User Research Synthesis skill for Claude Code

Synthesis is done well when every finding can be traced to specific participants and quotes, the strength of evidence is stated honestly, surprises and contradictions are kept rather than smoothed over, and each recommendation follows from a finding. The output is a decision aid, not a pile of quotes.

Ground rules

• Separate observation from interpretation. “P4 exported to Excel to add a column” is an observation; “users don’t trust our reports” is an interpretation that needs several observations.
• Count participants, not mentions. One person saying it five times is one participant.
• Every insight cites participant IDs and at least one verbatim quote.
• State confidence per finding (table below). Small samples support directional claims, not percentages. With 8 interviews, write “6 of 8”, never “75% of users”.
• Keep disconfirming evidence in the report next to the finding it challenges.
• Anonymize: participant IDs (P1…Pn) with role and segment, no names, emails, or company names unless consent covers it.

Workflow

- [ ] 1. Restate the research questions and the decision the research informs
- [ ] 2. Inventory sources; build the participant table (ID, segment, role, date)
- [ ] 3. Extract atomic observations from each source (one fact or quote per note)
- [ ] 4. Cluster observations bottom-up into themes (affinity mapping)
- [ ] 5. Build the evidence matrix: participants x themes
- [ ] 6. Write insight statements with confidence and counter-evidence
- [ ] 7. Turn insights into recommendations and open questions
- [ ] 8. Review traceability; write the readout

Step 3: atomic observations

One note = one observation, tagged with participant and source. Use this shape:

id,participant,segment,source,type,note
O-014,P3,SMB admin,interview-2026-09-12,behavior,"Exports invoices to Excel every month to add the cost-center column"
O-015,P3,SMB admin,interview-2026-09-12,quote,"""I don't trust the totals until I've rebuilt them myself."""
O-016,P3,SMB admin,interview-2026-09-12,pain,"Month-end close takes about 2 days, half of it reconciliation"
O-017,P5,Mid-market finance,usability-2026-09-14,behavior,"Could not find export from invoice list; used search after 90 s"

Types: behavior (what they did), quote, pain, goal, workaround, context, usability (task outcome). Aim for 30–80 observations per hour of interview. When working from transcripts, extract observations before reading anyone’s summary to avoid inheriting their framing.

Step 4: affinity clustering

1. Start bottom-up: group observations that are about the same underlying thing, without pre-set categories.
2. Name each cluster with a sentence that states the finding, not a topic label. “Reports” is a label; “Admins rebuild report totals because they can’t see how they are calculated” is a finding.
3. Clusters of 1–2 observations from one participant are parked, not deleted.
4. Group clusters into 4–8 higher-level themes. More than 8 means the themes are topics.
5. Compare with the research questions: note which questions got no evidence.

Step 5: evidence matrix

Theme	P1	P2	P3	P4	P5	P6	P7	P8	Count	Segments
Rebuild totals in Excel to trust them	x		x	x		x	x		5/8	SMB, mid-market
Export hard to find from invoice list		x			x			x	3/8	all
Cost-center column missing	x		x			x			3/8	SMB only
Prefer scheduled email of report				x					1/8	mid-market
Counter: trusts totals, no export needed		x			x				2/8	mid-market

The matrix shows breadth, segment skew, and the counter-evidence at a glance.

Step 6: insight statements

Pattern: who + does/feels what + because + so what, plus evidence and confidence.

Insight 1 (high confidence): SMB and mid-market admins rebuild invoice totals in Excel before month-end because the product doesn’t show how totals are calculated, adding hours to every close. Evidence: P1, P3, P4, P6, P7 (5/8). P3: “I don’t trust the totals until I’ve rebuilt them myself.” Counter: P2 and P5 (mid-market, ERP-synced) trust totals as-is.

Confidence	Evidence needed
High	Majority of participants in the relevant segment, observed behavior (not just opinion), consistent across sources
Medium	Several participants or one segment; mostly stated rather than observed; or one source type
Low / signal	1–2 participants or a single striking case; worth following up

Raise confidence by triangulation: does analytics, support data, or a survey show the same pattern? Cite it if so (see customer-feedback-analysis for ticket mining).

Step 7: recommendations

Insight	Recommendation	Type	Confidence
1. Rebuild totals to trust them	Show a calculation breakdown on totals; test whether Excel exports drop	Design + measure	High
2. Export hard to find	Add export to the invoice list toolbar; retest with 5 users	Usability fix	Medium
3. Cost-center column missing	Validate with a larger SMB sample before building	Further research	Low

Recommendations say what to do and how to know it worked. Low-confidence insights lead to more research, not to builds.

Personas and journey maps (when asked)

Build personas only from clusters of behavior and goals seen across several participants. Each persona lists the participants it is based on. A persona with no participant IDs is fiction. For journey maps, use observed steps and pains with participant evidence per stage.

Gotchas

• Confirmation bias. The team’s favorite hypothesis gets a theme of its own while contradicting notes vanish. Run a pass looking only for counter-evidence.
• Quote cherry-picking. A vivid quote from one participant drives the whole readout. Pair every headline quote with its count.
• Percentages from small samples. “62.5% of users” from 8 interviews implies precision that does not exist.
• Topic labels as findings. “Onboarding” tells the reader nothing; write the sentence.
• Mixing segments. A theme strong in SMB and absent in enterprise is two findings.
• Stated vs. observed. People say they read release notes; analytics says otherwise. Weight observed behavior above stated preference.
• Leaking PII into shared reports. Scrub names, emails, company names, and screenshots.

Verification

• Pick three findings at random and trace each back to observation IDs and raw sources.
• Every insight has a participant count, at least one quote, a confidence level, and any counter-evidence.
• Each research question is answered or explicitly marked as not answered.
• No percentages reported for samples under ~30.
• No names or identifying details remain.

Output contract

# Research readout: <topic> — <dates>
Decision this informs: <decision>   Method: <interviews/usability/survey>, n=<x>
Participants: <table of ID, segment, role>

## Top insights (3–6)
### Insight 1 (<confidence>)
<statement> — Evidence: <IDs, count> — Quotes — Counter-evidence
## Evidence matrix
## Recommendations
| Insight | Recommendation | How we'll know | Confidence |
## Unanswered questions and next research
## Appendix: observation log, method, discussion guide