The workflow
Step 1 — Prep the transcripts (30 min)
- Export each Otter / Fathom / Granola transcript as plain text.
- Create one concatenated file with clear
## Interview 1 — Jane, Senior PMheaders between each. - Strip any PII you don''t want in context: last names, employer details, specific URLs.
Step 2 — Theme extraction pass (45 min)
Paste the full concatenated transcript into a Claude conversation. Prompt:
You are helping synthesize 15 user interview transcripts. Your job is to surface recurring themes, not to summarize each interview.
Read all 15 transcripts. Then return:
1. Top 5 recurring themes, ranked by frequency
2. For each theme: 2-3 direct quotes (with interview number attribution)
3. Counter-evidence: any interviews where the theme was notably absent or contradicted
4. Unresolved tensions: places where interviews disagree
Step 3 — Interrogate the output (60 min)
This is where it gets good. For each theme:
- "Show me every mention of [theme] across all 15 interviews, even the tangential ones."
- "Which interviewees never mentioned this? Why might that be?"
- "Rewrite this theme in 10 words or less, for a non-design audience."
You're not accepting Claude''s first pass — you''re using it as a research assistant that has read everything.
Step 4 — Draft the report (45 min)
Prompt Claude to generate a first draft with your preferred report structure (I use: problem → themes → tensions → recommendations). Edit heavily. Add your own interpretation. Claude is bad at the "so what" — that''s your job.
Quality check before shipping
- Every theme has a minimum of 3 quotes from different interviews
- No quote has been paraphrased (use exact transcript text)
- You''ve read the final report end to end and agree with it
State of Design 2026 data: 47% of designers we surveyed in Nov 2025 said they used an LLM for research synthesis in the last 30 days. The top complaint: "it makes up insights." Solution: always force quote attribution.