The worst email I ever sent wasn't rude, badly written, or late.
It was accurate. That was the problem.
I asked ChatGPT to draft an email declining a request from a colleague, gave it a three-sentence description of the situation, and sent the result after a quick read. It was clean, professional, and completely wrong for the relationship. I had described the request but not the person. ChatGPT didn't know that this particular colleague had gone out of her way to help me two months earlier, and the email read like a polite no from a stranger.
She replied with one line: "Happy to help however I can, just let me know."
I felt terrible. Not because I'd done anything wrong, but because the email had made a decision I hadn't made—it treated a favor-adjacent request as a transaction, and that wasn't how I actually wanted to handle it.
The fix wasn't better writing. It was better questions. And the surprising part is that you can ask ChatGPT to ask you the questions before it writes anything.
Here's how that works, why it matters, and where it still falls short.
The Real Problem: Silent Assumptions
When you give ChatGPT a task, it fills every gap you leave with an assumption. Most of the time those assumptions are reasonable. Sometimes they're wrong in ways that matter, and you won't notice until someone reads what you wrote.
The problem isn't that ChatGPT guesses. It's that it guesses silently and then writes with total confidence, which makes the guess invisible.
The obvious fix is "give better context." But that puts the burden on you to anticipate every gap, and you can't. You don't know what you don't know until you see it.
So instead of trying to write a perfect prompt, I started doing something different: I ask ChatGPT to interview me first.
The Interview Prompt
Here's the prompt I use. It's short, and it's changed how I write almost everything that matters.
"I'm going to give you a task. Before you write anything, ask me 3–5 clarifying questions that would most change the output. Focus on things you genuinely can't infer from what I've told you.
Do not ask questions you can already answer. Do not ask about formatting or length—I'll specify those. Ask about the situation, the reader, and anything that would make your first draft wrong if you guessed.
Task: [your task here]"
The key line is "focus on things you genuinely can't infer." Without it, you get questions like "What tone would you like?" which is a question you could have answered yourself.
With it, you get questions like "Is this person expecting a yes?" or "Has this topic come up before?"—the kind of thing that actually changes the draft.
Why "3–5 Questions" and Not "As Many As You Need"
Because unlimited questions is unusable.
I tried "ask me anything you need" once and got eleven questions, most of them trivial. Sorting through them took longer than writing the email would have.
Three to five is the sweet spot. It forces ChatGPT to prioritize, which means it asks about the things that matter most. If a question matters and doesn't make the list, I'll usually catch it when I read the draft.
A Real Example
I'll use a version of the email that started this whole thing.
My Task
"Write an email declining a request from a colleague."
The Questions It Asked
What is the request, and why are you declining? (Necessary—I hadn't said.)
What is your relationship with this person? (This is the one that would have saved me.)
Is there an alternative you can offer, or is this a hard no? (Good question. I hadn't thought about it.)
Does this person know why you might decline, or will it come as a surprise? (Excellent question.)
Four questions. Every one of them changed the draft.
Question 2 is the one that mattered most. When I answered "someone who helped me recently and that I want to keep a good relationship with," the tone changed completely. It wasn't a polite refusal anymore—it was a warm one with an offer attached.
Question 3 made me realize I could offer something. I hadn't considered it because I was focused on the no.
That's the whole value. The questions didn't just improve the email. They improved my thinking about the email.
Where This Works Best
I've been using this approach for a while now, and it's not equally useful everywhere. Here's where it earns its keep.
Anything Involving a Relationship
Emails to coworkers, clients, vendors, or anyone you'll interact with again. The relational context is exactly what ChatGPT can't infer and exactly what changes the output.
Anything With History
If the task references something that happened before—a prior conversation, an earlier decision, a recurring issue—the interview surfaces it. Otherwise ChatGPT writes as if this is the first time.
Anything With Consequences
Declining a request. Delivering bad news. Pushing back on a deadline. Anything where getting it wrong costs something. The extra two minutes of questions is cheap insurance.
Anything You've Been Thinking About Loosely
Sometimes I know I need to write something but haven't fully figured out what I want to say. The interview forces me to articulate it. Half the time, I discover I didn't actually know what I wanted.
Where It Doesn't Help
Being honest about the limits.
Simple, Mechanical Tasks
If I need a summary or a reformat, there's nothing to clarify. The interview just adds a step.
When You Already Know What You Want
If I've already thought through the situation, the questions come back obvious and I skip them. That's fine—it's information too.
When the Questions Are Generic
Some tasks don't have hidden depth. "Rewrite this sentence" doesn't need an interview. If the questions come back as filler, that's a signal the task was simple enough to just do.
When You're in a Hurry
The interview costs about two minutes. If two minutes matters—and sometimes it does—skip it and accept the risk. Just know you're accepting the risk.
What This Doesn't Fix
I need to be clear about one thing, because it's the limit of the whole approach.
ChatGPT doesn't know what to ask about. It's guessing which gaps matter, based on patterns it's seen. Most of the time it guesses well. Sometimes it asks about something that doesn't matter and misses something that does.
The interview is better than no interview. It is not a substitute for your own judgment about what matters.
This is the same division of labor as everything else I write about. ChatGPT helps you think. You do the thinking.
The Test Card
Task: Get a first draft that reflects the real situation, not ChatGPT's assumptions about it.
Prompt or workflow: State the task, then ask for 3–5 clarifying questions that would most change the output, explicitly excluding questions it can already answer and questions about formatting or length. Answer the questions, then request the draft.
Starting conditions: A task involving a real relationship, some history, and consequences if handled badly.
Time before ChatGPT: Writing from scratch with no structure: 15–25 minutes, usually with at least one revision after rereading.
Time after ChatGPT: About 8 minutes—1 minute to describe the task, 90 seconds for questions, 2 minutes answering, 90 seconds generating, 3 minutes reviewing.
Editing required: Light. The draft is usually close on the first pass because the gaps got filled in before writing, not after.
What went wrong: Early versions asked generic questions ("What tone should I use?"). Fixed with explicit instructions to exclude inferable questions and formatting questions. Also: it will still miss things. Expect to add one detail it didn't ask about.
Privacy notes: The questions it asks are about the situation, not about people. I still strip names, clients, and identifiers before pasting anything. The interview process doesn't change the sanitization rules.
Who should not use this method: Anyone working on simple, mechanical tasks where there's nothing to clarify. Anyone who won't actually answer the questions thoughtfully—the prompt only works if you engage with it. And anyone whose workplace prohibits external AI tools, which no prompt technique changes.
The Bottom Line
The email I sent that day wasn't a ChatGPT problem. It was a "I described the task instead of the situation" problem.
Asking it to interview me first fixes that. Two extra minutes, four questions, and a draft that reflects what's actually going on instead of what I remembered to mention.
It doesn't make ChatGPT smarter. It makes the gap between what I said and what I meant visible before I hit send.
I tried it at my desk so you don't have to. No magic. Just four questions and one email I wish I could take back.
