Interactive eBook BuilderInteractive eBook Builder
$59 · unlimited access, unlimited eBooksGet it for $59

· Updated · By · 7 min read

How to Get AI to Ask Clarifying Questions Before It Writes

aipromptsworkflowebook

To get AI to ask clarifying questions, say so explicitly: tell it not to write yet, cap the number of questions, name the areas to probe, and tell it to wait for your answers. By default most assistants guess, because guessing produces something you can react to. For a quick email that's fine. For a chapter, a plan or a product, a guess usually means a generic draft.

Here is a prompt you can paste, a worked example with the questions and answers, the research behind the idea, and when to skip it.

What is the prompt that makes AI ask first?

Give it a role in the process (interviewer), a question cap, the topics to cover and a stop rule. Copy this and replace the brackets:

I want help with: [one-line task].
Before you write anything, interview me. Ask no more than 5 questions, numbered, covering what you most need to know about:
- who the reader is and what they already know
- what they should be able to do after reading
- scope: what is in and what is out
- the voice I want
- facts, sources or examples I must use
Do not start the draft. When I reply, summarize your understanding in 3-4 sentences and wait for me to say "go".

Each part does a job. The cap keeps it from interrogating you. The topic list stops it from asking trivia like "what font do you prefer?". The stop rule counters the model's urge to be helpful by producing text. The summary step catches misunderstandings while they are still cheap to fix.

What does it look like with a real task?

It looks like a five-minute exchange that replaces three rounds of rewriting. Below is an illustrative exchange I wrote to show the shape, not a transcript of a specific model.

You: I want help with: a short guide on meal prepping. [interview prompt above]

AI:
1. Who is this for (students, parents, athletes) and how much cooking do they already do?
2. What should they be able to do after reading: plan one week, cut costs, eat healthier?
3. How much time and money per week are you assuming?
4. Do you want recipes included, or just the system?
5. Any voice preference: friendly coach, no-nonsense, funny?

You:
1. Busy parents, cook a few nights a week.
2. Plan and prep 5 dinners in 2 hours on Sunday.
3. 2 hours, around a normal grocery budget.
4. System plus 5 sample dinners.
5. Friendly coach, no guilt.

AI: Here's my understanding: a guide for busy parents that teaches a 2-hour Sunday prep routine producing 5 dinners, in a friendly, guilt-free voice, with a system and 5 sample dinners. Say "go" to draft.

Compare that with the original request, "a short guide on meal prepping". Without the exchange the model has to choose a reader, a goal, a length and a voice on its own, and it will choose the most average ones. With the exchange, five decisions you already had in your head are on the page. The difference in output is the one covered in our post on how long a prompt should be: more relevant information beats more words.

Does this approach have evidence behind it?

Yes, in a limited way: research shows model-led questioning can pull out better task descriptions than users write alone. In "Eliciting Human Preferences with Language Models" (Li et al., 2023), the authors propose Generative Active Task Elicitation (GATE), where the model interacts with the user in free-form language to work out what they want. They tested it in three domains (email validation, content recommendation and moral reasoning) in preregistered experiments. Their results: LMs prompted to elicit preferences, for example by generating open-ended questions or synthesizing informative edge cases, got responses that were often more informative than user-written prompts or labels. Users reported that interactive elicitation took less effort than writing prompts or labeling examples, and that it surfaced considerations they hadn't anticipated.

That research isn't about writing books, and I don't want to oversell it. But it supports the practical point: a model asking "what about this edge case?" can show you a gap you didn't know your prompt had.

Documentation points the same way. Google's guidance on agent system instructions lists "ambiguity and permission handling" as a setting: when the model may make assumptions and when it must stop and ask the user. Anthropic's tip to imagine handing your prompt to a colleague with no context covers the same gap from the other side; a good colleague would ask questions.

Which questions are worth asking?

The ones that change the output. Here are the areas I'd probe for writing tasks, with an example question and what it prevents:

AreaExample questionWhat it prevents
ReaderWhat do they already know?Explaining basics to experts, or skipping steps for beginners
OutcomeWhat should they do after reading?Pleasant text with no point
ScopeWhat is out of scope?Sprawling coverage and padding
VoiceCan you share a sample or three adjectives?Default AI tone
EvidenceWhat facts or stories must be included?Invented examples
ConstraintsLength, format, things to avoid?Rewrites for shape

This is the thinking that goes into a book plan too. In Interactive eBook Builder, the Configuration Hub collects your title, audience and preferences up front, and its Quick Start randomizer can propose a starting point if you're stuck, so many of these questions are answered before any prompt is written. You can still use the interview prompt above for anything the hub doesn't cover.

How do I turn the answers into a draft?

Ask for a short brief, approve it, then draft from it. The brief is the interview's output. A good one is 80 to 150 words and states the reader, outcome, scope, voice and must-include facts. You then paste it into a fresh chat, or continue, with "Go". If you have a rough pile of notes rather than a task, start with how to turn a messy brain dump into a prompt, which converts raw notes into the same kind of brief. Once the brief exists, turning an outline into an ebook draft shows how to feed it chapter by chapter.

For a whole book, the interview becomes the first step in a chain. Prompt chaining for planning, drafting and editing shows the order: interview, plan, draft, edit. In Interactive eBook Builder, Agent A Spec Builder plays that front-of-chain role by turning your setup into the Chapter Plan that Agent B drafts from.

When should I skip the questions?

Skip them for small, clear tasks. "Fix the grammar in this paragraph" or "give me ten title ideas for a book about X" don't need an interview. The method pays off when the task is ambiguous, expensive to redo, or involves your preferences: chapter plans, positioning, voice and anything that you'd otherwise rewrite three times.

Before you hit send

The takeaway: a five-question interview moves the guessing from the model to you, and that's where the right answers are.

FAQ

Why would AI ask me questions instead of just writing?

Because your first prompt probably left things out that you know and the model doesn't, such as the reader, the goal and your limits. Researchers behind GATE (Li et al., 2023) found that when language models asked users open-ended questions, the elicited responses were often more informative than user-written prompts, and users said it took less effort.

How many questions should I allow?

Three to seven is a practical range (my suggestion, not a tested number). Fewer misses key gaps; more feels like a form and burns your patience. You can always say "ask three more".

Should it ask all questions at once or one at a time?

All at once is faster, and you can answer in a numbered list. One at a time works better when each answer changes the next question, for example when exploring a new topic.

What if the AI still guesses instead of asking?

Make the rule explicit and put it last: "Do not write any draft until I reply. If you start writing, stop." Some products and APIs also let you set assumption-versus-ask behavior in the system prompt.

Ship niche eBooks with agents already built.

Get the Interactive eBook Builder for $59