Agent Skills: Research Summarizer

Research Summarizer

A skill focused on synthesizing and communicating research — the part that comes after you've collected the data. Distinct from the research collection skills which guide interview design, recruiting, and protocol.

This skill assumes you have raw inputs (transcripts, notes, survey responses) and need to turn them into trustworthy insights that drive product decisions.

When to use this skill

• Synthesizing a batch of user interviews (typically 5-30)
• Pulling themes from open-text survey responses
• Synthesizing support tickets for product-truth analysis
• Building a findings brief for stakeholders
• Separating signal from anecdote in qualitative data
• Auditing existing research summaries for bias and reliability
• Preparing a research readout for execs / cross-functional teams

Inputs the advisor expects

• Type of research artifacts (interviews, surveys, tickets, observations, sales notes)
• Volume and recency
• Research question(s) the synthesis is answering
• Audience for the output (PM team / exec / engineering)
• Decision the output should inform

Clarify First

Before generating the synthesis or brief, confirm these inputs. If any is unknown or vague, ASK — do not assume:

• [ ] The research question — what decision this synthesis answers (drives the brief's lead and which themes matter)
• [ ] Audience and the decision it informs — PM team, exec, or engineering (sets brief altitude, length, and format)
• [ ] Artifact type and volume — interviews/surveys/tickets and how many (drives confidence, sample-size adequacy, and bias checks)

Stop rule: ask only the 2-3 that most change the output. If the user says "just draft it," proceed and list your assumptions at the top of the artifact.

Workflows

Workflow 1 — Organize and theme raw research

1. Capture raw research items (one per row) with source, date, segment.
2. Run research_synthesis_organizer.py to surface theme clusters based on tagging, computed frequencies, and segment cross-cuts.
3. Refine themes manually; promote to insights.

python3 research-summarizer/scripts/research_synthesis_organizer.py \
  --input research_items.json --format markdown

Workflow 2 — Score insight quality

1. List proposed insights with supporting evidence count + segment coverage.
2. Run insight_quality_scorer.py to grade each insight on Confidence, Specificity, Action-readiness, and Bias risk.
3. Keep High / Medium insights; demote Low to "questions for further research."

python3 research-summarizer/scripts/insight_quality_scorer.py \
  --input insights.json --format markdown

Workflow 3 — Generate a findings brief

1. Capture the question, top insights, supporting evidence, decisions.
2. Run findings_brief_generator.py to produce the structured brief.

python3 research-summarizer/scripts/findings_brief_generator.py \
  --input findings.json --format markdown

Decision frameworks

Observation → Pattern → Insight → Recommendation

A clean synthesis ladder:

1. Observation — direct quote or behavior ("user X said Y")
2. Pattern — repeats across users ("4 of 7 users said Y")
3. Insight — interpreted explanation ("users avoid Y because Z")
4. Recommendation — action implied ("redesign Z to address Y avoidance")

Each level requires more confidence than the last. Don't skip from observation directly to recommendation.

Insight quality dimensions

• Confidence: how many independent sources support it
• Specificity: is the insight specific enough to action?
• Bias risk: is the sample / interpretation biased?
• Decision impact: does this insight change anything?

A high-quality insight scores well on all four. Most rough notes are strong on confidence but weak on specificity (or vice versa).

Sample size for qualitative research

A rough heuristic for how many interviews are enough:

| Goal | Suggested N | |------|-------------| | Discover the space (early product) | 5-8 | | Validate hypotheses | 8-12 | | Persona definition | 12-20 | | Detect quantitative signal in qual | 20-30+ | | Validate cross-segment | 5-8 per segment |

Diminishing returns after the patterns repeat 2-3 times. If you keep hearing new things, you're not done.

When qualitative data lies (common biases)

• Confirmation bias — interviewers pull the quotes they expected to hear
• Acquiescence bias — participants agree to be polite
• Recall bias — what users remember vs what they did
• Selection bias — who agreed to interview is not representative
• Recency bias — recent interviews carry disproportionate weight
• Anchor bias — first interview shapes interpretation of later ones
• Demand characteristics — participants guess what you want to hear

Counter: use a second coder, structure your guide, sample diversely, and report negative evidence.

Common engagements

"Help me synthesize 12 user interviews"

1. Make sure you have transcripts (or detailed notes).
2. Tag each interview by demographic, journey stage, key behaviors.
3. Surface 5-10 themes from initial tagging.
4. For each theme, count: how many users? from which segments? evidence quality?
5. Promote 3-5 themes to insights; demote the rest to "questions for next round."
6. Add 1-2 unexpected findings (the "we didn't expect this" insight).

"Translate the research into a one-pager for execs"

1. Lead with the question being asked.
2. Lead with the answer (1-2 sentences); details follow.
3. 3-5 insights with evidence; not more.
4. Decisions / recommendations that follow.
5. What you don't know yet (research limits + next-step questions).
6. Methodology one-liner (N, segments, dates).

"Our research found contradictory things"

1. First: is one finding from a different segment? Often the contradiction is segment-based.
2. Second: was sample biased toward one side?
3. Third: maybe both are true and the system has tensions worth surfacing.

Anti-patterns to avoid

• Cherry-picked quotes. Always provide the count + context.
• Insight without evidence. "Users want X" without supporting observations.
• Anecdotal generalization. One angry user doesn't define the population.
• Reporting interview-by-interview. Synthesis means seeing across users.
• Hiding the negative evidence. Disconfirming evidence is valuable.
• Brief that's longer than needed. Briefer = better-read.
• Mixing facts and interpretations. Be clear which is which.
• Skipping methodology. Readers need to evaluate the trust level.

References

• references/research-synthesis-frameworks.md — affinity, thematic analysis, frameworks
• references/insight-quality-and-bias.md — quality dimensions, bias catalog, validation
• references/communicating-research-findings.md — brief formats, presentation patterns

Related skills

• product-team/ux-researcher-designer — research design + collection
• product-team/product-strategist — strategic input from insights
• product-team/product-analytics — quant complement to qual
• c-level-advisor/chief-customer-officer-advisor — VoC program context