interview-guide — From backlog to a discussion guide

Turn the graph's known gaps into a ready-to-use discussion guide for a real research round. This is the skill that closes the loop: synthesis → questions → real interviews → new Signals.

Trigger

/interview-guide Personas/Emma.md      ← guide targeting one persona's segment
/interview-guide                       ← guide from the whole open backlog
/interview-guide <topic>               ← guide scoped to one theme

Inputs — where questions come from

1. Open rows in Research backlog.md matching the scope (persona/theme). Prefer rows whose Kind is qualitative or mixed — a quantitative row ("what share of users…") belongs in analytics or a sized survey, and asking six people for a percentage produces a number that looks like data and isn't. Say so rather than silently dropping the row. Order by Priority where it is set — critical rows earn a place in the guide before major, and minor ones only fill time you have left; a session has room for maybe six real questions, so this column decides which ones they are.
2. Thin claims in the persona's data — anything at L2 (Evidence only, never confirmed in an interview) or L3 (single-source Signal, no external confirmation). Walk the persona's Pains/JTBD/Correlations and their linked files to find them. 2b. Open hypotheses (Hypotheses/ with status: open, matching the scope's feature:/tags) — each contributes questions that could verify or kill the bet's BECAUSE assumption. Mark them targets: Hypothesis <name> so findings can route back.
3. Known unknowns listed in the persona file.
4. Stale data (>3 months) worth re-verifying — re-asking beats assuming.

Output — the guide

A markdown document (offer to save as Research/Guide <persona/topic> <YYYY-MM-DD>.md — create the folder if needed), structured like a real discussion guide:

1. Research goals — 2–4 sentences: what this round should settle, in plain language.
2. Screener hints — who to recruit, from the persona's Product usage + Participants.md coverage gaps (segments under-represented so far).
3. Warm-up (2–3 questions) — low-stakes, from the persona's "Who they are" gaps.
4. Core sections by theme — each question annotated, for the researcher's eyes:
   • Why: which backlog row / thin claim it targets (cite the Signal/Evidence by name)
   • If confirmed: what it upgrades — e.g. "second independent source → moves Lost library from L3 to L4 (validated)"
   • open, non-leading phrasing only; follow-up probes indented underneath
5. Feature probes (optional) — candidate features from Product Context.md phrased as descriptions of behaviour, never internal names.
6. Closing — the "anything else?" space + a reminder to capture verbatim quotes.

Rules

• Non-leading phrasing is a hard requirement — the same anti-sycophancy standard personas are held to applies to the questions we ask real people. "How do you find saved songs?" not "Is your library hard to navigate?".
• Every core question must trace to a concrete gap (backlog row, thin claim, known unknown) — no filler questions.
• The guide states at the top: answers from this round come back via /extract-findings — that's what turns them into Signals and closes backlog rows (/backlog close).
• Generating a guide changes nothing in the graph — it's a plan, not data.