The Startup Research Prompt Library

A startup research prompt library is a set of reusable, copy-paste prompts that turn a language model into a research assistant for validation work: generating interview questions, extracting jobs-to-be-done, designing unbiased surveys, sharpening positioning, and tearing down competitors. Each prompt below ships with variables to fill and a check to run on its output before you trust it.

Quick Answer: Use AI to structure and accelerate research, never to invent findings. Give each prompt real context — your customer, their situation, a raw transcript — demand specifics over opinions, and verify every output against evidence before you act on it.

The prompts here are copy-paste starting points, not oracles. A model will happily generate a confident-sounding market analysis from nothing but your excitement, and that output is worthless. The value comes from feeding these prompts real material and treating what comes back as a first draft you interrogate. Read the leading-answers section first — it's the difference between a research assistant and a very articulate yes-man.

How to write a research prompt that avoids leading answers

Write research prompts the way a disciplined interviewer talks: anchor on the customer's actual behavior and past, not on whether your idea is good. A model mirrors your framing. Ask "is my idea for X a good one?" and it will find reasons to say yes, because that's the shape of the request you handed it.

This is the same failure mode Rob Fitzpatrick describes in The Mom Test: people — and models trained on how people talk — will tell you what you want to hear if you let them. His fix maps almost directly onto prompt design. Talk about the customer's life, not your idea. Ask about specifics from the past, not opinions about the future. And make the model do the listening work instead of the flattering.

A non-leading research prompt does four things:

• Anchors on the customer and their situation, never on your solution, so the model reasons about a real person's world.
• Asks for specific past behavior and evidence, not hypothetical "would you use this" opinions that predict nothing.
• Explicitly requests the strongest counter-case — the disconfirming answer, the objection, the reason this fails.
• Supplies raw material — a transcript, a landing page, a competitor's site — so the output is grounded in something, not generated from vibes.

Here's the difference in practice. A leading prompt reads: "Write interview questions to confirm that busy founders need a faster way to validate ideas." The model obliges, and every question smuggles the conclusion inside it. A non-leading version reads: "Write questions that uncover how founders currently decide whether an idea is worth pursuing, based on what they actually did last time." Same topic, opposite epistemics — the first hunts for agreement, the second hunts for evidence.

Two structural habits make the rest of this library work. Give the model a role and a constraint set, because "act as a skeptical positioning strategist, be direct" produces sharper output than an unframed request. And demand a labeled separation between what the model can verify and what it's inferring, so you never mistake a confident guess for a finding. Every prompt below bakes in at least one of these.

The last point matters most. A prompt with no attached evidence produces generic advice every time, because there's nothing specific for the model to reason over. These prompts are tools inside a larger process, not a substitute for it; if you don't have that process yet, start with the complete guide to startup idea validation and slot these in where they fit.

Prompt 1: Generate non-leading problem-interview questions

This prompt converts a fuzzy problem hunch into a set of interview questions that dig into past behavior instead of fishing for validation. It never mentions your solution, and it forces you to name what a disconfirming answer sounds like — so you can't unconsciously ignore one in the room.

You are a customer-research assistant trained on the principles of
The Mom Test. I am investigating this problem:

Problem hypothesis: [describe the problem in one sentence]
Target customer: [role, context, company size / life situation]
What I assume they currently do about it: [current workaround]

Generate 12 interview questions that:
- Ask about specific past behavior ("tell me about the last time...")
- Never mention or hint at my solution
- Avoid yes/no and hypothetical "would you" phrasing
- Surface how they solve this today, what it costs them, and what
  they have already tried and abandoned

For each question, add a one-line note describing what a
DISCONFIRMING answer would sound like — an answer suggesting the
problem is not real or not painful enough to act on.

Fill the three variables carefully; a vague target customer produces vague questions. Verify the output by scanning for any question that leaks your solution or asks about the future — delete those. The disconfirming-answer notes are the point: if you can't imagine hearing them, your questions are still leading. Treat the first batch as raw material, then iterate: ask the model to rank the questions by how likely each is to produce a concrete story rather than a generality, and keep the top handful. For a standalone version with worked examples, see the problem-interview question template.

Prompt 2: Extract jobs-to-be-done from interview transcripts

This prompt turns raw interview transcripts into structured job statements you can compare across conversations. It's an extraction task, not a summary task — the distinction matters, because summaries smooth over the specific quotes that carry the real signal.

You are analyzing a customer interview transcript to extract
jobs-to-be-done. Do not summarize; extract structured signal.

Transcript:
"""
[paste the raw transcript]
"""

Output a table with these columns:
- Job statement (verb + object + context; functional, social,
  or emotional)
- Trigger (the situation that set the job off)
- Current solution / workaround they described
- Stated frustration or cost (quote them verbatim)
- Evidence quality (actual past behavior, or an opinion about
  the future?)

Flag any job you INFERRED rather than heard directly. Do not
invent jobs that are not supported by a direct quote.

The verbatim-quote and evidence-quality columns are your fraud detection. A job with no supporting quote is a hallucination — cut it. Run the prompt one transcript at a time rather than pasting several at once; models blur details when the input gets long, and you want each job tied to the exact conversation it came from. When you run it across several interviews, the same job statements will start repeating, and that repetition is your first real signal of a pattern. The JTBD extraction template goes deeper on how to phrase job statements so they stay comparable.

Prompt 3: Design a survey that doesn't bias its own answers

This prompt designs a survey to test a problem rather than sell a solution, screening out the biased question shapes that quietly guarantee the result you wanted. Most founder surveys fail here: they ask "how important is this to you?" and everyone politely says "very," which tells you nothing.

Help me design a [N]-question survey to test this problem, not to
sell my solution.

Problem hypothesis: [...]
Audience: [...]
Decision this survey will inform: [what I will do differently based
on the result]

Constraints:
- No double-barreled questions (one idea per question)
- No leading language or loaded adjectives
- Prefer behavior and frequency questions over "how important is..."
- Include one screening question that filters out non-target
  respondents

For each question, state which assumption it tests and what result
would make me STOP pursuing this idea.

The "decision this survey will inform" line does the heavy lifting. A survey that can't change your behavior isn't research; it's theater. Verify the output by checking that every question maps to an assumption and that at least one plausible result would genuinely kill the idea. If nothing can, rewrite until something can.

Prompt 4: Sharpen positioning and stress-test the value proposition

This prompt rewrites your value proposition in customer language and then attacks it, so you find the weak claims before a skeptical buyer does. It's deliberately adversarial — you're asking the model to play the person you're least prepared to face.

Act as a skeptical positioning strategist. Here is my draft:

Product: [one sentence]
Target customer: [...]
Main alternative they use today: [...]
Claimed differentiation: [...]

Do three things:
1. Rewrite my value proposition in the customer's own language,
   focused on the job they are hiring the product to do.
2. List the three strongest objections a skeptical buyer would raise.
3. Name the single competitor or alternative I am most likely
   underestimating, and explain why.

Be direct. If the differentiation is weak or generic, say so, and
explain what evidence would change your assessment.

The instruction to "name what evidence would change your assessment" keeps the model honest rather than merely critical. Weak differentiation usually hides behind adjectives — "faster," "smarter," "easier." If the rewritten value proposition still leans on adjectives instead of a specific job and a specific alternative, your positioning isn't ready, and no amount of prompting will fake it into being ready. A useful follow-up is to paste the three objections back and ask the model which one, if left unanswered, would most likely lose the deal — that forces a priority order instead of a flat list.

Prompt 5: Run a structured competitor teardown

This prompt produces a founder's-eye teardown of a competitor — who they target, the job their messaging emphasizes, and the gaps a new entrant could exploit. Its key discipline is forcing the model to separate verifiable facts from inference, because a confident-sounding guess about a rival's strategy is a good way to build the wrong thing.

Do a structured teardown of this competitor from a founder's
perspective.

Competitor: [name / URL]
My product: [one sentence]
The customer we both want: [...]

Cover:
- Who they clearly target (and who they appear to ignore)
- The core job their messaging emphasizes
- Their apparent pricing and packaging logic
- Two gaps a new entrant could realistically exploit
- Two things they do well enough that competing head-on is a mistake

Separate what you can verify from public information from what you
are inferring. Label every inference clearly.

The "two things they do well" requirement stops you from wishful thinking. Every founder can list a rival's flaws; the useful teardown names where the rival is strong enough that you should route around them. If you can attach a competitor's real URL or paste its pricing page into the prompt, do — grounded input turns inference into analysis.

Prompt library at a glance: task, goal, and what to verify

The table below maps each prompt to its purpose and — critically — the single thing you must check in the output before trusting it. The verification column is not optional; it's the step that separates research from generated confidence.

The pattern across every row is the same: the prompt structures the thinking, and you supply the judgment about what's real. Skip the verification column and you've automated the production of plausible nonsense.

Common prompt mistakes that produce generic advice

The reason a research prompt returns bland, useless output is almost always the input, not the model. Generic in, generic out. When founders complain that AI "just tells them what they already know," it's usually one of a handful of avoidable mistakes.

Watch for these:

• No attached evidence. Asking for market analysis with no transcript, no landing page, and no data gives the model nothing to reason over, so it pattern-matches to platitudes.
• Leading framing. Prompts built around "why my idea is great" get you a defense attorney. Prompts built around "the strongest case against this" get you a research partner.
• Opinion questions instead of behavior questions. "Would you use this?" predicts nothing. "Tell me about the last time you dealt with this" surfaces real behavior.
• Treating the first answer as final. The best output usually comes from a follow-up: "Which of these claims is weakest, and what evidence would overturn it?"
• Skipping the verification step. An unverified extraction is a rumor with good formatting.

There's a deeper point here worth sitting with: AI shortens the research loop, it does not shorten the thinking, and if you're using it to skip the customer conversations entirely you've missed the tool's actual leverage — for where it genuinely fits across the workflow, see how AI tools speed up validation. The founders who get the most from this library are the ones already doing the work; the prompts just remove the busywork around it.

A subtler mistake is running each prompt exactly once and stopping. The first response is a draft, not a verdict. The highest-leverage move is almost always a single sharp follow-up — "which of these findings is weakest, and what would overturn it?" or "rewrite this assuming the customer has no budget for a new tool." Iterating twice on one good prompt beats firing off five shallow ones, because you're pushing the model past its first, most generic instinct and toward the specific edge cases where the real insight lives.

One more failure mode deserves its own line: outsourcing the decision. The model can generate questions, extract jobs, and stress-test positioning, but it cannot decide whether the evidence is strong enough to build. That call is yours, and it should stay yours. In CB Insights' widely cited analysis of startup post-mortems, "no market need" is consistently the top reason companies fail — cited in roughly 35% of cases. No prompt library fixes that; only listening to real customers and acting on what you hear does.

Key Takeaways

• AI structures research; it does not replace it. Every prompt in this library is a first draft you interrogate, not a finding you accept — the judgment stays with you.
• Leading prompts produce flattering garbage. Frame requests around the customer's past behavior and the strongest counter-case, the same discipline The Mom Test teaches for interviews.
• Grounded input beats clever wording. A prompt with a real transcript, landing page, or competitor URL attached will always outperform an eloquent prompt with nothing to reason over.
• Every output needs a verification step. For each prompt, know the one thing to check — an unsupported job statement or a leaked-solution question is a defect, not a result.
• Behavior questions beat opinion questions. "Tell me about the last time" surfaces evidence; "would you use this" surfaces politeness.
• The decision to build is never the model's. AI can extract and stress-test signal, but only you can decide whether the evidence clears the bar — and "no market need" remains the leading reason startups fail.

Frequently Asked Questions

What is a startup research prompt library?

A startup research prompt library is a curated set of reusable, copy-paste prompts that turn an AI model into a research assistant for validation tasks — writing interview questions, extracting jobs-to-be-done, designing surveys, sharpening positioning, and analyzing competitors. Each prompt includes variables you fill with your own context and a specific check to run on the output before you act on it.

Can I trust AI to validate my startup idea?

No — AI can accelerate validation, but it cannot validate an idea for you. A model will generate confident-sounding market analysis from nothing but your enthusiasm, and that output has no evidentiary value. Use these prompts to structure interviews, extract signal from transcripts, and stress-test positioning, then verify every result against real customer behavior. The validation comes from evidence, not from the model.

How do I stop AI from just agreeing with my idea?

Stop framing prompts around your solution. Instead of "is my idea good," ask the model to generate the strongest case against it, surface disconfirming evidence, and name what result would make you abandon the idea. Anchor prompts on the customer's past behavior rather than opinions about the future. This mirrors The Mom Test principle: never ask people to validate your idea; ask about their actual life.

What context should I give a research prompt?

Give it real, specific material: the problem in one sentence, a precisely defined target customer, and — most importantly — raw evidence like an interview transcript, a landing page, or a competitor's pricing page. A prompt with no attached evidence produces generic advice every time, because there is nothing concrete for the model to reason over. Specificity in the input is what makes the output useful.

Are these prompts a replacement for talking to customers?

No. These prompts help you prepare for customer conversations and process what you learn from them, but they cannot replace the conversations. The most valuable inputs — real interview transcripts, observed behavior, actual objections — come only from talking to people. AI shortens the loop around research; it does not remove the need to do research. Founders who skip the conversations get eloquent, empty output.