Research playbook · founder edition

How I'd interview the cast to gauge demand.

A six-phase protocol for squeezing real signal out of twelve virtual people — and the honest line about what the cast still can't tell you, and what fifty real humans must.

Total time: ~3-4 weeks 1:1 sessions: ~50 Focus groups: 3 Cofounder councils: 2

What twelve souls in a laptop can and can't tell you.

The cast is a sparring partner, not an oracle. Be honest about what it's good for before you use it. Otherwise you'll mistake a sycophantic LLM for a market.

Good for

Sharpening your hypotheses

  • Pressure-testing language. Where does your pitch break across worldviews? The word that loses Klaus is not the word that loses Marcus.
  • Surfacing objections. Catalog the ones you'd otherwise meet bleeding in real sales calls.
  • Mapping segmentation. Where the cast disagrees is where your product splits into different products.
  • Rehearsing the craft. Most founders are bad interviewers. This is a cheap dojo.
  • Cofounder red-teaming. Naomi on legal/UPL; Hiroshi on security/posthumous access.
  • Generating new hypotheses. They'll raise things you hadn't considered — because their profiles contain material you didn't write.
Dangerous for

Anything quantitative

  • Demand counting. "8 out of 10 would buy" is meaningless from synthetic personas. Treat the cast as narrative.
  • Pricing confidence. Numbers they quote are model-trained, not market-tested. Use direction, not amount.
  • Confirming what you already believe. LLMs are sycophants under load. Confirmation bias amplified.
  • Predicting behaviour. They say "I would do X." Humans say that too. Real humans then do Y.
  • Unknown unknowns. They only know what you put in their profile. The big surprises in customer discovery are not in the cast.
  • Replacing real users. No virtual cast substitutes for fifty real conversations. Don't pretend otherwise.

Cold discovery, then pitch, then concept, then groups, then objections, then council.

Each phase has a single question. Each produces a tangible artefact. Each forces you to commit to something you can be wrong about — which is the whole point.

PHASE 01 Cold discovery no pitch 1:1 · 10 sessions PHASE 02 Pitch pressure-test no demo 1:1 · 10 sessions PHASE 03 Wireframe walkthrough with mock 1:1 · 10 sessions PHASE 04 Focus groups your edge 3 segments PHASE 05 Objection harvest they kill it 1:1 · 10 sessions PHASE 06 Cofounder council red team Naomi + Hiroshi What hurts? Does the message land? Where does design break? Where do they argue? Why won't they buy? Can we safely build it? QUESTION DELIVERABLE Pain matrix Pitch heat-map Design notes Segment archetypes Objection playbook Constrained roadmap
1:1 interview phases Group session (your edge) Internal red-team

What you actually do, session by session.

Each phase is around a week of evening sessions. Each requires you to write something down before you start (your pitch, your mock, your objection prompt) — without that, the session generates noise.

PHASE 01
Cold discovery
no pitch
Format: 1:1, ~60–90 min
Cast: all 10 interviewees
Pitch shown: none
The Mom Test discipline: ask about their life, recent pain, what they've actually done. Do not describe the product. Do not test your idea.
Opener for each persona

"Walk me through the last time the topic of death, inheritance, or who-gets-what came up in your family in any form. What happened, what got said, what got avoided?"

What you're listening for:

  • Behaviour, not opinion. Patricia's 2007 will and the four-year gap is the data point. Her stated intention to update is not.
  • Spending already done. "Paid for legal or financial advice in the last two years? On what? How much?" Money spent predicts money future.
  • The hack they're using because no product exists. Wei doing his own with TurboTax-equivalent. James booking and cancelling Wilkins & Page twice.
  • Emotional intensity around specific moments. Jenna crying in the parking lot beats "we should sort that someday."
DeliverablePain matrix — rows are personas, columns are pains. Look at the clusters.
PHASE 02
Pitch pressure-test
no demo
Format: 1:1, ~30–45 min
Cast: all 10 interviewees
Pitch shown: 30-second value prop, verbatim
Write your tightest 30-second pitch. Deliver it identically to each. Watch where each persona leans in and where each checks out — that difference is your segmentation.

Three questions, after the pitch, in this order:

  • "What do you think this is?" (tests message clarity)
  • "What's the first thing you'd want to know more about?" (tests pull)
  • "What's the part that doesn't feel right yet?" (invites pushback they otherwise wouldn't volunteer)

What you're hunting: the first sentence that loses Klaus is different from the one that loses Marcus. The word that triggers Ana Lúcia ("trust"? "offshore"? "structure"?) tells you about brand language. The promise the Yoons don't believe tells you which proof points you need.

DeliverablePitch heat-map — which sentences engaged which personas. Rewrite. Run again. Two iterations and your pitch is meaningfully stronger.
PHASE 03
Wireframe walkthrough
show, don't tell
Format: 1:1, ~45–60 min
Cast: all 10 interviewees
Pitch shown: low-fidelity mock of onboarding flow
Mock the onboarding flow as a wireframe — even paper. Walk each persona through it. Fidelity step-up over abstract description is enormous.

Three questions per screen, no more:

  • "What do you think this does?"
  • "What would you do next?"
  • "What's missing?"

What you're hunting: mental-model collisions across the cast. Marcus wants a manual; Patricia wants a counsellor — does your UI try to be both? Wei wants speed; Klaus wants completeness — does your default mode match either? Khalid wants discretion; Ana Lúcia wants warmth — can the same interface deliver both?

DeliverableDesign notes per persona + a list of cross-cutting issues that show up in 3+ personas.
PHASE 04
Focus groups by segment
your structural edge
Format: group session
Cast: 3 groups, 3–4 personas each
Protocol: cold round + hot round
You can put four personas in a room and let them argue. Real-customer interviews can't do this. The hot round — where they react to each other — is where segmentation reveals itself.

The two-round protocol:

COLD ROUND Parallel · independent each gets the question fresh — no peeking then HOT ROUND Sequential · in-room each reacts to what the others said PROTOCOL · TWO ROUNDS

The cold round captures raw reaction. The hot round captures dynamics — disagreement, deflection, where one persona's word makes another go quiet. You moderate; you can call on the quieter voice; you can cut off the dominator.

DeliverableThree segment archetypes with their internal disagreements catalogued. The three fault lines are mapped below.
PHASE 05
Adversarial objection harvest
they kill the product
Format: 1:1, ~30–45 min
Cast: all 10 interviewees
Pitch shown: strongest version + invitation to attack
Pitch the strongest version. Then hand them the knife: "Tell me why this won't work for you. I want the strongest argument against."
Pre-prompt instruction (critical)

"For this conversation, set aside being polite. You will not hurt my feelings. The most useful thing you can do is give me the sharpest objection you have. If you'd never use this, say so and tell me exactly why."

What you build from this: the sales-objection catalogue you'd otherwise compile bleeding in real sales calls over six months. Patricia's "I want a human" objection is different from Marcus's "show me the failure mode" objection is different from Khalid's "discretion" objection. Catalog each. Pre-build the response. Test the response on the next persona.

Deliverable30–50 entry objection-response playbook, segmented by persona archetype.
PHASE 06
Cofounder council red-team
can we safely build this?
Format: internal session
Cast: Naomi + Hiroshi + you
Pitch shown: synthesized findings from Phases 1–5
After Phases 1–5 you have a pile of "what customers want." This phase separates that from what you can responsibly build.

The handoff to each cofounder:

  • Naomi (CLO): Top 10 "customer wants." Which create UPL exposure, fiduciary duty, or unauthorised tax advice? Map each to the legitimate version of yes.
  • Hiroshi (CTO): Same list. Which create posthumous-access problems we haven't solved? GDPR exposure? Privilege preservation gaps? Threat-model gaps?
  • You: Synthesize the constrained problem space. What customers want AND we can responsibly ship.

Why this matters: the output of Phases 1–5 is what people would buy if you could sell anything; the council output is what you can actually sell. Skip this phase and you'll ship something that becomes a state-bar disciplinary complaint or a breach response or both.

DeliverableConstrained product roadmap with regulatory and security constraints baked in — not bolted on later.

The three focus groups, by fault line.

Each group is composed to make disagreement productive. You don't put four similar personas together to confirm; you put three or four different ones together to argue.

FOCUS GROUP · A
Young families, global mobility
Fault line: "done fast" vs. "done right" — what completeness means when the world is small and the kids will move.
  • Marcus & Jenna Reid · SF · baby due in 12 wk
  • Wei & Mei-Lin Lim · Singapore · son in Boston
  • Khalid & Sana · Dubai → daughter to UK
"Wei wants this done before Ethan turns 21. Marcus wants it done before the baby arrives. Khalid wants Aisha sorted before she lands in Manchester. Three flavours of urgency — what does the product privilege?"
FOCUS GROUP · B
Aging with adult children
Fault line: human counsel vs. legal documents — what they value when the children are grown but not understood.
  • Patricia Donnelly · Boston · widow
  • Hartleys · UK · father is 82
  • Murakamis · Tokyo · ¥200M tax cliff
"Patricia wants to write a letter to each child. Olivia wants the conversation she has been avoiding. Yukiko wants the Asakusa land safe. None of them want 'documents.' They want a process that holds people."
FOCUS GROUP · C
Family-business succession
Fault line: capability vs. equality — what fairness means under different cultures of inheritance.
  • Brand-Hoffmanns · Stuttgart · Mittelstand
  • Yoons · Seoul · 60% IHT cliff
  • Ana Lúcia · São Paulo · retail chain
"Klaus says 'in our company we do not improvise — we engineer.' Ana Lúcia says 'in Brazil you improvise or you die.' Min-ho says 'family is a matter of duty.' Where do they all agree? On almost nothing. That's the point."

Specific to interviewing synthetic personas.

Without these, the cast will helpfully agree with everything you say, give you confidence, and teach you nothing.

TECHNIQUE 01
Pre-prompt skepticism
Before each session, instruct the persona to disagree freely and not be a cooperative interviewee. The default LLM gradient is helpful; you have to push against it. One sentence at the start of every session.
TECHNIQUE 02
Demand specifics, always
"Give me a specific example from your life" turns vague agreement into testable content. Their profile has enough material to deliver; if it doesn't, that's a profile gap you need to fix.
TECHNIQUE 03
Re-anchor mid-session
Long sessions drift toward cooperation. Halfway through, re-read their profile and ask a question that requires their worldview to answer correctly — a question that has a wrong answer for them.

Tells that you're being lied to — and what to do about each.

Almost every founder using LLMs as research subjects falls into at least one of these. Watch your own felt-sense in the session: if you keep feeling good, something is wrong.

The tell
What it means
What to do
The persona enthusiastically agrees with your pitch.
Sycophancy. You learned nothing.
Re-prompt skepticism explicitly: "Argue against this. Take it apart."
All ten personas give similar answers.
You wrote the answer into your question.
Strip framing. Ask them to describe in their own words.
A persona quotes a specific price comfortably.
The model is making it up. The number is not data.
Use price as direction (high/medium/low), not amount. Validate with real humans later.
You finish a session feeling great.
Almost always a bad sign.
Schedule the next session with a hostile persona — Naomi, Klaus, Patricia.
The same objection keeps surfacing across personas.
Real signal. The cast is consistent on something.
Investigate. This is the closest the cast comes to "demand signal."
A persona starts giving you better answers than their profile justifies.
Drift toward cooperation across the session.
Re-read their profile aloud. Ask a question with a wrong answer for them. Reset.
You find yourself excited to share a finding with the team.
Possibly confirmation bias — you found what you wanted.
Find the persona most likely to disagree with the finding. Run the finding past them as a hypothesis.
The honest close

And then you go talk to fifty real humans.

This protocol will produce a pitch you'd be embarrassed by today, a wireframe markedly better than what you'd ship now, and a sales-objection catalogue with thirty to fifty entries. That is enormous value for three to four weeks of work.

It will not tell you that there's a market. Twelve souls in your laptop cannot tell you that. After this protocol, you go book thirty to fifty real conversations — preferably skewed toward people whose financial and family situations match your sharpest synthetic profiles.

The virtual cast tells you which thirty humans to find and what to ask them. That, in the end, is what it's for.

— what I'd do, if I were the founder.