Extract Insights

/extract-insights

Extract insights from interview transcripts, following the insight-extraction skill.

Input: $ARGUMENTS

Workflow

1. Resolve sources. If files or a persona were named, use those transcripts from product/interviews/. With no arguments, use the most recent transcript — or, if the current conversation contains a just-finished interview, use that directly. Real-interview transcripts work too: pasted text or a file path anywhere on disk. When a real transcript lives outside product/interviews/, offer to archive a copy there as {YYYY-MM-DD-HHMM}-{interviewee-slug}.md with a source: real header, so the whole interview corpus stays in one place.

2. Extract 3–5 insights per the insight-extraction skill: grounded in quotes, actionable, quantified when possible, prioritized by impact, in the transcript's language.

3. Cross-interview synthesis. When multiple transcripts are involved, flag insights that recur across interviewees.

4. Save to product/insights/{YYYY-MM-DD-HHMM}-{slug}.md with a header listing sources and source: synthetic or source: real (mixed corpora list both per source). For synthetic sources, include the one-time reminder that synthetic insights are hypotheses to verify with real users.

5. Contrast against the overview's beliefs. Read the unverified beliefs in product/overview.md (skip silently if the file or section doesn't exist) — every scope alike: [product], [opportunity: {slug}], [feature: {slug}]. If an insight confirms or contradicts one, say so in one line, citing the insight, and offer to annotate the belief on its own line in the overview: — confirmed/contradicted/weakened by [insights file] (date) (status keywords stay in English; no status = still unverified). Only real evidence confirms: insights from source: real transcripts can earn confirmed or contradicted; synthetic insights only signal — call the belief promising in conversation, or propose at most weakened when they contradict it, never confirmed. Nothing is written without the user's approval.