synthesize-research

SKILL.md

Synthesize Research

If you see unfamiliar placeholders or need to check which tools are connected, see CONNECTORS.md.

Synthesize user research from multiple sources into structured insights and recommendations.

Usage

/synthesize-research $ARGUMENTS

Workflow

1. Gather Research Inputs
Accept research from any combination of:

Pasted text: Interview notes, transcripts, survey responses, feedback

Uploaded files: Research documents, spreadsheets, recordings summaries

~~knowledge base (if connected): Search for research documents, interview notes, survey results

~~user feedback (if connected): Pull recent support tickets, feature requests, bug reports

~~product analytics (if connected): Pull usage data, funnel metrics, behavioral data

~~meeting transcription (if connected): Pull interview recordings, meeting summaries, and discussion notes

Ask the user what they have:

What type of research? (interviews, surveys, usability tests, analytics, support tickets, sales call notes)

How many sources / participants?

Is there a specific question or hypothesis they are investigating?

What decisions will this research inform?

2. Process the Research

For each source, extract:

Key observations: What did users say, do, or experience?

Quotes: Verbatim quotes that illustrate important points

Behaviors: What users actually did (vs what they said they do)

Pain points: Frustrations, workarounds, and unmet needs

Positive signals: What works well, moments of delight

Context: User segment, use case, experience level

3. Identify Themes and Patterns

Apply thematic analysis — see Research Synthesis Methodology below for detailed guidance on thematic analysis, affinity mapping, and triangulation techniques.

Group observations into themes, count frequency across participants, and assess impact severity. Note contradictions and surprises.

Create a priority matrix:

High frequency + High impact: Top priority findings

Low frequency + High impact: Important for specific segments

High frequency + Low impact: Quality-of-life improvements

Low frequency + Low impact: Note but deprioritize

4. Generate the Synthesis

Produce a structured research synthesis:

Research Overview

Methodology: what types of research, how many participants/sources

Research question(s): what we set out to learn

Timeframe: when the research was conducted

Key Findings

For each major finding (aim for 5-8):

Finding statement: One clear sentence describing the insight

Evidence: Supporting quotes, data points, or observations (with source attribution)

Frequency: How many participants/sources support this finding

Impact: How significantly this affects the user experience or business

Confidence level: High (strong evidence), Medium (suggestive), Low (early signal)

Order findings by priority (frequency x impact).

User Segments / Personas

If the research reveals distinct user segments:

Segment name and description

Key characteristics and behaviors

Unique needs and pain points

Size estimate if data is available

Opportunity Areas

Based on the findings, identify opportunity areas:

What user needs are unmet or underserved

Where do current solutions fall short

What new capabilities would unlock value

Prioritized by potential impact

Recommendations

Specific, actionable recommendations:

What to build, change, or investigate further

Tied back to specific findings

Prioritized by impact and feasibility

Open Questions

What the research did not answer:

Gaps in understanding

Areas needing further investigation

Suggested follow-up research methods

5. Review and Extend

After generating the synthesis:

Ask if any findings need more detail or different framing

Offer to generate specific artifacts: persona documents, opportunity maps, research presentations

Offer to create follow-up research plans for open questions

Offer to draft product implications (how findings should influence the roadmap)

Research Synthesis Methodology

Thematic Analysis

The core method for synthesizing qualitative research:

Familiarization: Read through all the data. Get a feel for the overall landscape before coding anything.

Initial coding: Go through the data systematically. Tag each observation, quote, or data point with descriptive codes. Be generous with codes — it is easier to merge than to split later.

Theme development: Group related codes into candidate themes. A theme captures something important about the data in relation to the research question.

Theme review: Check themes against the data. Does each theme have sufficient evidence? Are themes distinct from each other? Do they tell a coherent story?

Theme refinement: Define and name each theme clearly. Write a 1-2 sentence description of what each theme captures.

Report: Write up the themes as findings with supporting evidence.

Affinity Mapping

A collaborative method for grouping observations:

Capture observations: Write each distinct observation, quote, or data point as a separate note

Cluster: Group related notes together based on similarity. Do not pre-define categories — let them emerge from the data.

Label clusters: Give each cluster a descriptive name that captures the common thread

Organize clusters: Arrange clusters into higher-level groups if patterns emerge

Identify themes: The clusters and their relationships reveal the key themes

Tips for affinity mapping:

One observation per note. Do not combine multiple insights.

Move notes between clusters freely. The first grouping is rarely the best.

If a cluster gets too large, it probably contains multiple themes. Split it.

Outliers are interesting. Do not force every observation into a cluster.

The process of grouping is as valuable as the output. It builds shared understanding.

Triangulation

Strengthen findings by combining multiple data sources:

Methodological triangulation: Same question, different methods (interviews + survey + analytics)

Source triangulation: Same method, different participants or segments

Temporal triangulation: Same observation at different points in time

A finding supported by multiple sources and methods is much stronger than one supported by a single source. When sources disagree, that is interesting — it may reveal different user segments or contexts.

Interview Note Analysis

Extracting Insights from Interview Notes

For each interview, identify:

Observations: What did the participant describe doing, experiencing, or feeling?

Distinguish between behaviors (what they do) and attitudes (what they think/feel)

Note context: when, where, with whom, how often

Flag workarounds — these are unmet needs in disguise

Direct quotes: Verbatim statements that powerfully illustrate a point

Good quotes are specific and vivid, not generic

Attribute to participant type, not name: "Enterprise admin, 200-person team" not "Sarah"

A quote is evidence, not a finding. The finding is your interpretation of what the quote means.

Behaviors vs stated preferences: What people DO often differs from what they SAY they want

Behavioral observations are stronger evidence than stated preferences

If a participant says "I want feature X" but their workflow shows they never use similar features, note the contradiction

Look for revealed preferences through actual behavior

Signals of intensity: How much does this matter to the participant?

Emotional language: frustration, excitement, resignation

Frequency: how often do they encounter this issue

Workarounds: how much effort do they expend working around the problem

Impact: what is the consequence when things go wrong

Cross-Interview Analysis

After processing individual interviews:

Look for patterns: which observations appear across multiple participants?

Note frequency: how many participants mentioned each theme?

Identify segments: do different types of users have different patterns?

Surface contradictions: where do participants disagree? This often reveals meaningful segments.

Find surprises: what challenged your prior assumptions?

Survey Data Interpretation

Quantitative Survey Analysis

Response rate: How representative is the sample? Low response rates may introduce bias.

Distribution: Look at the shape of responses, not just averages. A bimodal distribution (lots of 1s and 5s) tells a different story than a normal distribution (lots of 3s).

Segmentation: Break down responses by user segment. Aggregates can mask important differences.

Statistical significance: For small samples, be cautious about drawing conclusions from small differences.

Benchmark comparison: How do scores compare to industry benchmarks or previous surveys?

Open-Ended Survey Response Analysis

Treat open-ended responses like mini interview notes

Code each response with themes

Count frequency of themes across responses

Pull representative quotes for each theme

Look for themes that appear in open-ended responses but not in structured questions — these are things you did not think to ask about

Common Survey Analysis Mistakes

Reporting averages without distributions. A 3.5 average could mean everyone is lukewarm or half love it and half hate it.

Ignoring non-response bias. The people who did not respond may be systematically different.

Over-interpreting small differences. A 0.1 point change in NPS is noise, not signal.

Treating Likert scales as interval data. The difference between "Strongly Agree" and "Agree" is not necessarily the same as between "Agree" and "Neutral."

Confusing correlation with causation in cross-tabulations.

Combining Qualitative and Quantitative Insights

The Qual-Quant Feedback Loop

Qualitative first: Interviews and observation reveal WHAT is happening and WHY. They generate hypotheses.

Quantitative validation: Surveys and analytics reveal HOW MUCH and HOW MANY. They test hypotheses at scale.

Qualitative deep-dive: Return to qualitative methods to understand unexpected quantitative findings.

Integration Strategies

Use quantitative data to prioritize qualitative findings. A theme from interviews is more important if usage data shows it affects many users.

Use qualitative data to explain quantitative anomalies. A drop in retention is a number; interviews reveal it is because of a confusing onboarding change.

Present combined evidence: "47% of surveyed users report difficulty with X (survey), and interviews reveal this is because Y (qualitative finding)."

When Sources Disagree

Quantitative and qualitative sources may tell different stories. This is signal, not error.

Check if the disagreement is due to different populations being measured

Check if stated preferences (survey) differ from actual behavior (analytics)

Check if the quantitative question captured what you think it captured

Report the disagreement honestly and investigate further rather than choosing one source

Persona Development from Research

Building Evidence-Based Personas

Personas should emerge from research data, not imagination:

Identify behavioral patterns: Look for clusters of similar behaviors, goals, and contexts across participants

Define distinguishing variables: What dimensions differentiate one cluster from another? (e.g., company size, technical skill, usage frequency, primary use case)

Create persona profiles: For each behavioral cluster:

Name and brief description

Key behaviors and goals

Pain points and needs

Context (role, company, tools used)

Representative quotes

Validate with data: Can you size each persona segment using quantitative data?

Persona Template

[Persona Name] — [One-line description]

Who they are:
- Role, company type/size, experience level
- How they found/started using the product

What they are trying to accomplish:
- Primary goals and jobs to be done
- How they measure success

How they use the product:
- Frequency and depth of usage
- Key workflows and features used
- Tools they use alongside this product

Key pain points:
- Top 3 frustrations or unmet needs
- Workarounds they have developed

What they value:
- What matters most in a solution
- What would make them switch or churn

Representative quotes:
- 2-3 verbatim quotes that capture this persona's perspective

Common Persona Mistakes

Demographic personas: defining by age/gender/location instead of behavior. Behavior predicts product needs better than demographics.

Too many personas: 3-5 is the sweet spot. More than that and they are not actionable.

Fictional personas: made up based on assumptions rather than research data.

Static personas: never updated as the product and market evolve.

Personas without implications: a persona that does not change any product decisions is not useful.

Opportunity Sizing

Estimating Opportunity Size

For each research finding or opportunity area, estimate:

Addressable users: How many users could benefit from addressing this? Use product analytics, survey data, or market data to estimate.

Frequency: How often do affected users encounter this issue? (Daily, weekly, monthly, one-time)

Severity: How much does this issue impact users when it occurs? (Blocker, significant friction, minor annoyance)

Willingness to pay: Would addressing this drive upgrades, retention, or new customer acquisition?

Opportunity Scoring

Score opportunities on a simple matrix:

Impact: (Users affected) x (Frequency) x (Severity) = impact score

Evidence strength: How confident are we in the finding? (Multiple sources > single source, behavioral data > stated preferences)

Strategic alignment: Does this opportunity align with company strategy and product vision?

Feasibility: Can we realistically address this? (Technical feasibility, resource availability, time to impact)

Presenting Opportunity Sizing

Be transparent about assumptions and confidence levels

Show the math: "Based on support ticket volume, approximately 2,000 users per month encounter this issue. Interview data suggests 60% of them consider it a significant blocker."

Use ranges rather than false precision: "This affects 1,500-2,500 users monthly" not "This affects 2,137 users monthly"

Compare opportunities against each other to create a relative ranking, not just absolute scores

Output Format

Use clear headers and structured formatting. Each finding should stand on its own — a reader should be able to read any single finding and understand it without reading the rest.

Tips

Let the data speak. Do not force findings into a predetermined narrative.

Distinguish between what users say and what they do. Behavioral data is stronger than stated preferences.

Quotes are powerful evidence. Include them generously, with attribution to participant type (not name).

Be explicit about confidence levels. A finding from 2 interviews is a hypothesis, not a conclusion.

Contradictions in the data are interesting, not inconvenient. They often reveal distinct user segments.

Recommendations should be specific enough to act on. "Improve onboarding" is not actionable. "Add a progress indicator to the setup flow" is.

Resist the temptation to synthesize too many themes. 5-8 strong findings are better than 20 weak ones.