Research Synthesis skill for Codex

Synthesis is finished when a decision-maker can act on one page: the recommendation, the two or three findings it rests on, how strong that evidence is, the main risk, and what evidence would reverse it. Summarising each source in turn is not synthesis.

Inputs check

Before writing anything, inventory the evidence:

| ID | Source | Type | Date | n / scope | Quality notes |
|---|---|---|---|---|---|
| S1 | 12 customer interviews | qualitative | 2026-09 | mid-market admins | recruited from churned accounts |
| S2 | In-app survey | quantitative | 2026-10 | 840 responses, 6% rate | opt-in |
| S3 | Usage analytics | behavioural | last 90 d | all 4,100 accounts | export events only |
| S4 | Two competitor teardowns | secondary | 2026-08 | - | public pages |

Then pin the decision: what choice, by whom, by when, and what options are on the table. If no decision is defined, ask; synthesis without a decision drifts into a report.

Workflow

- [ ] 1. Inventory sources and rate their quality
- [ ] 2. Extract atomic findings with source IDs
- [ ] 3. Cluster findings into themes
- [ ] 4. Weigh: convergence, quality, relevance
- [ ] 5. Resolve or explain contradictions
- [ ] 6. Write insights (finding + implication)
- [ ] 7. Map insights to options; recommend
- [ ] 8. State confidence, risks, and reversal conditions

2. Atomic findings

One observation per line, in neutral language, tagged with its source. Keep the raw evidence pointer (quote ID, chart, table).

F1 [S1] 7 of 12 admins export to spreadsheets weekly to build reports for finance.
F2 [S3] 31% of active accounts used CSV export at least once in 90 days.
F3 [S2] "Reporting" ranked 4th of 6 in requested improvements.
F4 [S4] Both main competitors ship scheduled reports on mid-tier plans.

3-4. Themes and weighing

Group findings that point the same way. For each theme, weigh the evidence:

Factor	Stronger	Weaker
Convergence	Different methods agree (interviews + analytics)	One method only
Behaviour vs opinion	What people did (logs, purchases)	What people say they would do
Sample	Representative of the decision’s population	Convenience or skewed recruit
Recency	After the last relevant product or market change	Before it
Effect size	Large and practical	Statistically detectable but small
Independence	Separate teams, sources, methods	Same source reused

Triangulation is the core move: a theme supported by behavioural data and qualitative explanation from different sources is worth more than ten quotes saying the same thing.

5. Contradictions

When findings conflict (survey ranks reporting low; interviews say it is painful):

1. Check populations: do the sources describe the same people? (S1 recruited churned accounts; S2 is active users.)
2. Check measures: stated priority vs observed workaround behaviour measure different things.
3. Check timing and definitions.
4. If explained, state the explanation as an insight. If not, carry it forward as an open question with the evidence that would settle it.

Never drop the inconvenient finding.

6. Insights

An insight pairs a finding with why it matters for the decision:

**Reporting pain drives churn among admins who answer to finance, not the general base.**
Workaround exports are common (31% of accounts, F2) and interviewees describe them as a
weekly chore (F1), yet active users rank reporting 4th (F3). The pain concentrates in
accounts with a finance stakeholder, the segment over-represented in churned interviews.
Confidence: moderate (two methods; interview sample skewed toward churned accounts).

Label inferences as inferences. Keep facts (“31% exported”) separate from interpretations (“they export because reports are missing”).

7. Options and recommendation

Lay out 2-4 real options, including “do nothing” when plausible, against the criteria that matter for this decision:

Option	Addresses	Cost / effort	Risk	Evidence for	Evidence against
A Scheduled reports	F1, F2, F4	~6 weeks	Low adoption outside finance-linked accounts	S1, S3	S2
B Better CSV export	F2	~1 week	Leaves the chore in place	S3	S1
C Do nothing	-	0	Continued churn in that segment	-	S1

Recommend one, say why it beats the runner-up, and name the cheapest way to test it before full commitment (pilot, fake door, limited rollout).

8. Confidence and reversal

• Overall confidence: high / moderate / low, with the one or two reasons.
• Key assumption: the belief that, if false, changes the recommendation.
• Reversal signal: the observation that should make the team change course, and when to check for it.

Gotchas

• Frequency is not importance: the most mentioned theme may be the easiest to articulate, not the costliest problem.
• Vivid quotes persuade beyond their weight. Use them to illustrate a theme already supported by broader evidence, never as the evidence.
• Confirmation pull: list the evidence against the recommendation explicitly; if that column is empty, look again.
• Mixing segments hides the signal; split findings by segment when sources differ.
• Stale research: findings from before a major product or market change need re-checking.
• For a formal decision document format, hand the result to $decision-memos; for interview-specific affinity mapping use $user-research-synthesis.

Verification

• Every insight cites findings, and every finding cites a source ID.
• Each option row lists evidence for and against.
• A reader who disagrees can see exactly which finding they would need to dispute.
• The recommendation answers the decision stated at the start, by the stated deadline.

Output template

# <Decision>: synthesis and recommendation
**Recommendation:** <option> - <one-sentence why>  **Confidence:** moderate
**Decide by:** <date> | **Owner:** <name>

## What we know (insights)
1. <Insight> (F#, F#) - confidence
2. ...

## Where evidence conflicts

## Options compared
<table>

## Risks and the assumption that matters most
## What would change this recommendation
## Next step to de-risk
## Appendix: sources and findings