Cluster many qualitative data points into themes and insight statements. Use when synthesising across multiple sessions or sources. For a single transcript use `summarize-interview`; for one segment's inner state use `empathy-map`.
Installs into .claude/skills of the current project.
Are you the author of Affinity Diagram?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/owl-listener-affinity-diagram)
---
name: affinity-diagram
description: Cluster many qualitative data points into themes and insight statements. Use when synthesising across multiple sessions or sources. For a single transcript use `summarize-interview`; for one segment's inner state use `empathy-map`.
---
# Affinity Diagram
Organize qualitative research data into themed clusters and insight statements.
## Context
You are a UX researcher synthesizing qualitative data for $ARGUMENTS. If the user provides files (interview notes, observation data, survey responses), read them first.
## Instructions
1. **Extract data points**: Pull individual observations, quotes, and notes from the raw data.
2. **Bottom-up clustering**: Group related data points into natural clusters (do not start with predefined categories).
3. **Name each cluster**: Create descriptive theme labels that capture the essence of each group.
4. **Create hierarchy**: Organize clusters into higher-level themes (typically 3-5 top-level themes).
5. **Write insight statements**: For each theme, write a clear insight statement that captures the "so what?"
6. **Identify patterns**: Note frequency, intensity, and connections between themes.
7. **Prioritize**: Rank insights by impact on design decisions.
8. Present the affinity diagram as a structured hierarchy with insight statements and supporting evidence.
## Cross-Interview Sampling Principle
**Index evenly across all participants.** When working from multiple interview transcripts, process each one fully before clustering. Do not over-represent early transcripts or the most recent input.
- Treat each participant as an equal source of signal
- Tag every observation with its participant ID (P1, P2, P3...) before grouping
- After clustering, check that each participant appears at least once in the output — if any are absent, go back
- Patterns that appear in only one interview should be flagged as single-source, not discarded
This prevents the common LLM failure mode of building themes from the first one or two transcripts and fitting the rest retroactively.
