Tell AI to ask you questions first
One instruction, and the draft finally sounds like your business.
By Chris Battis
A friend of mine runs a two-person bookkeeping shop. Tuesday morning, a client's invoice hit day 60 unpaid, and she needed the collection email out before lunch. She pasted the situation into ChatGPT: client owes $4,200, sixty days late, needs a firm but friendly note.
Thirty seconds later: a template. Firm-ish. Friendly-ish. The kind of email any bookkeeper could send any client about any overdue invoice.
She almost sent it. Then she deleted the draft and typed one line above her original request:
"Before you answer, ask me what you need to know to write this well. Wait for my answers."
The model asked back: how many times has this client paid late before, is this the first notice or a later one, does the contract mention a late fee, and does she want to leave the door open or signal this is serious. She answered in two sentences: third notice, good client otherwise, no late fee in the contract, wants firm without threatening.
The redraft dropped the generic reminder language and opened with the actual history: this is the third time we've had this conversation. It read like her, because for once the model had what it needed to sound like her.
- Your request
- It guesses
- You rewrite it
- Your request
- It asks
- You answer
- It writes
The blanks get filled with average, not with true
A bare request has unstated details in it: the relationship, the history, the tone that fits this specific client. The model can't leave those blank, so it fills each one with whatever's most common across every similar request it's seen. Average client, average payment history, average tone. That's why the first draft reads like it was written for nobody in particular. It was.
The usual fix is to read the draft, notice what's wrong, and rewrite it yourself. That fixes this one email. It doesn't fix the next one, because the model never learned what it was missing. You corrected around it and moved on.
Asking it to interview you first flips the order. Instead of guessing and hoping you'll fix it, it names exactly what it doesn't know, and you fill only those slots. The model still writes the whole thing. You just stopped being the one supplying the missing context after the fact.
The instruction. Paste it above anything you'd actually rewrite by hand
Before you answer, ask me what you need to know to write this well.
Don't guess. Wait for my answers, then write it.
Save it for requests where a wrong guess costs you something: a client email, a proposal, an about-us page. Skip it for a brainstorm list; ten podcast topic ideas don't need an interview. And answer every question it asks, even the one that feels obvious. The obvious one is usually the assumption it would have gotten wrong.
This works the same in Claude, ChatGPT, or Gemini. It has nothing to do with which model is sharper. It's about whether you let it ask before it writes.
Same invoice. Same model. The only thing that changed was who filled in the blanks
My friend didn't switch tools, and she didn't write a better prompt. She answered four short questions instead of rewriting one bad email. That trade gets better every time she does it, because the questions get faster to answer and the drafts get closer to right on the first pass.
Next time you're about to paste in a request you actually care about, add the line first. Let it ask. Then let it write.
