| name | customer-interview-synthesizer |
| description | Two-phase customer discovery tool. Phase A: generate a Mom Test-compliant interview script from a hypothesis and target persona. Phase B: synthesize 1-N interview transcripts into pain points, JTBD, recurring quotes, and recommended pivots. Triggers on: interview script, customer discovery, synthesize interviews, jobs to be done, JTBD, user research, validate hypothesis. |
| triggers | ["interview script","customer discovery","synthesize interviews","jobs to be done","JTBD","user research","validate hypothesis"] |
customer-interview-synthesizer
Two-phase customer discovery skill. Use it before building any feature,
after forming a hypothesis, or when you need to understand why customers
behave the way they do.
When to use this skill
- Pre-PMF: before committing engineering time to a feature.
- After any pivot hypothesis — validate the new direction cheaply.
- When onboarding a new target persona you haven't interviewed before.
- After reading a batch of support tickets and wanting to find the JTBD underneath.
When NOT to use this skill
- Post-PMF scaling: at scale, surveys and product analytics give faster signal.
- Quantitative research with N > 100: use structured surveys (Typeform, Qualtrics).
- B2B enterprise deals: use SPIN Selling or Challenger Sale methodology instead —
those prospects expect a different interaction style.
Phase A — Interview Guide Generator
Inputs:
hypothesis: one-sentence belief you're testing (e.g., "Engineering managers at
Series A startups spend >4h/week manually aggregating deployment metrics").
persona: role, company stage, industry, typical daily workflow.
Outputs: 5–15 interview questions following Mom Test rules:
- Focus on past behavior, not future intent.
- Ask about specific incidents, not generalities.
- Never pitch your solution; probe the problem.
Question quality heuristics:
| Bad question | Why bad | Good alternative |
|---|
| "Would you use a product that…?" | Asks about future intent | "Tell me about the last time you…" |
| "Don't you think X is a problem?" | Leading | "How do you currently handle X?" |
| "How much would you pay for…?" | Premature pricing | "What do you currently spend on this?" |
| "Is this a big pain for you?" | Invites validation bias | "Walk me through last week's X workflow" |
See references/mom-test-rules.md for the full six-rule reference.
Example
Phase A — Sample Output
Hypothesis: Engineering managers spend >4h/week manually aggregating
deployment metrics.
Persona: EM at a Series A startup, 5–15 engineers, using GitHub + Datadog.
Generated questions:
- Walk me through how you prepared your last engineering update for leadership.
- What data did you need that you couldn't get in under 5 minutes?
- How long did that process take? What was the biggest time sink?
- What would have happened if you'd skipped that update?
- Have you tried to automate any part of it? What happened?
Phase B — Transcript Synthesis
Inputs: 1–N interview transcript blobs (freeform text, any format).
Outputs: Structured synthesis report using references/synthesis-template.md:
- Pain Points (ranked by frequency across transcripts)
- JTBD Statements in the format: "When I [situation], I want to [motivation],
so I can [outcome]."
- Recurring Quotes (verbatim — the exact words that capture a pattern)
- Recommended Pivots (one per major theme: what to build, cut, or change)
- Persona Refinement (how the real respondents differed from the target persona)
Phase B — Sample Output
Input: 3 interview transcripts with engineering managers.
Pain Points (by frequency):
- Manual Slack export + spreadsheet aggregation — mentioned in 3/3 interviews
- No single source of truth for deployment frequency — 2/3
- Metrics definitions differ between engineering and product — 2/3
JTBD Statements:
- When I prepare a weekly engineering update, I want reliable deployment metrics
in under 10 minutes, so I can spend the meeting on decisions, not data gathering.
Recurring Quotes:
- "I copy-paste from three tabs every Friday. It's embarrassing."
- "The dashboard exists but nobody trusts the numbers."
Recommended Pivots:
- Build the Friday digest as a Slack bot, not a dashboard — EMs already live in Slack.
Persona Refinement:
- Target persona said "Series A." Reality: all three were Series B. Adjust ICP.
Outputs
Files written to docs/discovery/<project>/ when WALTER_CURRENT_PROJECT is set:
interview-guide.md — Phase A output
synthesis-report.md — Phase B output
If the project variable is not set, output is written to the chat only.
How it composes with other Walter-OS skills
- Upstream of
pricing-experiment: synthesis reveals willingness-to-pay signals;
hand the synthesis report to pricing-experiment as the ICP input.
- Upstream of
vc-evaluator: a completed synthesis report strengthens the
customer evidence section of a pitch deck.
- With
weekly-review-coach: each synthesis session counts as a KR checkpoint
("completed N customer interviews this week").
- With
architect: use the synthesis output to evaluate whether a proposed
feature addresses a real, frequent pain point before committing to a spec.