SKILL.md
Customer Research
Ground positioning, product, and copy in what customers actually say, not what the team assumes. Verbatim quotes beat paraphrases: "we were drowning in spreadsheets" is copy fuel, "manual process inefficiency" is not.
Setup
Read .agents/product-marketing.md or .claude/product-marketing.md if either exists. Then establish: the goal (messaging, personas, product gaps, churn), what raw material exists, the target segment, and the deliverable wanted. Lead with goal and material, follow up on the rest.
Two modes. Mode 1: analyze assets the user already has. Mode 2: go mine online sources. Most engagements combine both.
Mode 1: Analyze Existing Assets
Per asset type: interview and sales transcripts yield pains, triggers, and exact language. Surveys: segment before concluding, and flag where open-ended answers contradict multiple choice. Support tickets: categorize first (bug vs confusion vs missing feature vs expectation mismatch), they skew toward problems. Win/loss and churn notes: segment by reason, never average across causes. NPS: passives and detractors carry more improvement signal than promoters.
Extract six things from every asset:
- Jobs to be done: functional (the task), emotional (how they want to feel), social (how they want to be seen).
- Pain points: prioritize ones mentioned unprompted with emotional language.
- Trigger events: what changed that started the search (new hire, missed target, embarrassing incident).
- Desired outcomes: success in their words, exact quotes.
- Language: the vocabulary customers use for the problem.
- Alternatives considered: competitors, DIY, hiring, doing nothing.
Then synthesize: cluster by theme, score frequency times intensity, segment by profile, pick 5-10 money quotes per theme, and flag contradictions between what customers say and do.
Mode 2: Mine Online Sources
Communities are where customers speak without a filter. Pick sources by ICP, capture verbatim quotes with source, date, context, sentiment, theme tag, and any profile signals. Read references/sources.md when choosing where to look: it maps ICP types to platforms and gives per-platform extraction tips.
Confidence Guardrails
Label every insight before presenting it:
| Confidence | Criteria |
|---|---|
| High | 3+ independent sources, unprompted, consistent across segments |
| Medium | 2 sources, or prompted only, or one segment only |
| Low | Single source, possible outlier, needs validation |
Weight recent sources more heavily, markets shift. Correct for sample bias: reviewers skew opinionated, tickets skew negative, developer forums skew technical and skeptical. Do not build personas or messaging conclusions from fewer than 5 independent data points per segment.
Deliverables
Offer and confirm before generating: synthesis report (themes, quotes, implications), VOC quote bank, personas, JTBD map, competitive intel summary, or a research gap analysis. Read references/persona-template.md when building personas: it has the structure, the synthesis report template, and the anti-patterns.
Verification
For each top theme, count its independent sources and check the representative quotes are verbatim with citations. Expect every High-confidence theme to show 3+ sources and exact quotes. If a theme cannot cite 3 sources, downgrade its label or gather more before anyone acts on it.
Good vs Bad
Judgment call: reporting a theme.
Bad: "Customers find onboarding difficult." No source, no quote, no segment.
Good: "Onboarding friction, High confidence: appears unprompted in 4 of 6 interviews and 11 tickets, mid-market segment only. 'I gave up on the import twice before support walked me through it.' Implication: guided import flow before more features."
Footguns
- Paraphrasing quotes. Cleaned-up quotes lose the copy value and can distort meaning. Fix: capture exact words, mark edits with brackets.
- Averaging across segments. A persona representing everyone represents no one. Fix: segment first, synthesize within segments.
- Inventing persona details. Filled-in blanks read as data later. Fix: leave unknown fields empty, list them in the gap analysis.
- Treating loud as representative. One angry reviewer is not a theme. Fix: confidence labels on everything, minimum sample enforced.
Completion Checklist
- [ ] Goal, material, segment, and deliverable confirmed
- [ ] All six extraction dimensions covered per asset
- [ ] Every insight carries a confidence label
- [ ] Quotes are verbatim with source and date
- [ ] Persona built only where 5+ consistent data points exist
Any box unchecked: not done. Fix or say so.