ChatGPT is very good at answering questions.
It is much worse if you want it to ask you questions.
By default, ChatGPT assumes you want a fast, neat answer and a polite goodbye. That’s useful when you already know what you’re asking. It’s terrible when you don’t. The moment your question is vague, underspecified, or half-formed (which is most real-world problems), ChatGPT happily fills in the gaps and charges ahead, probably agreeing with you, without really knowing what you meant.
Normally this means that the missing information is filled in with “most likely guess” which leaves a world of room for misunderstanding. If you want better results, you need to flip the interaction around. You need ChatGPT to behave less like a search box and more like an interviewer, consultant, or therapist — someone who refuses to answer until they actually understand what you mean.
Here’s how to do that, step by step.
The fix is not “better prompting” in general. The fix is telling ChatGPT, very clearly, that part of the task is to ask you questions.
Step 1: Tell ChatGPT to Delay Its Answer Until It Understands You
This is the single most important step.
You need to override ChatGPT’s instinct to answer immediately. Do that explicitly, in plain language.
Example prompt:
“Don’t answer yet. First, ask me clarifying questions so you fully understand what I’m trying to achieve.”
This instruction does two things:
- It gives ChatGPT permission to not answer.
- It reframes success as understanding, not completion.
Without this step, everything else is weaker.
If you only remember one trick from this article, remember this one.
Step 2: Give ChatGPT a Role Where it’s normal to ask you questions
ChatGPT copies patterns. If the role normally asks questions, ChatGPT will too.
Some roles that reliably trigger good follow-ups:
- investigative journalist
- consultant
- therapist
- hiring manager
- editor
- researcher designing a study
Example prompt:
“Act as a consultant. Before giving advice, ask the questions you would normally need answered.”
This works because consultants don’t give answers blind. Neither should ChatGPT — once you tell it to behave like one.
Avoid vague roles like “expert” or “assistant.” They tend to answer, not interrogate.
Step 3: Specify the Type of Follow-Up Questions You Want
If you don’t specify the kind of questions, ChatGPT will ask safe, generic ones. You can do much better by constraining the direction of its curiosity.
Here are examples that work well:
- “Ask questions about constraints, not preferences.”
- “Ask questions that reveal missing information.”
- “Ask questions that would change the answer if answered differently.”
- “Avoid yes/no questions.”
Example combined prompt:
“Before answering, ask follow-up questions that focus on constraints, edge cases, and missing context. Avoid yes/no questions.”
This turns ChatGPT from polite conversationalist into something closer to a diagnostic tool.
Step 4: Force it to ask you questions a minimum number of times
ChatGPT is surprisingly obedient to numbers.
If you say “ask me questions,” you might get one. If you say “ask five,” you will almost always get five — and the quality improves as it’s forced to dig deeper.
Examples:
“Ask at least 5 clarifying questions before answering.”
or
“Keep asking questions until you can summarise my problem without ambiguity.”
This is especially useful for complex tasks (strategy, writing, technical problems), where the first two questions barely scratch the surface.
Step 5: Gamify Curiosity (This Works Better Than It Should)
This sounds silly. It works anyway.
ChatGPT responds very strongly to feedback loops. If you frame good questions as “points” or “progress,” it leans into them.
Example prompt:
“You get 1 point for every genuinely useful clarifying question. Try to reach 10 points before answering.”
Or:
“Your goal is to maximise understanding, not speed. Ask as many high-quality questions as needed.”
You’re not training the model. You’re nudging its behaviour. And for reasons that aren’t entirely clear, this nudge is very effective.
A Simple Reusable Prompt You Can Save
If you want a one-paste setup, this works reliably:
“Don’t answer yet. Act as a consultant. First ask at least 5 clarifying questions focusing on constraints, missing information, and edge cases. Avoid yes/no questions. Only answer once you’re confident you understand my goal.”
References:
Park, C et al, 2025, “When to Ask a Question: Understanding Communication Strategies in Generative AI Tools” https://dl.acm.org/doi/full/10.1145/3708319.3733711