Question-Mode: The LLM as a Sparring Partner That Asks Questions Instead of Handing You Answers

STAGE 1 — FINDING ALTERNATIVES (Novelty Phase):
Action: the LLM asks questions + offers analogies from other fields
You: generate 2-3 alternative directions
You: pick one to develop

STAGE 2 — DEVELOPING THE IDEA (Elaboration Phase):
Action: the LLM asks targeted questions to make it concrete
You: develop the idea yourself by answering the questions
Result: a final, well-developed idea

Both stages happen in one conversation. The LLM guides you through the process but doesn't do the work for you.

Task: You're rebranding a local coffee brand called "Brewhaus". Right now its positioning is fuzzy — "quality coffee for everyone." You need a sharp, memorable idea.

Prompt:

Help me develop an idea through questions, not ready-made answers.

STAGE 1 — Find alternative directions:
My current idea: "Brewhaus — quality coffee for office workers"

Ask questions and offer analogies from OTHER fields (not coffee, not food)
that will help me see non-obvious directions.
Don't give me ready-made options — help ME come up with them.

After that I'll pick one direction and we'll move on to stage 2.

Result:

The LLM will ask 3-5 questions like: "Which brands from completely different categories remind you of the state of mind coffee gives you?", "If Brewhaus weren't a drink but a place, what would it be?", and offer an analogy: "Spotify for music = personal playlists for every mood. How could you carry that principle over to coffee?".

You come up with 2-3 directions yourself (for example: "coffee as a switching-gears ritual", "anti-stress coffee for the burned out", "coffee as a status symbol for people who never sleep"). Then you pick one.

In stage 2 the LLM asks: "Who exactly is your audience in this direction?", "What specific moment in their day do you want to own?", "What should change in how they see things after they interact with the brand?". You answer and develop the idea yourself.

The finish line: you have well-developed positioning that feels yours, because you built it yourself through the conversation.

The LLM's weakness: When a model generates ideas on its own, it averages. It produces whatever is statistically "right" based on its training data. Result: ideas that are solid but similar to each other. That's homogenization — everyone ends up thinking inside the same corridor the model drew.

The LLM's strength: The model is great at setting context, knows many domains, can find analogies in unexpected places, and can phrase questions that open up new angles.

How the method plays to that strength: Question-mode keeps you in the author's seat and the LLM in the role of a conversation partner who stretches your thinking. You generate the content yourself, drawing on your own context and expertise (which the LLM doesn't have), while the model highlights blind spots and sparks your thinking with cross-domain analogies.

Prompt control levers:

• Domains for analogies: "Only use analogies from [music/sports/architecture]" → you can point the LLM at the field you want inspiration from
• Question depth: "Ask only 1-2 questions" VS "Keep asking 5-7 questions until I land on an unexpected direction" → you control how deep the exploration goes
• Stage 2 focus: "In stage 2, focus on practical implementation" VS "focus on emotional impact" → you change the direction of the deep dive
• Question format: "Ask open-ended questions" VS "Ask questions with concrete examples" → you control the style of the conversation

If you drop the questions and keep only the analogies, you get suggestion-mode (which doesn't work). If you drop the two phases, you lose the structure. The "questions + analogies + two stages" structure is critical.

Help me develop an idea through questions, not ready-made solutions.

STAGE 1 — Finding alternative directions:
My current idea: {describe your starting idea in 1-2 sentences}

Your tasks at this stage:
- Ask questions that will help ME see alternative directions
- Offer analogies from {name the domains: "other business areas" / "technology" / "natural processes" / etc}
- DO NOT give ready-made options — stimulate my thinking

I'll generate 2-3 alternative directions and pick one.

---

STAGE 2 — Developing the chosen idea:
Once I've picked a direction, ask targeted questions to help me:
- Pin down {audience/mechanics/value/etc}
- Work out {practical implementation/uniqueness/etc}
- Find {constraints/opportunities/etc}

Your role is NOT to generate solutions but to guide my thinking through questions.

What to fill in:

• {describe your starting idea} — your current idea in 1-2 sentences
• {name the domains} — which fields to pull analogies from (the further from your topic, the more surprising the associations)
• {audience/mechanics/value} — what exactly needs working out in stage 2

🚀 Quick start — paste into the chat:

Here's a question-mode template for developing ideas through dialogue.
Adapt it to my task: [describe your task].

Ask me questions to fill in the fields and tune it to my context.

[paste the template above]

The LLM will ask about your idea, the focus you want, and the domains for analogies — because the method only works well once it understands your specifics. The model will take the "questions + analogies + two stages" pattern and adapt it to your task.

⚠️ Cognitive load: Question-mode takes more mental effort than just getting a ready answer. If the task is simple ("name the capital of France") or you're already fried, it's overkill. The method is for situations that call for deep development and originality.

⚠️ You have to be willing to think: If you want a fast result without thinking, question mode will get on your nerves. Model-led is faster, but it sacrifices diversity and your authorship.

⚠️ Not for every kind of task: The method proved effective on creative tasks (ideas, concepts, strategies). For purely mechanical tasks ("format this table", "fix the grammar") the structure is overkill.

⚠️ Depends on your expertise: The LLM's questions can stretch your thinking, but the final quality still depends on how well you know the context. In an unfamiliar field, the questions can lead you into a dead end.

The researchers recruited 486 people from the UK via Prolific and gave them a creative task: come up with a new use for car sensors (an adaptation of the Alternative Uses Test). Participants were randomly assigned to 5 groups: question-mode (the LLM asks questions + gives analogies), suggestion-mode (the LLM gives advice), model-led (the LLM rewrites the ideas itself), vanilla ChatGPT (a regular chat), and a control group (no AI).

Design logic: All the LLM groups first generated an idea on their own, then interacted with the model according to their scenario. That matters — the researchers were testing the effect of the interaction mode, not just "AI vs. no AI."

What they measured:

1. Idea quality — 4 automotive-industry experts rated originality and usefulness (the Consensual Assessment Technique, the gold standard for judging creativity)
2. Diversity — semantic distance between ideas, measured with embeddings (OpenAI text-embedding-3-large)
3. Perceived ownership — how much people felt the idea was "theirs" (a 5-item questionnaire)

Unexpected findings:

• Suggestion-mode flopped on almost every metric. You'd think giving advice = helping, but people either ignored the tips or applied them superficially without really working them through.
• Vanilla LLM (a regular chat) produced high idea diversity but didn't improve quality and lowered ownership. The paradox: no structure → a wide range, but no systematic improvement.
• Model-led delivered the highest quality, but at a price: ideas became 9% less diverse within the group, and people felt less like the authors.

The key practical insight: The study showed that homogenization is not an inevitable property of LLMs but a consequence of interaction design. Question-mode didn't just preserve diversity — it increased it by 7% over the course of the work (within-subject analysis). That means a well-designed dialogue can actively expand the idea space instead of shrinking it.

It was tested on GPT-4.1 via the API, but the principles are universal — they work with any sufficiently capable LLM.

Partnering with Generative AI: Experimental Evaluation of Human-Led and Model-Led Interaction in Human-AI Co-Creation
Sebastian Maier, Manuel Schneider, Stefan Feuerriegel
LMU Munich & Munich Center for Machine Learning (MCML)

💡 Adaptation for writing (an article, a pitch, a script):

Classic question-mode is built for ideas, but it adapts easily to text:

Help me develop [an article/a pitch/a script] through questions.

STAGE 1 — Find non-obvious angles:
My current thesis: {the main point of the text}

Ask questions that will help ME see:
- Alternative narrative structures
- Unexpected analogies from {other genres/formats}
- Weak spots in the logic

Don't suggest ready-made wording — stimulate my thinking.

STAGE 2 — Working out the details:
Once I've picked a direction, ask questions to make it concrete:
- Which examples will strengthen the thesis?
- Where might the reader "drop out" of the text?
- What objections do I need to anticipate?

Example: You're writing an article about burnout in tech. Instead of "rewrite this better," the LLM asks: "Which brands from completely different industries solved the burnout problem for their customers? How could their approach apply to tech workers?". You find an unexpected angle yourself (say, the fitness industry and the principle of "progressive overload"), and the LLM asks questions about how to weave it into the text.

🔧 Technique: Add a critic role → sharper development

The original question-mode leads to a solid deep dive, but it can be a bit "soft." If you need tough criticism of an idea:

STAGE 2 — Development through criticism:

Once I've picked a direction, step into the role of {a skeptic/an investor/a competitor}.

Ask questions FROM THE CRITIC'S POSITION:
- "Why won't this work in the real world?"
- "Where's the weak spot that will kill the idea?"
- "What's being overlooked?"

Help ME find and plug these holes myself, without giving me ready-made solutions.

Effect: The questions get sharper, and you work through not only the strengths but also protection against failure. Useful for business ideas, strategies, and high-stakes decisions.

💡 Extrapolation: Question-mode + Self-Refine for iterations

The combo: question-mode (for generation) + a self-refine loop (for improvement):

LOOP:

1. Question-mode → I generate an idea through your questions

2. You, as a critic, score the idea on [criteria]

3. If the score is <8/10 → a new round of questions to improve it

4. Repeat until we hit 8/10 or 3 loops

Use case: For tasks that need iterative polishing (an investor pitch, a brand's key message, a product value proposition). You don't just answer the questions once — you go through several rounds, deepening the work each time.