LLM prompts and methodology for turning customer interview transcripts into product roadmap recommendations. Point an agent at a folder of transcripts and get: What should we build next, and why?
Analyzes interview transcripts to produce:
- Frequency-ranked problems (X of N transcripts mentioned...)
- Frequency-ranked solution desires
- Verbatim quote evidence with attribution
- Concrete build recommendations
Explicitly avoids:
- Per-interview summaries
- Vague themes ("bad UX")
- Opinion-driven prioritization
- Mention-counting (counts transcripts, not mentions)
Analyze the interviews in ./transcripts/ using the methodology in PROMPT_core.md
The agent reads all transcripts, applies the methodology, and produces a report matching TEMPLATE_output.md.
| File | Purpose |
|---|---|
INTERVIEW_script.md |
20-minute interview script optimized for later analysis |
PROMPT_core.md |
Core methodology: ingest → extract → normalize → count → synthesize |
PROMPT_with_quotes.md |
Quote-heavy variant for stakeholder presentations |
TEMPLATE_output.md |
Expected output structure |
RUBRIC_scoring.md |
7-dimension quality rubric (1-5 scoring) |
EXAMPLE_usage.md |
Detailed usage guide |
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Conduct │ │ Analyze │ │ Build │
│ Interviews │ ──▶ │ Transcripts │ ──▶ │ Recommendations│
│ │ │ │ │ │
│ INTERVIEW_ │ │ PROMPT_core.md │ │ "What to build │
│ script.md │ │ + transcripts │ │ next and why" │
└─────────────────┘ └─────────────────┘ └─────────────────┘
INTERVIEW_script.md provides a tight 20-minute structure:
| Phase | Time | Focus |
|---|---|---|
| Open | 2 min | Setup, purpose frame |
| Context | 5 min | Background, current focus |
| Stack | 5 min | Tech choices, buy vs. build |
| Problems | 6 min | Challenges, gaps, workarounds |
| Magic Wand | 2 min | Aspirational solutions |
| Close | 2 min | Follow-up, referral ask |
Consistent interviews yield better cross-transcript pattern detection.
Note: This script is optimized for B2B and developer-facing interviews. The "Stack" phase and technical probing will need adaptation for B2C or non-technical contexts.
Interview selection matters more than interview volume. Key principles:
- Interview your best customers — those who intuitively see value, are profitable, and recommend you to peers. Not just anyone with a heartbeat.
- Pursue variation — across engagement levels (power users, new users, churned), roles, company sizes, and use cases. Variation beats "representative samples" in qualitative research.
- Match selection to goals — acquisition insights come from prospects; retention insights from current/churned customers.
- Recent switchers are gold — people who just started or stopped using a solution remember their decision context vividly.
- Ask for referrals — warm intros from interviewees have 80-90% success vs. 10% for cold outreach.
Common mistakes:
- Over-optimizing for the "perfect" sample (starting anywhere beats paralysis)
- Interviewing whoever responds instead of deliberately recruiting
- Mixing personas in the same analysis (CTOs and junior devs have different problems)
See: Teresa Torres on selecting customers, Customer Dev Labs B2B scripts
A good analysis lets someone:
- Write a feature spec from the top problem
- Defend any ranking with "X of N transcripts said..."
- Verify any claim by checking the cited source
Use RUBRIC_scoring.md to score outputs (target: 32+/35).