--- name: discovery-interview description: "Use when running discovery interviews — question-bank build, bias audit, insight extraction. Triggers on 'audit my guide', 'extract insights from transcript', 'is my hypothesis falsifiable'." status: active tier: senior source: package domain: product context_spine: [product] --- # discovery-interview ## When to use - A discovery slice has been framed (`customer-research` ran), but the interview guide is still rough or untested. - A transcript exists and the team needs structured insight extraction, not narrative summary. - An interview round produced surprising findings; a bias audit is needed before the team acts on them. Do NOT use for the upstream framing of the discovery slice (frame sentence, recruit criteria, JTBD focal job) — that is [`customer-research`](../customer-research/SKILL.md). Do NOT use for quantitative survey design or scale-bound research. ## Cognition cluster - **Mental model 22 — Data-informed, not data-driven.** Interview data is signal at low N; treat it as evidence to reason with, not a vote count. See [`docs/contracts/mental-models.md`](../../../docs/contracts/mental-models.md) § 22. - **Mental model 15 — Signal vs noise.** A vivid quote from one articulate user can swamp three muted but consistent signals; frequency-rank by distinct people, never by quote count. See `mental-models.md` § 15. - **Mental model 28 — Eisenhower matrix.** Sort post-interview insights into urgent / important quadrants so the team acts on high-importance signals, not the loudest ones. See `mental-models.md` § 28. - **Product context-spine slot.** Read **product** for the focal job + competitor names; do **not** re-derive these inside this skill. See [`context-spine`](../../../docs/contracts/context-spine.md). ## Procedure ### 1. Build the question bank 1. Anchor on the **switch event** from `customer-research`. The first question is always *"Walk me through the day you decided to ..."* — never *"would you ..."*. 2. Three layers: - **Past behaviour** (what did you do? when? alternative considered?) - **Anxiety / habit** (what feared in switching? what habit died?) - **Outcome** (what changed for you? expected? unexpected?) 3. Cap at 8 open questions per 45-min slot. Beyond that, the interview becomes a survey delivered in person. ### 2. Audit the bank for bias Before running, inspect each question and review the bank against the four common biases: - **Leading** — *"Don't you think X is annoying?"* → rewrite as past behaviour. - **Hypothetical** — *"Would you use Y?"* → replace with *"Last time you needed Y, what did you do?"*. - **Confirmation** — every question presupposes the team's hypothesis is correct. At least two questions must be able to **disconfirm** it. - **Recall ceiling** — questions that ask for events ≥ 90 days back produce confabulation; bound the timeframe. A bank that survives the audit unchanged is suspect — re-read. ### 3. Run-time discipline 1. Open with the switch event. Stay silent for 8 seconds after the user finishes; the second answer is usually the truer one. 2. Capture verbatim, not paraphrase. *"It made me anxious"* is data; *"the user expressed concern"* is interpretation. 3. Probe with *"tell me more about X"* on any anxiety / habit mention; do not switch topics until the thread is exhausted. ### 4. Extract insights For each transcript: 1. One quote per insight. Tag: switch / anxiety / habit / expected-outcome / unexpected-outcome / disconfirmation. 2. Frequency-rank by distinct interviewees (≥ 3 = signal; 1 = anecdote). 3. Mark **disconfirmations** explicitly — these are the most valuable rows, because they are the cheapest to ignore. ### 5. Hand back Produce the three artifacts (see `## Output`); hand the disconfirmation log to whoever owns the original hypothesis. ## Related Skills **WHEN to use this** - The discovery slice is framed and the interview guide / transcript is the unit of work. - A round of interviews ran and the insights need structured extraction (frequency-ranked, bias-audited, disconfirmations highlighted). **WHEN NOT to use this** - The slice itself is unframed — start with [`customer-research`](../customer-research/SKILL.md); this skill inherits its frame, never re-derives it. - The signal needs to come from existing artefacts (issues, PRs, errors) rather than a live interview — route to [`voc-extract`](../voc-extract/SKILL.md). - Insights translate into AC for a ticket — hand off to [`refine-ticket`](../refine-ticket/SKILL.md). - The output is a quantitative funnel — route to [`funnel-analysis`](../funnel-analysis/SKILL.md). ## When the agent should load this - "Hilf mir den Interview-Leitfaden auditieren." - "Welche Fragen sind biased?" - "Extract die Insights aus diesem Transkript." - "Wir haben 6 Interviews geführt — was ist Signal, was ist Anekdote?" - "Ist meine Hypothese widerlegbar mit dem aktuellen Frageset?" ## Output 1. **`question-bank.md`** — 8 open questions, each tagged past-behaviour / anxiety-habit / outcome; bias-audit notes per rewritten question. 2. **`insight-log.md`** — one row per insight: quote · interviewee ID · tag · distinct-people frequency. Sorted descending. Verbatim. 3. **`disconfirmation-log.md`** — each row names the original hypothesis, the interview-derived disconfirmation, and the named owner who must respond before the team acts on the round. ## Gotcha - A bank that survived the audit unchanged is rare; usually means the audit was rushed, not that the bank was perfect. - One articulate interviewee biases insight-extraction toward their vocabulary — frequency-rank by people, not quotes. - Disconfirmations are the cheapest insight to ignore and the most valuable to act on; the log exists so they survive the round. ## Do NOT - Do NOT re-derive the frame inside this skill — read the **product** spine slot or hand back to `customer-research`. - Do NOT translate insights into AC inside this skill — that is `refine-ticket`. - Do NOT collapse disconfirmations into "we also heard X" prose; they earn their own log.