Turn Every User Call into Structured Research Notes

Workflow area: Discovery & Research · Tier: Foundation · Hands-on time: ~20 minutes

The bottleneck in continuous discovery was never talking to users — it was the hour of note cleanup after each call that nobody did. An AI notetaker plus a fixed extraction template collapses that hour to five minutes. The catch: raw AI summaries flatten what users actually said into what the model expects users to say. The template and the review pass exist to stop that.

When to use — and when not to

• Use it for every scheduled user or customer conversationdiscovery interviews, usability sessions, churned-customer exit calls, sales calls you join as a listener.
• Use it when notes feed a shared repotemplated extraction makes calls comparable across months and interviewers, which is the whole point of a research repository.
• Skip it for calls the participant hasn't consented to recordrecording consent is non-negotiable, and in many jurisdictions it is a legal requirement, not a courtesy.
• Skip it for sensitive escalationslegal disputes, security incidents, or an angry enterprise account where a recording bot joining the call reads as hostile. Take manual notes.
• Skip the extraction for casual hallway chatsa two-line Slack note is fine; not every conversation is research.

The recipe

1. 1.

   Set the notetaker to auto-join everything on your calendar. In Fireflies, Zoom AI Companion, or Granola, enable auto-join for meetings you own and configure the consent announcement so every participant hears that recording is on. Do this once; per-call setup is where the habit dies.

2. 2.

   Open your consent script and use it. Thirty seconds at the top of the call: "I have an AI notetaker on this call so I can focus on you instead of typing — the transcript stays inside our research team. Are you comfortable with that?" If they say no, turn it off and type. Never bury consent in a calendar invite footer.

3. 3.

   After the call, run the transcript through your extraction template. Paste the full transcript into Claude or ChatGPT with this prompt:

   Copy

   You are extracting research notes from a user interview transcript.
   Product context: [ONE LINE — e.g. "B2B expense-approval SaaS, buyer
   is a finance manager"]
   Interviewee: [ROLE, COMPANY SIZE, ANONYMIZED — e.g. "ops lead,
   ~200-person logistics firm"]
   
   Extract, using ONLY what the transcript supports:
   
   1. PAINS — each pain in one line, with the verbatim quote that
      evidences it and its timestamp.
   2. VERBATIM QUOTES — 5-8 quotes worth keeping, exactly as spoken,
      with timestamps. Do not paraphrase inside quotation marks.
   3. JOBS-TO-BE-DONE — situations and desired outcomes the
      interviewee described, phrased as "When... I want... so I can...".
   4. FOLLOW-UPS — questions we should have asked, commitments we made.
   5. SURPRISES — anything that contradicts what we currently believe.
   
   Rules: if a section has no evidence in the transcript, write "none
   found" — do not fill gaps. Never merge two statements into one quote.
   Flag any quote you are less than certain is verbatim with [CHECK].
   
   Transcript:
   [PASTE FULL TRANSCRIPT]
   

4. 4.

   Read the output against the transcript for two minutes. Good output looks like this:

   Copy

   PAINS
   - Approval chains stall when the approver is traveling — "I've had
     reimbursements sit for three weeks because Priya was in Singapore"
     (14:22)
   
   JOBS-TO-BE-DONE
   - When an approver is out of office, I want requests to reroute
     automatically, so I can close the books without chasing people.
   
   SURPRISES
   - Interviewee does NOT want mobile approvals — "if it's on my phone
     I'll approve without reading" (21:05) — contradicts our Q3 bet.
   

   Spot-check two or three quotes at their timestamps. Notetaker transcripts garble names and numbers, and extraction models occasionally "improve" a quote into something cleaner than what was said. Anything marked [CHECK], verify or delete.

5. 5.

   File the reviewed note in the repo, not the raw transcript. Tag it with persona, date, and product area. The transcript is the source of truth to link back to; the reviewed extraction is what teammates read and search.

Make it yours

• Claude Project as your standing extractor: create a Project called "Interview notes," put the template and your product context in the project instructions, and just paste transcripts. The context travels automatically, and the extraction stays consistent across every call you process.
• NotebookLM for the backlog of old calls: upload a batch of past transcripts as sources and query across them — "what did anyone say about approvals on mobile?" — with inline citations back to each source. Good for mining calls that happened before you had this habit.
• Team scale: agree the template as a team so every PM's notes have identical sections, then set up your notetaker's native "channels" or a Zapier step to drop each reviewed note into one Dovetail or Notion database. Comparable structure is what makes pb-02's cross-interview synthesis possible.

Quality check

• Every quote is verbatimspot-check at least two against their timestamps; a paraphrase inside quotation marks is a corrupted source.
• No section was padded"none found" appearing sometimes is a health signal; an extraction that always fills every section is inventing.
• The interviewee is anonymized correctlyrole and segment in the note, name and company only where your consent terms allow.
• Surprises made it throughthe model's strongest bias is toward coherence; if a call challenged your roadmap and the note doesn't say so, the note failed.
• The source link worksevery note links back to its transcript and recording so any claim can be re-verified in pb-02's citation audit.