You have probably done this. Someone you love is hurting, or you owe someone an apology, and you open an AI tool and type: Write a message to my friend whose dad died, something comforting but not cheesy. Eleven words. The reply arrives in two seconds, and it is fine. It is also nothing like her, nothing like him, and nothing like the friendship you actually have. You read it, feel a small flatness, and close the tab.

It's easy to blame the tool. But the tool answered the question you asked. The problem is that the question you typed held about one percent of what you know.

The gap between what you know and what you typed

In 1990, Stanford graduate student Elizabeth Newton ran a now-famous experiment. She asked people to tap out the rhythm of a well-known song, like "Happy Birthday," on a table while a listener tried to name it. Before the test, the tappers predicted the listeners would guess correctly about half the time. In fact, listeners guessed right only a few times out of 120 songs, roughly 2.5 percent.

The tappers could hear the melody in their heads. They couldn't imagine not hearing it. This is the curse of knowledge: once you know something, it becomes very hard to simulate the mind of someone who doesn't.

An AI model is the listener at the table. It hears your taps and nothing else. It doesn't know your friend is the one who drove four hours when your car died. It doesn't know the two of you never say "I love you" and just trade bad puns. It doesn't know her dad taught you to parallel park. All of that is the melody in your head. Eleven typed words are the taps.

Why we type so little

Here is the quiet part. We write short prompts because typing costs effort, and people are very good at minimizing effort without noticing. Every extra sentence you type is friction, so you stop at the point where your request feels complete to you. And it does feel complete, because you're filling in all the missing context automatically. That is the curse of knowledge working in your own favor, and against you.

Speech doesn't have this problem to the same degree. When you talk to a person, you ramble. You circle back, add a "well, the thing is," and throw in the detail that explains why it matters. Linguists have long noted that spontaneous speech is looser and more redundant than writing. In a prompt, that looseness is not a flaw. Redundancy and tangents are exactly the raw material a model needs: the stories, names, tone, and stakes.

The richest prompt is often not a well-engineered one. It's the thing you'd say if a thoughtful friend asked, "Okay, tell me about her."

What a model can't guess

Generic output comes from underspecified input. When a request leaves out the important details, the model fills the gap with the most statistically common version of that kind of message. For condolences, that means "thinking of you during this difficult time." It isn't wrong. It's just average, and the average is the one thing you were trying to avoid.

Four kinds of information do most of the work in a personal prompt:

  • The relationship. Not "my friend," but how long, how you talk, what you joke about, what you've never said out loud.
  • The situation. What actually happened, in plain order, including the part you're embarrassed by.
  • The goal. What you want the person to feel or do after reading it. "I want her to know I'm not going anywhere, without her having to reply" is a goal. "Be comforting" is not.
  • The guardrails. What would make it wrong. "Don't say 'at peace.' Don't make it about me. No exclamation points."

Notice that none of these is clever prompt engineering. They're just facts you already have, and a naive typed prompt omits all of them.

The one-minute ramble

So the fix isn't a magic template. It's changing the medium. Instead of composing a prompt, talk one.

Press record, or dictate straight into the prompt box, and speak for sixty to ninety seconds as if explaining the situation to a friend who has never met these people. Don't organize. Don't edit as you go. Say the awkward bits. Say "I don't know how to put this, but..." Then stop and read what you've got. It will probably be messy and full of the good stuff.

A typed version might be 11 words. The spoken version tends to be a few hundred, and it contains the drive in the dead-car story, the pun habit, and the fact that you're afraid of saying the wrong thing. The model now has something to work with, and what comes back sounds like it knows who it's writing for.

Show it what "you" sounds like

Context about the situation is half of it. The other half is voice. If you want output that sounds like you, give the model a sample of you. Paste in a text or two you've sent this person, or a few lines you've written that you like. Models are strong at matching style when they have an example to match, and weak at it when all they get is the adjective "warm."

Then treat the first draft as a draft. Ask for a rewrite in a different register: shorter, plainer, less formal. Compare versions the way you'd compare takes, and end by reading the winner out loud. If you stumble over a phrase, you wouldn't say it, so cut it.

Your next moves

Try these today, on a message you've actually been putting off:

  • Record a 90-second ramble before you type anything. Pretend you're explaining the situation to a friend who knows none of the people involved. Use your phone's voice memo or any dictation tool, and don't stop to fix mistakes.
  • Make sure four things are in it: your relationship to the person, what happened, what you want them to feel afterward, and at least one phrase you refuse to use.
  • Paste a sample of your own writing. Find one real message you've sent that sounded like you, and include it with the line "Match this voice."
  • Ask the model to question you. Add: "Before you write anything, ask me up to three questions about what I left out." Answer them out loud. The answers are usually the best material.
  • Read the final draft aloud once, and fix every line you wouldn't say. Then send it before you talk yourself out of it.

Where this leaves you

The frustrating thing about generic AI writing is that it feels like a limit of the technology. Often it's a limit of the channel. We hand over a few typed words and expect a model to reconstruct a relationship that took years. When you speak the context instead, the same tool behaves very differently, because you finally gave it something to hear.

This is the habit Quill is built around. It's dictation that works in any app, runs on your device so your private ramble about a grieving friend stays private, and turns clean text into a rewrite in whatever style you want with one tap. Talk first, shape second. If you'd like to try it, you can find it at quill.lumenlabs.works. But the idea works with any microphone: the next time you're tempted to type eleven words, say three hundred instead.