Insight Synthesis

Insight Synthesis

> Part of the discovery-phase skill pack · synthesis group · reads discovery-context.md (run profile-builder first if missing).

Turns whatever evidence the BA managed to gather into structured insights — weighted by source quality so the team doesn't treat one SME's hunch as equal to five user interviews.

Step 1 — Read discovery context

Read discovery-context.md (section 4. Access & Data) to know what evidence to expect. If missing, ask the BA inline: "which evidence types should I expect — interviews / SMEs / tickets / secondary / competitive / mix?" — or weight all sources equally and tag the output [NO-WEIGHTING-CONTEXT]. Never block; recommend profile-builder for high-stakes work.

Then scan ./discovery/ for actually-present artifacts:

• interview-notes/*.md (direct evidence — highest weight)
• sme-notes/*.md (proxy evidence — medium weight)
• support-data-analysis.md (unsolicited signal — high weight)
• secondary-research.md (published data — variable weight by source)
• competitive-scan.md (market signal — context, not direct evidence)

If fewer than 2 sources are present, tell the BA: "Synthesis with single-source evidence is fragile. Consider running another evidence skill before this. Proceed anyway?"

Step 2 — Affinity mapping (sticky-notes phase)

For every distinct observation across all sources, create a one-line entry in ./discovery/_observations.md:

- [P03 | direct] "I keep a spreadsheet because the tool doesn't filter by region"
- [P01 | direct] 15 min/day lost to manual status updates
- [SME-Maria | inferred] "Users get stuck at the third step" (no specific user cited)
- [tickets | n=23] complaints about CSV export breaking
- [secondary | Gartner 2025] 60% of similar tools lack X

Tag each with [source | confidence]. Aim for 30-80 observations across all sources.

Step 3 — Cluster into themes

Group by underlying job, not surface vocabulary. Target 5-8 themes. More than 10 = clusters too narrow; fewer than 4 = too broad.

Step 4 — Score themes (confidence-weighted)

For each theme, compute:

• Frequency = count of distinct DIRECT sources mentioning it (weight: 3) + count of distinct PROXY sources (weight: 1). Aggregate = direct × 3 + proxy × 1.
• Intensity (1-5) = highest emotional intensity observed. Use [!] markers from interview notes; for tickets/NPS, language strength.
• Strategic fit (1-5) = alignment with client's quarterly priorities (per discovery-context.md section 1 Stage + section 2 Initiative). NOT agency's strategic fit.

Score = Frequency × Intensity × Strategic fit.

Step 5 — Top 3 pain points

The 3 highest-scoring themes. For each, write:

• Who feels it (specific segment from observations)
• How often (daily / weekly / monthly)
• What breaks today (process step or moment)
• What "good" would look like in their words
• Representative quote (verbatim, with [source] tag)
• Confidence (high / med / low based on source mix)

Step 6 — Surprises

Anything that contradicts problem-canvas hypothesis or discovery-context.md section 2 (client's stated initiative). Surprises often outrank confirmed pains in strategic importance — flag them prominently even if low-frequency.

If the client already proposed a solution and the data contradicts it, this is the most important section in the document. Spend extra care here — the proposal/sow-draft skills will need this evidence.

Step 7 — Dropped themes

Low-frequency observations preserved here so they don't get re-discovered later in the engagement.

Step 8 — Confidence summary

End with one paragraph: "Confidence in these insights is ` because . The weakest claim in the matrix is and would benefit from `."

This is critical for outsourcing — the BA must be able to say to the client "we believe X with high confidence; we believe Y with medium confidence and recommend Z to upgrade".

Output

• ./discovery/insight-matrix.md per ./template.md (scored table + top-3 detail + surprises + dropped)
• ./discovery/themes.md (1-page narrative summary, for stakeholder readout — written for client, not internal team)

Append to _log.md: [insight-synthesis | YYYY-MM-DD] sources: ; themes: ; top3: ; confidence: ; surprises: .

Anti-patterns

• Treating proxy evidence as primary. SME hunch ≠ user quote. Confidence weighting must show through in the final ranking.
• Confirmation bias. Re-read problem-canvas hypothesis AFTER clustering, not during.
• Ignoring contradictions. Especially when the client proposed the solution. Hardest section to write; most important.
• Single-source synthesis. If only one source ran (e.g., only secondary research), confidence is structurally limited; say so.
• Deferring "what would good look like" to scoping. Capturing it in user-language now prevents feature-list scope creep later.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: stanislavnianko
• Source: stanislavnianko/product-discovery-claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.