Research Synthesis

🔵 Optional

Use this when: You have raw research data — interview transcripts, survey results, support tickets, analytics — and need to turn it into actionable insights that drive the backlog.

Quick version: Use the /research-synthesize skill in Claude Code for one-command execution.

Why this matters

Research is only valuable if it changes what you build. Raw data sitting in a folder doesn't do that. Synthesis — identifying patterns, extracting insights, and connecting them to the backlog — is the bridge between learning and action.

AI agents excel at processing large amounts of text and surfacing patterns. They're fast, thorough, and don't get anchored by the first interview they read. But they lack your domain knowledge and judgment about what matters most.

What to prepare

Before running this workflow, gather:

Raw research data: interview transcripts, survey responses, support tickets, analytics screenshots, usability test notes
Research questions: what were you trying to learn?
Current context: what does the team believe about users right now? What assumptions are you testing?
Current backlog: so the agent can connect insights to existing work

The workflow

Step 1: Surface patterns

I conducted (number) user interviews about (topic/feature).

Research questions:
(What we were trying to learn)

Here are the transcripts/notes:

(Paste or upload interview notes)

Analyze these and tell me:

1. What are the top 3-5 themes or patterns across participants?
2. For each theme, include specific quotes from participants that illustrate it
3. Where do participants agree? Where do they disagree?
4. What surprised you — things that came up that I didn't directly ask about?
5. What questions remain unanswered that we should investigate further?

Step 2: Connect to the backlog

Here's our current backlog:

(Paste backlog or key stories)

Based on the research synthesis above:

1. Which existing stories are validated by this research? (Users confirmed the need)
2. Which existing stories are challenged by this research? (Users didn't express this need, or expressed it differently)
3. What new stories should we consider adding?
4. How should this research affect our priority order?

Step 3: Generate a planning inbox entry

Based on the research synthesis and backlog analysis above, draft a planning inbox entry in this format:

## Source
(Where the data came from, date, participants)

## Key findings
(Bulleted list of the most important insights, with supporting evidence)

## Impact on current plan
(What should change in the backlog based on these findings)

## Recommended actions
(Specific next steps — stories to add, priorities to shift, follow-up research to plan)

## Confidence level
(How confident are we in these findings? What would increase confidence?)

Step 4: Validate with your judgment

The agent's synthesis is a starting point. Before sharing with the team:

Check pattern accuracy — Did the agent identify real patterns, or force-fit themes? Reread the source data for the most surprising findings.
Assess quote quality — Are the quotes representative, or cherry-picked to support a theme?
Validate connections — Does the backlog mapping make sense, or is the agent stretching?
Add your observations — You were in the room (or on the call). What did the agent miss about tone, body language, or context?

What to review

Check	Why
Themes are grounded in evidence	The agent should cite specific quotes, not make unsupported claims
Surprising findings are real	Cross-check against the raw data — agents can hallucinate patterns
Backlog connections are genuine	Not every insight maps to an existing story; that's fine
Confidence levels are honest	If you talked to 3 people, the confidence level is low regardless of how clear the patterns seem

Tool-specific tips

Tool	When to use it
Claude Projects	Best when you have ongoing research across multiple sessions; the project maintains context over time
Claude with file upload	Good for batch processing multiple transcripts in one session
ChatGPT Projects	Same as Claude Projects; strong at data analysis if you have quantitative survey results
NotebookLM	Best for audio recordings; can process interview audio directly and generate notes
Gemini + Google Workspace	Good if your research data lives in Google Docs/Sheets

Common mistakes

Synthesizing without research questions — If you don't tell the agent what you were trying to learn, it'll tell you what it finds interesting, which may not be what matters.
Trusting small samples — 3 interviews can reveal themes worth investigating, not themes worth building on. Be honest about sample size.
Confirmation bias — If the agent's synthesis confirms what you already believed, be extra skeptical. Ask it: "What in this data contradicts our current assumptions?"
Synthesis without action — If the synthesis doesn't change the backlog, it was wasted effort. Always end with specific recommended actions.
Skipping the planning inbox — Research insights should flow into the planning inbox so they're captured and considered alongside other signals. Don't let insights die in a document.

AI in practice: research synthesis beyond user interviews

This workflow applies to many client scenarios beyond product research:

Market intelligence reports — Synthesize earnings calls, press releases, and internal data into competitive landscape briefings. A strategy consultant turns a 2-day task into 30 minutes of review.
Cross-system reporting — Pull from email, Teams/Slack, CRM, and project tools to generate unified status reports that would otherwise require manual compilation across 4+ systems.
Customer feedback clustering — Process hundreds of NPS responses and support tickets to surface themes. Teams find patterns like "3 of the top 5 pain points are onboarding-related" that aren't visible in individual ticket review.

See the full Research & Knowledge use cases for more scenarios.

Artium open-source tools for research synthesis

Insight Extractor — automated pipeline that ingests Zoom meeting recordings/transcripts, extracts AI-powered insights, and writes results to Notion and Google Drive. Useful for teams that want to automate the first pass of meeting synthesis.
Feedback Facilitator — collect structured feedback on prototypes and designs with embedded content support (Figma, YouTube, Loom, PDFs). AI synthesizes responses into themes, sentiment, and action items.

Frequency

After every batch of research — interviews, survey rounds, usability tests. Don't let raw data age. Synthesize within 24-48 hours of the research while context is fresh.