Lesson 19

Why Asking Questions Is a Superpower (Yours, Not Just the Model's)

Most people treat a language model like a vending machine: you put in a prompt, you get a snack. If the snack is wrong, you shake the machine and try again. This works for simple tasks — "write a haiku about coffee" — but it collapses the moment your project has real complexity. A vague request like "help me plan a product launch" produces a generic, five-bullet-point answer that any intern could have written. The model guesses what you mean, and because it is optimized to be helpful, it confidently delivers a polished guess. The result is not wrong in a way you can easily detect; it is wrong in a way that wastes your time.

The professional fix is not to write a longer prompt. It is to make the model stop guessing and start asking. This chapter shows you how to force that behavior, how to break a large project into sub-tasks the model can handle sequentially, and how to build a prompt that treats missing information as a feature, not a bug.

The Core Technique: Instruct the Model to Ask Before It Answers

By default, most models will answer immediately with whatever they can infer. You must explicitly override this default. The instruction is simple, but it must be precise. Do not write "ask me if you have questions" — the model will interpret that as optional. Write a hard rule.

Here is the pattern that works consistently across GPT-4, Claude, and Gemini:

You are a senior project strategist. Before you provide any plan, analysis, or recommendation, you must ask me a minimum of 5 clarifying questions about the project scope, constraints, audience, and success metrics. Do not provide any content until I have answered all questions. After I answer, produce the deliverable in full.

Notice three things. First, the role is specified ("senior project strategist") — this sets the tone and depth. Second, the number of questions is fixed ("minimum of 5") — this prevents the model from asking one lazy question and then proceeding. Third, the sequence is locked: questions first, content second. The model cannot skip ahead.

Worked Example: A Multi-Step Project Prompt

Let us apply this to a realistic scenario. Suppose you are a marketing manager at a mid-sized B2B software company. You need a comprehensive go-to-market plan for a new feature called "Automated Invoice Reconciliation" that your product team just finished. You have no internal template, and you want the model to act as an external consultant.

Here is the full prompt you would paste into ChatGPT, Claude, or Gemini:

You are a B2B SaaS go-to-market consultant with 15 years of experience launching features for companies with 50–500 employees. I need a complete launch plan for a new feature called "Automated Invoice Reconciliation" in our existing accounting software product.

CRITICAL RULE: Before you write anything, ask me at least 6 clarifying questions. Do not output any part of the plan until I have answered all 6. Your questions must cover: target customer segment, current user base size, pricing model, competitive landscape, available launch budget, and internal team capacity.

After I answer, structure your response as follows:
1. Executive summary (150 words max)
2. Target customer profile (with specific job titles and pain points)
3. Positioning statement (one sentence)
4. Launch timeline (week-by-week for 8 weeks, with specific deliverables per week)
5. Channel strategy (which 3 channels to prioritize and why)
6. Success metrics (with specific numbers: adoption rate, activation rate, churn impact)
7. Risk register (top 5 risks with mitigation plans)

Be specific. Use real-world benchmarks for B2B feature adoption (typical activation rates, typical pilot durations). Do not use generic advice like "leverage social media" — name the platforms and the exact type of content.

When you run this, the model will stop and ask you six questions. It might ask: "What is the current size of your active user base?" or "What is your pricing model — per seat, per feature, or bundled?" You answer each one in plain text. Then, and only then, does the model produce the full plan.

Why This Works: The Sub-Task Decomposition

Notice what happened in that example. You did not just ask for a plan. You decomposed the project into seven distinct sub-tasks (executive summary, customer profile, positioning, timeline, channels, metrics, risks). Each sub-task has its own format and its own depth requirement. This is called task decomposition, and it is the single most reliable way to improve output quality on complex projects.

Here is why it works: a language model generates text token by token. When you ask for "a plan," the model has to decide on the fly what a plan includes. It will default to a generic structure. But when you specify the seven sections, the model allocates its attention (and its token budget) across those sections. It cannot spend 800 tokens on the executive summary and leave 50 for the risk register, because you have constrained the structure.

You can push this further by asking the model to work section by section in separate turns. For the launch plan above, you might run the first prompt to get the clarifying questions, answer them, get the full plan, and then say:

Now expand section 4 (the 8-week timeline). For each week, give me the specific tasks for the product marketing manager, the content designer, and the sales enablement lead. Use a table format with columns: Week, Task, Owner, Deliverable, Due Date.

This is a second-level decomposition. You are drilling into one sub-task and asking for a finer granularity. Each turn builds on the previous context, so the model remembers the answers to your clarifying questions and the structure it already produced.

Handling Truly Ambiguous Requests

Sometimes the ambiguity is not about missing facts — it is about conflicting goals. For example, you might ask for "a blog post that is both persuasive and neutral." Those goals fight each other. A professional prompt should surface that conflict explicitly.

Here is a pattern for that situation:

I need a 1,200-word blog post about the benefits of switching to cloud-based accounting. I have two conflicting requirements: it must be persuasive enough to drive demo sign-ups, but it must also be balanced enough to pass our legal review. 

Before writing, ask me:
1. Which requirement takes priority if they conflict?
2. Do you have specific customer testimonials I should include, or should I write placeholder quotes?
3. What is the primary call-to-action — a demo request, a whitepaper download, or a free trial?

Do not write until I answer.

This prompt does two things. It names the conflict explicitly, and it forces the model to ask which side wins. Without this, the model will produce a bland middle-ground piece that persuades no one and fails legal review anyway.

Step-by-Step Application: What You Literally Do

Here is the exact sequence to follow for any complex project, whether it is a business plan, a software architecture design, or a research synthesis.

1. Open a fresh conversation. Do not reuse an old chat — the context will contaminate the new task.
2. State the role. First sentence: "You are a [specific role] with [specific years of experience] in [specific domain]."
3. State the deliverable. Second paragraph: "I need [specific artifact] for [specific purpose]."
4. Impose the question rule. Third paragraph: "Before you produce anything, ask me at least [N] clarifying questions. Do not proceed until I answer."
5. Define the output structure. List the sections you want, in order, with word counts or format constraints.
6. Run it. Paste the prompt. The model will ask questions. Answer them in full sentences, providing real numbers and real constraints.
7. Iterate section by section. After the full output, pick the weakest section and ask for a deeper expansion with specific format requirements.

Here is a concrete example of step 6 and 7 in action. Suppose the model asks: "What is your current active user base?" You answer: "We have 2,400 active accounts, of which about 800 are on the enterprise plan. The feature is included in all plans, but we want to upsell the enterprise tier." Then, after the full plan is generated, you say: "Expand section 5 (channel strategy). For the top channel, give me a week-by-week content calendar with specific post titles and target platforms."

Common Mistakes

Practice Task: Build Your Own Clarifying Prompt

Your task is to write a prompt for a complex project that requires the model to ask for missing details. Choose one of these scenarios:

• Scenario A: You need a 6-month content marketing strategy for a niche B2B consultancy that specializes in helping logistics companies reduce fuel costs.
• Scenario B: You need a technical architecture proposal for migrating a legacy on-premise CRM (with 50,000 customer records) to a cloud-based system, with a budget of $80,000.
• Scenario C: You need a training curriculum for a 2-day workshop teaching non-technical managers how to use AI tools for data analysis.

Write a single prompt that includes all of the following elements:

1. A specific role with domain expertise.
2. A clear deliverable description.
3. A hard rule requiring at least 5 clarifying questions before any output.
4. A structured output format with 5–7 named sections.
5. A constraint that prevents generic advice (e.g., "name specific tools" or "use real industry benchmarks").

Self-verification: Paste your prompt into any major AI tool (ChatGPT, Claude, Gemini). The model should stop and ask you at least 5 questions. It should not produce any plan content. If it does produce content, your rule was not strong enough — revise it and try again. Then answer the questions with realistic details and confirm the final output follows your requested structure. This entire exercise should take under 15 minutes.