It's Not That AI Is Bad — You're Just Not Giving It Enough

You ask AI to build you a budgeting app. What comes back looks like it was designed by a random number generator — the colors clash, the layout is off, and the categories are hardcoded to four options you can't change. You click around and it gets worse: empty states aren't handled, there's no delete confirmation, and if you add an entry then switch tabs, the data is gone.

You send a few more messages, a few more rounds of fixes. Thirty minutes later, it's still full of problems. You conclude: AI just isn't smart enough.

But if you'd said everything upfront — who the user is, what the core use case is, how data should be stored, whether categories are fixed or custom, whether you need charts, whether it needs to work across devices — AI wouldn't have received a single sentence. It would have received a complete picture. No guessing needed, and the odds of getting it right the first time go way up.

A 50-word prompt and a 500-word prompt produce completely different results.

The question is: who wants to type 500 words?

Typing quietly deletes your own thoughts

A moderately complex requirement — background, specifications, constraints, expected output format — easily runs to three or five hundred words when fully expressed. But halfway through typing, what you meant to say at the beginning is already getting fuzzy. You had a complete chain of logic in your head; typing broke it apart.

So you instinctively cut corners: fine, I'll just send a couple of sentences.

This isn't your fault. Psychologist Kellogg studied what the brain does during writing: when you type, you're not just "expressing" — you're also organizing phrasing, operating the keyboard, and checking what's on screen. These tasks compete for your limited attention, leaving less room for "figuring out what you actually want to say."

Your brain has a "thinking budget." Typing is spending that budget. So your output gets compressed — not because you didn't think enough, but because typing itself is stealing part of your capacity to think.

Speaking doesn't just express thought — it shapes it

There's a deeper difference between speaking and typing.

Have you ever had this experience: an idea is vague and formless in your head, but when you try explaining it to someone, it suddenly becomes clear? You're speaking and you surprise yourself — "Oh, so that's what I think."

Vygotsky wrote in Thought and Language that language isn't just a container for thought — language itself participates in forming thought. The ideas in your head are compressed, jumpy, incomplete. When you speak them, you're forced to expand the compressed logic, fill in the gaps between leaps, and make vague concepts concrete. The act of speaking is itself an act of thinking.

Typing can do this too, but because it's slow and cognitively expensive, the expansion keeps getting interrupted. You unfold an idea halfway, then stop to type. By the time you're done typing, half the context is lost.

Speaking doesn't have this problem. It's fast enough and natural enough that your thoughts can unfold continuously — one point leads to the next, which leads to the next, the chain unbroken. This is why, when you describe a requirement to a colleague out loud, you naturally include background and details that you'd skip when typing.

A study from the University of Mannheim confirmed this directly: given the same open-ended questions, voice responses were more than twice as long as typed responses, and covered significantly more topics. It's not that the speakers were more talkative — it's that typing forces people to self-edit.

This is how AI loses its context

Back to the AI scenario.

When you write a prompt, the most important part isn't the instruction itself — "build me a budgeting app" is something anyone can write. What matters is the context around the instruction: who the user is, what the scenario is, what constraints exist in the data structure, what interaction preferences you have, how edge cases should be handled.

You have all of this in your head. But while typing, it gets shaved away layer by layer: this detail is too granular, I'll skip it; AI should be able to figure that out on its own; I was going to type this but forgot what I was saying three sentences ago.

What you end up sending is the version that made it through the "keyboard filter" — a truncated, incomplete version. AI can only work with what it receives — and when it guesses wrong, you blame it for not being smart enough.

What if you could speak all that context instead? In two minutes of speaking, you convey far more information than two minutes of typing. Background, constraints, preferences, edge cases — they come out naturally when you're speaking, instead of being weighed one by one against "is this worth typing?"

From speech to prompt

The problem is that ordinary voice input produces text you can't paste straight into an AI — filler words, repetitions, no punctuation, one giant unbroken block. You'd have to clean it up first, and that raises the barrier again.

Flow Keyboard turns what you say into text you can use immediately: filler words removed, punctuation and paragraphs in place, logic smoothed out. You can pause, repeat, change direction mid-sentence — the output is ready to paste into an AI conversation, send to a colleague, or drop into a document.

Two minutes of speaking produces roughly two to three hundred words of clean prompt. Compared to the two lines you'd have squeezed out on a keyboard, the context AI receives is entirely different.

AI didn't get smarter

The old workflow: you want to build something → type a prompt (2 minutes, 50 words) → AI misunderstands → another round of clarification → another round → thirty minutes gone.

The new workflow: you want to build something → speak for two minutes → a complete requirement description comes out → paste it into AI → one round, done.

The difference isn't on AI's end. The difference is on yours — you finally said everything that needed to be said.