Data synthesis

Data Synthesis -- Research Intelligence

Skill pack for transforming raw research data into structured insights. Covers affinity mapping, thematic analysis, quantitative aggregation, triangulation, insight scoring, gap analysis, and integration with the Memoire research pipeline.

Core Loop

COLLECT RAW DATA --> ORGANIZE --> CODE --> THEME --> TRIANGULATE --> SCORE --> OUTPUT INSIGHTS

Every synthesis pass must produce actionable insights, not summaries. An insight is a non-obvious finding that changes a design decision.

Affinity Mapping Methodology

What It Is

A bottom-up clustering technique. Individual observations (sticky notes, quotes, data points) are grouped by natural similarity until themes emerge.

Process

Step	Action	Output
1. Harvest	Extract individual observations from raw data. One fact per unit.	50-300 discrete data points
2. Spread	Lay all data points out without grouping. Randomize order.	Ungrouped pool
3. Cluster	Silently group related points. Move items that feel similar together.	8-20 raw clusters
4. Label	Name each cluster with a descriptive header that captures the shared meaning.	Named clusters
5. Supergroup	Group clusters into 3-6 high-level themes.	Theme hierarchy
6. Prioritize	Rank themes by frequency, severity, and opportunity.	Prioritized theme map

Rules

• One observation per data point -- never combine two findings into one note
• Group by meaning, not by source (do not cluster "all interview 3 data" together)
• If a cluster has more than 15 items, split it
• If a cluster has fewer than 3 items, consider merging or flagging as an outlier
• Relabel clusters after regrouping -- initial labels often drift from final content

In Memoire

memi research from-stickies       # Pull sticky notes from Figma canvas
memi research synthesize           # Run affinity mapping + thematic analysis
memi research report               # Generate structured insight report

Thematic Analysis

Three Coding Phases

Phase 1: Open Coding

Read through all data and assign descriptive codes to each meaningful segment.

Raw Data	Open Code
"I always forget where I saved the file"	file-location-confusion
"The search never finds what I need"	search-ineffective
"I end up asking a colleague where things are"	workaround-social

Rules:

• Code everything relevant, skip nothing on first pass
• Use gerunds or descriptive phrases (not single words)
• One segment can have multiple codes
• Target 30-80 unique codes from a typical study

Phase 2: Axial Coding

Group open codes into categories. Identify relationships between categories.

Category	Open Codes
Navigation Failure	file-location-confusion, search-ineffective, breadcrumb-ignored
Social Workarounds	workaround-social, slack-channel-for-files, ask-manager
Mental Model Mismatch	expected-folder-structure, alphabetical-assumption, recency-bias

Relationships to identify:

• Causal: A leads to B ("navigation failure causes social workarounds")
• Contextual: A happens during B ("mental model mismatch occurs during onboarding")
• Strategic: A is a response to B ("social workarounds compensate for navigation failure")

Phase 3: Selective Coding

Identify the core theme that connects all categories. This becomes the central narrative.

Core Theme: "Users build parallel information architectures through social channels
             because the system's structure does not match their mental models."

Code Quality Checklist

• Every data segment is coded (no orphan data)
• Codes are applied consistently across all sources
• Categories have clear boundaries (no code belongs to 3+ categories)
• Core theme is supported by data from multiple participants/sources

Quantitative Aggregation

Standard Metrics

Metric	Formula	Use Case
Mean	Sum / Count	Average satisfaction, task time
Median	Middle value (sorted)	Task time (resistant to outliers)
Mode	Most frequent value	Preferred option selection
Std Dev	sqrt(variance)	Spread of responses
NPS	%Promoters - %Detractors	Overall loyalty/satisfaction
SUS Score	((Sum of odd - 5) + (25 - Sum of even)) * 2.5	System usability
Task Success Rate	Successes / Total Attempts * 100	Effectiveness
Time on Task	Median completion time	Efficiency
Error Rate	Errors / Opportunities * 100	Error-proneness

NPS Calculation

Responses on 0-10 scale:
  Detractors: 0-6
  Passives:   7-8
  Promoters:  9-10

NPS = (Promoter Count / Total) * 100 - (Detractor Count / Total) * 100
Range: -100 to +100

NPS Range	Interpretation
-100 to 0	Critical -- systemic issues
0 to 30	Below average -- improvement needed
30 to 50	Good -- competitive
50 to 70	Excellent -- strong loyalty
70 to 100	World-class

SUS Score Interpretation

SUS Score	Grade	Percentile
< 51	F	Bottom 15%
51-68	D	15-50%
68-74	C	50-70%
74-80	B	70-85%
80-90	A	85-97%
> 90	A+	Top 3%

Statistical Significance

• For A/B comparisons: use chi-squared test (categorical) or t-test (continuous)
• Minimum sample size for usability: 5 per segment (qualitative), 30+ per variant (quantitative)
• Report confidence intervals, not just point estimates
• Flag any metric based on fewer than 5 data points as unreliable

Triangulation Across Sources

Triangulation Matrix

Cross-reference findings across at least 3 data sources to validate strength.

Finding	Interviews	Surveys	Analytics	Usability Tests	Strength
Users abandon search after 2 attempts	Mentioned by 4/6	62% agree	Avg 1.8 queries before exit	3/5 gave up	STRONG
Mobile nav is confusing	2/6 mentioned	28% agree	Mobile bounce 2x desktop	Not tested	MODERATE
Dark mode is desired	1/6 mentioned	71% want it	N/A	N/A	WEAK (conflicting)

Strength Classification

Level	Criteria
STRONG	Confirmed by 3+ independent sources with consistent direction
MODERATE	Confirmed by 2 sources, or 3 sources with some inconsistency
WEAK	Single source, or multiple sources with conflicting signals
CONTRADICTED	Sources actively disagree -- requires deeper investigation

Insight Quality Scoring

Rate every insight on four dimensions before including it in a deliverable.

Dimension	Score 1	Score 3	Score 5
Evidence	Single anecdote	Multiple data points, one source	Multiple data points, multiple sources
Actionability	Vague observation	Suggests a direction	Points to a specific design change
Novelty	Already known by team	Partially known	Genuinely new understanding
Impact	Affects edge cases	Affects a segment	Affects most users

Threshold: Only insights scoring 12+ (out of 20) go into the final report. Insights scoring 8-11 go into an appendix. Below 8, discard or flag for follow-up research.

Insight Statement Format

[WHO] experiences [WHAT PROBLEM/BEHAVIOR]
because [WHY/ROOT CAUSE],
which means [IMPLICATION FOR DESIGN].

Example:

New users abandon the onboarding wizard at step 3
because the required fields feel invasive for a first interaction,
which means we should defer profile completion to post-activation.

Gap Analysis Frameworks

Coverage Gap Matrix

Map research questions against data sources to identify blind spots.

Research Question	Source A	Source B	Source C	Gap?
Why do users churn at day 7?	Partial	No data	Partial	YES
What drives upgrade decisions?	Full	Full	Partial	No
How do teams collaborate?	No data	No data	Full	YES

Types of Gaps

Gap Type	Description	Resolution
Data gap	No data collected on this question	Plan follow-up study
Segment gap	Data exists but not for a key user segment	Recruit missing segment
Depth gap	Quantitative data but no qualitative understanding	Conduct interviews
Recency gap	Data is older than 6 months	Refresh study
Context gap	Lab data but no field observation	Run diary study or contextual inquiry

Integration with Memoire Research Pipeline

Pipeline Stages

RAW DATA (Excel, CSV, stickies)
  |
  v
memi research from-file <path>          # Ingest raw data
  |
  v
memi research from-stickies             # Pull Figma sticky notes
  |
  v
memi research synthesize                # Affinity map + thematic analysis
  |                                      # Outputs coded themes + scored insights
  v
memi research report                    # Structured insight report
  |                                      # Markdown + Figma sticky output
  v
memi spec component <name>              # Insights feed component specs
memi spec page <name>                   # Insights feed page specs
memi compose "dashboard from research"  # Agent builds dashboard from insights

Data Formats

Input Format	Command	Notes
Excel (.xlsx)	from-file	One sheet per study; row 1 = headers
CSV (.csv)	from-file	UTF-8 encoding required
Figma stickies	from-stickies	One observation per sticky
JSON	from-file	Array of { text, source, participant?, timestamp? }

Output Artifacts

Artifact	Format	Destination
Coded themes	JSON	.memoire/research/themes.json
Insight cards	JSON + Markdown	.memoire/research/insights/
Affinity map	Figma stickies	Canvas (grouped by theme)
Quantitative summary	JSON	.memoire/research/metrics.json
Final report	Markdown	.memoire/research/report.md

Connecting Insights to Specs

When generating component or page specs, reference insight IDs:

{
  "name": "OnboardingWizard",
  "level": "organism",
  "researchInsights": ["INS-007", "INS-012"],
  "designRationale": "Deferred profile completion based on day-1 drop-off finding (INS-007)"
}

This creates traceability from research through spec to generated code.