Speak anything long; type only one-line replies
We don't do everything by voice, either.
| Situation | How we handle it |
|---|---|
| Short replies of about one line ("That works," "Go with option 2") | Sometimes typed |
| Longer messages that explain background or thinking | Almost always spoken |
Short replies can be quicker to type. The longer the message, the easier speaking is, and the more you pass on. A simple rule, "if it's going to be long, say it," saves you deciding every time.
Dictated text works best for AI when the ums are cleaned up but your hesitation is kept
If spoken words are turned into text exactly as said, filler words and restarts like "um" and "no wait, I mean" stay in. They're hard to read and add noise for the AI too. A dictation tool is more useful if it can tidy things up to a degree.
But tidying can go too far. What you want to give the AI is not only what you said, but also where you're confident and where you're unsure.
| Example text (written for this article) | |
|---|---|
| As spoken | So, um, I feel like a monthly plan might be easier to, uh, keep going, but oh, actually it could be a burden for users |
| Tidied into firm statements | A monthly plan is easier to sustain. However, it is a burden for users. |
| Tidied with the nuance kept | I feel like a monthly plan might be easier to sustain, but it could be a burden for users. |
When "I feel like" becomes a plain statement and "could be" becomes "is," the AI takes them as settled facts. You wanted help thinking through your doubt, and the doubt itself has disappeared.
Leave what you meant to be uncertain, uncertain. Then the AI answers knowing where you're unsure.
SayonaraType tidies your speech while keeping subtle nuance
SayonaraType is a dictation tool built on this idea: a Mac dictation tool that turns what you say into text ready to send.
It tidies filler words, restarts, punctuation and line breaks, number formatting and how sentences connect, and is built to keep your meaning and nuance as you said them. It doesn't add things you didn't say. You can speak English and 12 other languages besides Japanese, but how well it formats English is still being researched.
We built it on the idea that, especially when talking to AI, the nuance in what you say shouldn't be cut away
For long requests to AI, email and chat: for people who write a lot every day. It's free to try (the paid plan is currently available only in Japan). It runs on Apple silicon Macs (M1 or later) with macOS 14 or later.
If you want to compare it with other dictation tools, see “Dictation apps compared”.
Read nextDictation apps: 13 compared and ranked (November 2026)You want what you say to come out as text you can send. If that's why you're choosing a dictation app, four things matter most. We checked those four points for 13 apps on Mac and Windows, using only what each company's official pages say. This article does not compare how well they hear you or how well they clean up your speech.
To start, talk to a paid-plan AI before your thoughts are organized
For people just starting to work with AI by voice, we suggest three things.
- 1Talk to an AI on a paid planClaude's Opus, ChatGPT's paid models, Gemini's paid plan, etc.
- 2Start talking before your thoughts are organizedLet the AI do the sorting; say things as they come
- 3Say the things you wouldn't tell a personUnease, gut feelings, stray ideas
Talk to an AI on a paid plan
We recommend models available on paid plans, for example Claude's Opus, ChatGPT's paid models or Gemini's paid plan. You'll be handing over long, loosely organized talk, so a partner that can take it in and sort it out makes the brainstorming go further. It doesn't have to be about work; any topic is fine.
Start talking before your thoughts are organized
The trick is not to organize your thoughts before you speak. If you try to organize first, you end up leaving things out, just as when typing.
For example, you might start like this (an example written for this article):
I want to talk through the new service we're launching next month. My thinking isn't clear yet, but I'm torn between a monthly subscription and a one-time purchase. A subscription feels easier to sustain, but something tells me users might find it a burden. I can't quite put my finger on what's bothering me.
Even if it isn't organized, the AI separates out the issues in its reply. Keep talking and adding to that, and you're brainstorming.
Say the things you wouldn't tell a person
Precisely because it's an AI, try saying things you'd hesitate to tell a person:
- A sense that something is off, even if you can't explain why
- A feeling deep down that you can't quite name
- An idea you're not sure holds together
Because it's an AI, you can share half-formed feelings as they are. We find that this sometimes leads to unexpected insights, or a depth we couldn't reach on our own.
Also keep these in mind:
- Read it once before sending: check that names and numbers weren't misheard
- For confidential content, check data handling in both the dictation tool and the AI: your words pass through both. SayonaraType doesn't store your audio or text on our servers, and processes them with settings that keep them out of AI training
- Say things you wouldn't want overheard only where no one is around
If you're unsure, try sending one long request by voice
- AI answers feel shallowGive more context: purpose, background, constraints and doubts
- Typing leaves only the main pointsDictate, including background, doubts and passing thoughts
- Your thoughts aren't organized yetDon't organize; start talking as things come to mind
- Unsure whether to speak short replies tooType one-liners; speak anything long
- Your hesitation disappears from the textChoose a tool that tidies the ums but keeps "I feel like"
If you'd like to try without spending anything first, you can start with the dictation built into your Mac. The steps are in “How to use dictation on Mac”.
Read nextHow to use dictation on Mac: setup, punctuation and limitsYour Mac already has dictation built in, at no extra cost. Here's how to turn it on, how punctuation works, and how to switch languages, following Apple's official help. We also cover where the built-in feature falls short.
ChatGPT, Claude or Gemini: start by handing the AI everything only you know.
SayonaraType
A voice input app for Mac. Free to try — formatting for English is still being researched.
How we checked
Based on general principles of working with AI and on our development team's own experience of dictating to AI in Japanese. This is not a study that measured the effect.
