2.2 · Sharpen Your Question List with AI Personas

Module 2 · Step 2 of 4 · From Idea to First Paying Customer

Input: your draft Mom Test question list (5-8 questions from Ch 2.1) + 3 ICP characteristics (ICP = Ideal Customer Profile - the specific kind of person your hypothesis names, introduced in Ch 1.1)

Output: a sharpened question list (5-7 solid questions) + top 3 objections, ready to take into Ch 2.3 (a + b) recruitment and real interviews

Cost: $0 (free tier on Claude or ChatGPT)

Skip this if you’ve interviewed before. If you’ve run customer interviews in the past and your questions produced concrete past-tense answers, go straight to Ch 2.3a: Find 10 People. This chapter catches broken question shapes before they waste real interview slots - useful for first-timers, unnecessary if you’ve already calibrated your question technique.

TL;DR: A 90-minute AI rehearsal catches broken questions before you spend real interview slots on them. Claude personas expose hypothetical phrasing that generates polite yeses from anyone.

Already using Perplexity Pro / Claude / ChatGPT Deep Research for Module 1? The same tools work here for objection rehearsal - swap the deep-research prompts for the in-character persona prompts below; the cost line is already paid.

You drafted 5-8 Mom Test questions in Ch 2.1. Before you spend a real interview slot on a question that turns out to be pitch-shaped, run the question list past a Claude persona that matches your ICP. The persona answers in character; you ask Claude (out of character) which question generated which kind of answer and why.

The failure shape the rehearsal catches: a question like “Would you use a tool like this?” reads fine on paper, generates a polite “sounds great” from any persona, and absorbs five real interview slots before you notice the pattern. The rehearsal flags the question shape before you spend the slot. Same applies to questions that smuggle in your solution, ask for a hypothetical purchase, or bury the past-tense ask under three clauses.

The pattern: rehearsal tells you whether the question is broken; real interviews tell you whether the hypothesis is right. Catching a broken question with a free Claude session is cheaper than catching it on call 5 of 10.

Real interview slots are scarce. You only get about 10 a round, every one took outreach to book, and a hypothetical question burns the slot - the interviewee says “sure, I’d use it” to be kind, you hang up thinking you got a signal, and you got nothing usable.

An AI rehearsal costs nothing. A short pass through Claude before you pick up the phone, and you find out which questions collapse the moment a real human deflects them.

Real interviews stay irreplaceable for the things rehearsal cannot simulate: the noncommittal shrug on question three, the mention of a workaround you never imagined, the silence after Q4 that tells you more than ten polite yeses. The rehearsal sharpens your questions before you spend a real customer’s hour on them.

This chapter is the companion polish step between Ch 2.1 (where you learned the Mom Test technique and drafted 5-8 rough questions) and Ch 2.3 (a + b) (where you recruit 10 real interviewees). You don’t validate anything here - the real interviews do that. You catch the broken question shapes before they reach a real human - one focused rehearsal session saves 5 wasted interview slots. Here’s the rehearsal flow at a glance:

Why rehearse with AI at all

Think of it as a trial attorney prepping cross-examination with a paralegal standing in as the witness. The paralegal can’t give testimony, but the rehearsal finds the question that falls apart under any pressure at all.

Build 3 ICP personas in Claude

You don’t need a sophisticated prompt setup. You need to describe the person you’re trying to interview in terms Claude can work with, and then ask Claude to stay in character while you practice questions.

Here’s the persona setup prompt. The placeholder mapping table below tells you where each [BRACKETED] value comes from in your existing artifacts - read it AFTER you scan the prompt, not before.

Prompt 1 - Persona Setup

You are [PERSONA NAME], a [ROLE] at a [COMPANY SIZE] company in [INDUSTRY].

Your situation:
- You deal with [PROBLEM DESCRIPTION] about [FREQUENCY].
- You currently handle it by [CURRENT WORKAROUND].
- You've been doing this for [DURATION].
- Your budget authority for tools in this category is [BUDGET RANGE].
- You're skeptical of new software because [REASON - e.g. "the last three tools we bought sat unused"].

Stay in character for this entire conversation. Do not break character to explain your reasoning. Answer as [PERSONA NAME] would, not as an AI. If a question is vague, give the kind of vague polite answer a busy professional gives when they're not sure what you're asking.

Placeholder mapping - where each value comes from:

Fallback for the 3 fields not in your hypothesis ([FREQUENCY], [DURATION], [REASON]): make your best informed guess. The rehearsal is calibrated; the persona doesn’t have to be perfect. After your first real Ch 2.3 (a + b) interview, you will know whether your guess was too mild (“monthly” when reality is “daily”) or too aggressive. Revise BETWEEN interview 1 and interview 2. If any placeholder above is empty, the Ch 1.1 hypothesis is not specific enough - tighten it before rehearsing.

Heads up: Claude is trained to be helpful, which means it tends to give reasonable answers even to broken questions. Don’t read a coherent persona answer as proof the question works. Read Claude’s out-of-character diagnosis instead - the in-character answer reflects what Claude thinks a polite persona would say; the out-of-character note reflects what the question is actually asking.

Build 3 distinct personas before you start - not 3 variations of the same person. If your ICP is “ops managers at B2B SaaS companies,” your three personas might be: one at a 10-person seed startup (different budget, different urgency), one at a 60-person Series A (different process maturity), and one at a 200-person growth-stage company (different buying committee).

Each persona has different objections, different workarounds, and different reasons to care. A question that works cleanly on one persona will collapse on another - and that collapse tells you something before you spend real calendar slots finding out.

Run the rehearsal session

Once the persona is set, run your draft questions. Here’s the sequence.

Prompt 2 - Opening question test

I'm going to ask you a question about [PROBLEM AREA]. Answer as [PERSONA NAME] would if a stranger asked you this at a conference. Question: "[YOUR DRAFT QUESTION 1]"

After each answer, ask this diagnostic.

Prompt 3 - Question diagnosis

Now break character for 30 seconds. As an AI assessing that question: Was that a question that would produce useful data in a real customer interview? What would a real busy [ROLE] do with that question that I wouldn't predict from your answer? What's the version of that question that would make you open a real memory instead of giving a general response?

Repeat for each question. The in-character answer is plausible by default; the out-of-character diagnosis is where you learn whether the question would actually produce useful data on a real call.

Prompt 4 - Objection surface

Back in character as [PERSONA NAME]: What would make you want to end this conversation early? What question would feel like a pitch in disguise? What would make you worry I'm about to sell you something?

This prompt surfaces the three objections you need to test (not discover) in real interviews. An interviewee who feels sold to shuts down. If your question list has any question that a persona reads as a pitch, it will also read as a pitch to a real human - and real humans stop giving real answers the moment they sense a sales agenda.

Prompt 5 - Sharpening

Rewrite [QUESTION X] so that it anchors in a specific past event rather than a hypothetical. Keep it under 20 words. Give me three versions.

Use this on every question that produced a vague answer in Prompt 2. Concrete past-anchored questions are the whole point of the Mom Test. If you can’t make the question past-anchored in 20 words, the question isn’t ready.

Run all three personas against the same question list before you judge any single question. A question that holds up across all three is probably solid. A question that collapses on persona 2 but not persona 3 tells you something about the ICP segment, not just the question.

What to keep, what to throw out

After the rehearsal, you’ll have a mix of responses. Use these three signals to sort:

Judgment is still yours. The diagnostic only tells you which questions are obviously broken before you find out the expensive way. A question with a plausible in-character answer but a “this is hypothetical” out-of-character flag still gets cut - a coherent answer to a hypothetical question tells you nothing about real customer behavior.

What changed in your real interview slate

After the rehearsal, you have two deliverables.

The sharpened question list. Take your original questions, apply the revisions from Prompt 5, cut the ones flagged in Prompt 4. You should end the session with 5-7 solid questions where you started with 8-12 loose ones. That’s the list you take into booking real interviews with the full outreach stack.

The top 3 objections to test in real interviews. Prompt 4 will surface 3-5 things that make your persona want to end the conversation. Pick the 3 that appeared across at least 2 of your 3 personas. These are the objections you’re listening for in real interviews - not discovering them for the first time, but noticing whether and how they show up. There’s a difference between a real customer who raises objection #2 early (strong signal that the objection is real) and one who never raises it at all (either it’s not real for this person, or your questions didn’t give them space to surface it).

Objection Tracker - fill this in after the rehearsal, before your first real interview:

Print it. Put it next to the Mom Test interview script on your second monitor. After each real interview, tick the column. By interview 5, you’ll know which objections are real and which were just AI pattern-matching.

What to do next

Cost: $0 (the free tier on Claude or ChatGPT covers a focused rehearsal session; no paid plan required).

Failure signal: if all 3 personas produce nearly-identical answers to your questions, your persona descriptions are too generic - rewrite them with sharper role / company-size / urgency differences before continuing. The rehearsal only works when the 3 personas are genuinely different people with different objections.

The two deliverables: a sharpened question list (5-7 solid questions) and an Objection Tracker (top 3 objections to listen for in real interviews).

Going further

Reuse the rehearsal stack when a round of real interviews ends in partial signal.

One constraint worth naming: the rehearsal only surfaces signal that’s already in your mental model of the customer. Claude constructs the persona from what you tell it.

If your ICP description is wrong - the wrong role, the wrong company size, the wrong industry detail - the persona will be wrong in the same direction, and the rehearsal will give you false confidence.

This is the other reason real interviews stay irreplaceable: a real customer can tell you your ICP description is wrong, while Claude can only simulate the ICP you described.

Module 2 AI critic/simulator block - This chapter IS the block.

What AI can help with at this stage:

• Simulate 3 ICP personas answering your draft Mom Test questions in-character
• Flag hypothetical questions that generate polite-yes responses from anyone
• Surface the top 3 objections to listen for in real interviews
• Sharpen vague questions into past-anchored versions under 20 words

What AI cannot prove or substitute:

• Whether your ICP description is accurate (only a real interviewee can correct you)
• What a real customer will actually say (Claude simulates the persona YOU describe)
• Whether the problem is real (only 10 Mom Test interviews can falsify the hypothesis)

The real gate: 10 Mom Test interviews with real humans, ≥7/10 strong signal (Ch 2.1 technique + Ch 2.3 (a + b) recruitment).

Advanced: AI ensemble stress-test (after your interviews). Once your 10 Mom Test interviews are done and you have a refined hypothesis, you can cross-validate the business logic using multiple AI models simultaneously. Paste your validated problem statement into IdeaProof (70 free credits, no credit card) - it runs your hypothesis through 4 different models (Claude 4, GPT-4.1, Gemini 3, Grok 4.1) and flags contradictions between them. A claim that passes one model but fails another is a blind spot worth investigating before you build. The ensemble approach catches what a single-model rehearsal misses: each model has different training biases, and consensus across four is stronger signal than one model saying “sounds good.” This is not a substitute for the Mom Test interviews - it validates the logic AFTER the interviews validated the problem. Think of it as the final sanity check before you commit to building.

The rehearsal does not validate the hypothesis. It validates that your questions are ready to validate the hypothesis. Skip it and you burn real interview slots on questions that fail in minute one.

Further reading

• Rob Fitzpatrick, The Mom Test (book site) - the canonical reference for past-anchored interview questions.
• Y Combinator, How to Talk to Users - YC’s short essay on why this conversation has to happen.
• Anthropic, Claude prompting guide - persona setup, role-play, and breaking character cleanly.

Done when: You have a sharpened question list (5-7 solid questions) and an Objection Tracker with the top 3 objections to listen for.

Next click: 2.3a · Find 10 People: Where to Look - build the 30-name list first, then 2.3b · What to Say sends the messages.

If blocked: If all 3 personas produced identical answers, your persona descriptions are too generic. Rewrite them with sharper role, company-size, and urgency differences before continuing.

Case Study: Tomas & Mia

Tomas: Runs his draft questions through a Claude persona - a skeptical controller who’s been pitched 3 automation tools and rejected all of them. Persona flags 2 leading questions. Sharpens them to anchor in specific past reconciliation events.

Mia: Runs her draft questions through a Claude persona - a parent of a 10-year-old with ADHD burned by a tutoring app before. Persona flags 1 question that assumes the parent has time to search. Adds: “What happened the last time you tried to book a tutor during a workday?”

Built by JetThoughts as part of the From Idea to First Paying Customer curriculum.