What it is
Most prompts leave gaps. You ask for a job description and forget to say how senior the role is, so the model picks the most likely answer and writes on. The draft reads well, and the guess is buried in it where you may not notice.
Asking it to ask questions first reverses the order. You add one instruction: before you write, ask me about anything you would otherwise have to guess. The model replies with a numbered list of questions. You answer each in a line, and only then does it draft.
This works because a model is good at seeing what a task needs and has no way of knowing your situation. It can tell that a price-increase email depends on whether customers on annual plans are affected. It cannot know the answer. The interview splits the job along that line: the model brings the checklist, you bring the facts.
When it helps
Use it when the task depends on facts about your company, your people or your situation that you have not written down: a job ad, a policy, an event plan, a delicate email. Those are the requests where a plausible guess does the most damage, because it looks finished.
It also helps when you are not sure what matters yet. A model that asks “Who approves exceptions?” can show you a decision you had not made. And it pays off when the first draft would be slow to redo, such as a long document, a presentation or anything that goes to many people.
Finally, use it the first time you brief a task you will repeat. The questions the model asks are the blanks your saved prompt needs. Once you have answered them, turn the chat into a prompt and the questions become part of the brief.
When it does not
It costs a round trip. For a quick question, or a draft you will rewrite anyway, two turns are slower than one, and a decent guess is fine.
If you already know what the model needs, write it in. Giving it the facts up front is faster than waiting to be asked for them, and the model cannot ask about something it does not know exists, such as a customer who complained last month.
It does not work where nobody is there to answer. A prompt that runs in a spreadsheet, an automation or an API pipeline needs a rule instead: tell it what to do when unsure, for example to write [CHECK] rather than guess.
Left open, the questions can be generic. Ask for questions with no limit and you may get ten, half of them answered in your prompt already. Cap the number, ask for the most important first, and tell it to skip anything it can work out. Some newer models already ask when an answer would change the result, and can stop to ask when you expected them to carry on, so say which you want.
Before and after
Three requests from different kinds of work. The scores are the Prompt Grader’s free estimate with Written for set to Any chat model. A score counts what the prompt states; it does not grade the answer.
Write a job description for a customer success manager.
Write a job description for a customer success manager. It will be posted on our careers page and on LinkedIn, and it is the first thing candidates will read about the role. We are a 40-person software company that sells scheduling tools to dental clinics. Before you write anything, ask me up to five questions about what you would otherwise have to guess, such as seniority, which accounts the person would own, and the salary range. Ask them all in one message, numbered, most important first, and wait for my answers. Skip anything you can work out from what I have told you.
The first prompt gets a complete job description with an invented seniority, a generic list of duties and no pay range. The second gets five questions back, and the job description that follows names the real accounts and salary band.
Write an email to our customers about the price increase.
Write an email to our existing customers announcing a price increase. The goal is that they understand the change and do not feel ambushed. Facts: our Standard plan goes from $29 to $34 a month on March 1. Annual plans renew at the old price until their renewal date. Before you draft, interview me: ask me one question at a time about anything that would change what the email says or how it sounds, and wait for each answer. Stop after five questions at most. Then write a subject line and a body of under 200 words. Do not apologize for the increase.
Without the interview, the model has to invent the amounts and the date, or leave brackets. With it, the model can ask about what customers will want to know, such as whether anything new comes with the price, and the email answers that before anyone writes in.
Plan our team offsite.
Help me plan a one-day offsite for my operations team of 12. The goal is to agree on next year's three priorities and fix how we hand work between shifts. Before you suggest an agenda, ask me the questions you need answered, numbered, no more than six. Leave out anything you can work out from what I have told you. After I answer, give the agenda as a table with times, sessions and who leads each one.
A bare request returns a generic offsite with icebreakers and a team lunch. The second asks about location, budget and which shifts can be away at once, and the agenda is built around those answers.
The line to add
Put it at the end of the prompt, after the task and any facts you already have:
Before you start, ask me up to {{number}} questions about anything you would otherwise have to guess. Ask them all in one message, numbered, most important first, and wait for my answers. Skip anything you can work out from what I have already told you.
Three to five questions is enough for most emails and documents. If the answers depend on each other, as in planning an event, ask for one at a time instead:
Interview me one question at a time. After each answer, ask the next most useful question. Stop when you have enough for a good first draft, or after {{number}} questions, then write it.
You do not have to answer everything. “You decide” is a fine answer to a question that does not matter to you, and it tells the model where it has room.
In ChatGPT, Claude, Gemini and Copilot
The line asks only for a reply in text, so it works the same way in all four. What differs is how often each asks unprompted. OpenAI’s guide to its newest GPT models says the model “asks focused questions when the answer could change the outcome”, and notes that this can make it stop when you expected it to make reasonable assumptions and carry on (Using GPT-6). The line removes the uncertainty in either direction.
Anthropic’s guidance describes Claude as a brilliant but new employee who lacks context on your norms, and suggests a test: show the prompt to a colleague with little context, and if they would be confused, Claude will be too (Prompting best practices). Asking for questions runs that test with the model as the colleague.
Microsoft’s guide to Copilot prompts says to expect some back-and-forth to get the result you want (Get started writing prompts). For a task you do every week, put the line in a Custom GPT, a Claude Project or a Gem, and every new conversation opens with the questions. The assistant instructions template shows the layout.
Key takeaways
- Ask the model to interview you when the task depends on facts only you have.
- Cap the number of questions and ask for the most important first.
- Tell it to skip what it can work out from your prompt.
- Where nobody can answer, such as an automation, give a rule for gaps instead.
- The questions it asks are the blanks your reusable prompt needs.
Questions
How do I get ChatGPT to ask me questions before it answers?
Say so at the end of the prompt: “Before you start, ask me up to five questions about anything you would otherwise have to guess, and wait for my answers.” Without a line like that, most chat models write a full answer from their best guess. Capping the number and asking for the most important first keeps the list short and relevant.
Why does AI make things up instead of asking?
A chat model is built to answer, so when a detail is missing it fills the gap with the most likely value and carries on. Some newer models ask more often when an answer would change the result, but none can know which facts you have not shared. Ask for questions first, or tell it to write [CHECK] instead of guessing, as in tell it what to do when unsure.
Should the AI ask all its questions at once or one at a time?
All at once is faster and works for most emails and documents: you answer a numbered list in one reply. One at a time suits tasks where each answer changes the next question, such as planning an event or a project. Either way, set a limit, such as five questions.
Prompts to use with this
- Prepare questions to ask at the end of a job interviewYou want thoughtful questions that help you judge the job and show you did your homework, instead of generic ones.
- Draft landing page copy from product factsYou have a product or feature to launch and a page to fill, and you want copy built from what is true, not from hype.
- Answer a customer question from your help docs and policyYou need an accurate, friendly reply that only uses what your documentation actually says.