Interview synthesis

Interview synthesis

Invocation: user only. Produces a synthesis artifact — doesn't publish without approval.

Takes N user interviews and produces a doc that turns raw listening into signal. A good synthesis surfaces patterns the PM didn't already see and grounds them in specific quotes.

Principles

• Patterns, not anecdotes. A theme needs 2+ independent mentions. One user saying something vivid is a quote, not a theme.
• Quote liberally. The user's actual words beat your paraphrase. Include short direct quotes for every theme.
• Contradictions are signal. If users split on something, name the split — don't average it away.
• Surface-area matters. An observation from one segment isn't generalizable. Note who said what when the segment differs.
• Themes are descriptive, not directive. "Users struggle with onboarding" is a theme. "We should redesign onboarding" is a recommendation — goes at the end.

Workflow

1. Collect the interviews

From $ARGUMENTS, identify the set:

• Explicit list of Fathom URLs or meeting IDs
• "last N interviews about X" — use Fathom MCP search
• Or a batch the user pastes / links

If Fathom MCP isn't connected, or the user points at external transcripts, ask for links or content.

2. Pull transcripts and summaries

For each interview via Fathom MCP:

• Attendees + participant role (customer, prospect, user segment)
• Key quotes flagged by Fathom's extraction
• Full transcript — pull but don't re-inject into main context; scan for additional quotes

If the batch is large (>8 interviews), do this in a forked subagent context to avoid flooding.

3. Extract observations

Per interview, extract:

• What problem they described — their own words
• What they're currently doing to solve it — workarounds, existing tools
• What they liked / disliked in the product or concept
• Asks — explicit requests
• Surprises — things they said that contradicted assumptions
• Behavioral signals — did they lean forward? go quiet? change tone? (Fathom sometimes notes these; if not, skip rather than fabricate)

4. Cluster into themes

Group observations that recur. A theme needs:

• 2+ interviews supporting it
• A clean one-sentence articulation
• 2-3 supporting direct quotes

Also track:

• Singletons — interesting-but-lonely observations worth noting
• Contradictions — where users split

5. Segment the patterns

If the interview set crosses segments (e.g. new users vs power users, SMB vs enterprise), break themes by segment where it matters. Not every theme needs segmentation — only when the pattern actually differs.

6. Output

# Interview synthesis: <topic or batch name>

**Interviews:** N  ·  **Window:** <date range>  ·  **Segments:** <if relevant>

## TL;DR
<3 sentences. The single biggest pattern and the single biggest surprise.>

## Themes

### Theme 1: <one-sentence articulation>
**Evidence:** N interviews
**Quotes:**
> "<direct quote>" — <participant role, interview date>

> "<direct quote>" — <participant role, interview date>

**What it suggests:** <1-2 sentences, descriptive not prescriptive>

### Theme 2: ...

(Aim for 3-7 themes. If you have more, they're not clustered tightly enough.)

## Contradictions
- **<dimension>**: Some users said <X>, others said <Y>. The split appears to track <segment / usage pattern / etc.>.

## Surprises
- <singleton observation, interesting, worth noting>

## Unchanged assumptions
<Assumptions we walked in with that the interviews did NOT change. Worth stating — it tells the reader which of the PM's existing hypotheses survived.>

## What the interviews do NOT tell us
<Explicit. "These interviews were with existing users, so they tell us nothing about why non-users don't sign up." Honest scoping keeps the synthesis from being over-applied.>

## Recommendations (optional section — only if the user asks)
1. <specific, grounded in themes>
2. ...

7. Publish

Show the synthesis. Common edits:

• A theme needs to be split or merged
• A contradiction is actually a segment difference
• A recommendation should be cut if the user wants to decide later

On approval:

• Save to Notion under the research path (default from CLAUDE.md; ask if unset)
• Offer to create follow-ups: feature-triage entries, a Notion tag cluster, a Slack summary for the team

Don't

• Don't manufacture themes to hit a round number. Three real themes beat seven thin ones.
• Don't average out contradictions. Name them and segment if possible.
• Don't include quotes without attribution. The quote is only useful if we know who said it and when.
• Don't jump to recommendations by default. Synthesis is about pattern-finding; the PM decides what to do with the patterns.
• Don't claim a theme based on one vivid interview. That's an anecdote — useful, but label it honestly.