A 40-minute synthesis workflow for cross-functional teams. Take a stack
of interviews, support tickets, and NPS responses — and use Claude to
turn it into a structured framework ready for a product strategy
conversation. The workflow below is what we ran in the room.
The 5-step workflow
01
Set context
Business goal, data sources, your specific question. Before any data.
02
Load data
4–5 interviews, then tickets + NPS. Orient Claude on each column.
03
Extract surprises
Cross-source patterns, segment disagreement, missing voices. Before any framework.
04
Build framework
Default: Opportunity Solution Tree. Alt: JTBD, HMW + journey map.
05
Pressure-test
Argue against your #1. Reformat as a 1-page exec summary.
Corrective prompts
When Claude returns patterns dressed as surprises
"That's a pattern, not a surprise. Tell me something I wouldn't
have guessed from reading two transcripts."
When Claude returns summary, not synthesis
"That's a summary of each source, not a synthesis across them.
What do these transcripts, tickets, and NPS comments
collectively reveal that no single source shows on its own?"
Before trusting any quote
"Can you cite which transcript this comes from? How many data
points support this finding?"
To kill generic recommendations
"If this recommendation could apply to any SaaS company, it's too
vague. Push for evidence-grounded specifics — name the segment,
the source, the quote."
Where human judgment is required
- Weight, don't count. One quote from your largest enterprise customer may matter more than 15 Starter-tier tickets. Frequency ≠ importance.
- Read the absence. The dog that didn't bark. Churned customers can't answer your survey. The quietly dissatisfied don't open tickets. Ask who isn't in the room.
- Hold strategic tension. When the "obvious" answer is complicated by a single sharp interview, good synthesis holds both sides — not just the majority.
- Validate emotional inference. Claude infers beliefs and feelings from text. If its read doesn't match yours, trust your judgment and push back with specific evidence.
- Spot-check the raw. Before you trust a theme, open 2–3 verbatims behind it. Averaging tickets into bullets is useful — and quietly erases the specific person who was frustrated.
If your org restricts AI tools
- Synthetic first. Build muscle memory on a fictional dataset (like Launchpad), then apply within approved guardrails.
- Anonymize. Strip names and identifiers before upload. Synthesis still works — Claude doesn't need to know who said it.
- Enterprise instance. Many orgs already have a compliant Claude or ChatGPT path. Check before you build a workaround.
Hallucination traps to watch for
⚠ Invented quotes
Claude may paraphrase or fabricate plausible-sounding quotes. Always demand the transcript source.
⚠ False consensus
7 transcripts say one thing, 1 says another — and Claude weighs them equally. Ask: "How many data points support this?"
⚠ Hypothesis confirmation
Share your hypothesis with the data and Claude leans toward confirming it. Force the opposite: "Argue against this."
⚠ Generic recommendations
"Improve mobile experience" applies to any SaaS company. Push for the Launchpad-specific evidence, segment, and quote.
Four takeaways to leave with
Context first, always.
The synthesis context you set in Step 1 determines everything. Generic input → generic output. Anchor Claude to your specific business goal before sharing a single data point.
Demand cross-source synthesis.
Push beyond single-source summaries. The most valuable insights only emerge when interviews, tickets, and NPS are cross-referenced — including contradictions and absences.
You are the final judge.
Claude surfaces patterns; you weigh them. Frequency isn't importance. Strategic tensions require human judgment. Always pressure-test before bringing output to a stakeholder.
Under 40 minutes.
What used to take 2–3 days now fits in a focused session — leaving time for the decisions that actually require your expertise. Run it weekly on new interviews and NPS data.