Synthetic User Interview

Purpose

This skill simulates discovery interviews with generated personas. It is calibration, not research. Use it to sharpen your thinking before talking to real people, to catch bad questions before you ask them in a real call, or to explore segments you cannot reach in time.

The skill is designed to fail loud: every output ends with explicit warnings about what it can and cannot replace.

When to use

• You have a hypothesis and want to pressure-test it before scheduling real interviews
• Your interview guide needs a rehearsal and you want to catch leading questions
• Target segment is small, busy, or expensive to reach (CTOs, regulated industries)
• Strategic exploration of a buyer you don't yet have access to

When NOT to use

• For shipping decisions. Never ship based on synthetic interviews alone.
• As a substitute for ongoing customer development
• For sensitive segments where lived experience is the whole point (e.g. medical patients, marginalized groups). The model will produce plausible-sounding nonsense.
• For quantitative claims ("47% of CTOs say..."). The output is qualitative simulation, not a survey.

Inputs

Required:

• ICP description (industry, role, company stage, key context)
• Hypothesis or question to probe

Optional:

• Geography or regulatory context
• Specific user behaviors you want to explore
• Interview style (open-ended JTBD vs targeted feature feedback)

Process

Phase 1: Persona calibration

Generate 3-5 personas, each with:

• Name, role, company context (size, stage, industry)
• Current tools and workflow
• Top 3 pressures and goals this quarter
• Constraints (budget, time, political)
• Sources used to inform the persona (general industry knowledge, prior research, etc.)

Personas should differ meaningfully. Two personas that are basically the same with different names are wasted.

Phase 2: Build the interview script

• 5-8 open-ended questions following The Mom Test principles
• No leading questions, no pitching, focus on past behavior
• Mix: context, current workflow, pain points, attempted solutions, deal-killers
• Adjustable per persona based on their context

Phase 3: Run the interviews

For each persona, simulate a transcript:

• 8-12 question-answer pairs
• Realistic voice (interrupted, tangential, sometimes contradictory)
• Include details the persona would naturally bring up
• Mark explicitly where a real human would likely push back or refuse to answer

Phase 4: Cross-cut analysis

After all transcripts:

• Patterns: themes that appeared in 2+ interviews
• Contradictions: where personas disagreed and why
• Surprises: things you (the PM) probably didn't expect
• Hypothesis verdict: weakly supported / mixed / contradicted (with reasoning)
• Better questions for real interviews: what should you actually ask

Phase 5: Calibration warnings

Mandatory closing section:

• What this simulation likely got right (areas where public knowledge is strong)
• What it likely got wrong (anything requiring lived experience, current pricing, internal politics)
• What you should still go ask real users about
• Specific red flags: any persona answer that felt suspiciously articulate or "marketing-ready"

Output

# Synthetic Interview Report: [Hypothesis]

## Personas (3-5)
## Interview script
## Transcripts
## Cross-cut analysis
## Hypothesis verdict
## Calibration warnings - READ BEFORE ACTING

Common failure modes to avoid

• Persona = mouthpiece: the persona just agrees with the PM's hypothesis. Force at least one persona to be skeptical or hostile.
• Articulate user trap: real users are messy, hesitant, and contradictory. If transcripts sound like a TED talk, they're broken.
• Treating output as data: synthetic interviews never count as evidence in a PRD. Use them to sharpen real research, not replace it.
• Skipping calibration warnings: the warnings are the most useful part. Never omit them.

Why this skill exists

Synthetic interviews are controversial. The honest framing: they're useful for thinking, dangerous for deciding. This skill is structured to keep that line visible.