Interview Guide

Purpose

A discovery interview is most valuable when the interviewer walks in knowing what the graph currently doesn't know. This skill reads the graph and produces a guide pointed at the gaps that would change the most if filled — not a generic discovery questionnaire that any PM could have written.

Each question is tied to a named gap in the graph, so the PM understands why it's on the list. The questions follow interview-craft rules: they ask about past behaviour, not hypothetical futures; they are open-ended and non-leading; they avoid pitching the product.

It does NOT script a sales conversation, prepare answers, or anticipate what the interviewee "should" say. The interviewer enters open; the guide just makes sure they enter pointed at the right gaps.

Procedure

1. Read graph/graph.json. If it does not exist, tell the PM to run

graphify first and stop.

1. Read meta.health. Cold-start branch: if health === "thin",

state that on a thin corpus the gaps are everywhere — the guide should focus on the highest-leverage first gap (usually: which segment has the most weight on the team's current direction?). One well-targeted interview against a thin corpus is more valuable than a long guide chasing every gap.

1. Identify the load-bearing gaps in the graph. Sources of gaps:

a. Segments named in the corpus but represented by ≤1 source. b. Assumption nodes adjacent to a major decision (i.e. the PRD intent doc or memo references them). c. Pains or Needs supported by only one source — single-source pains pointing at scope are high-leverage to triangulate. d. contradicts edges — the graph itself records that two sources disagree; resolving the contradiction often unlocks the decision.

1. If the PM has named a specific segment or persona to interview,

filter the gaps to those that segment alone can fill. If not, pick the top 3 load-bearing gaps in the graph regardless of segment.

1. For each chosen gap, write 1–3 interview questions. Rules:

• Behavioural, not hypothetical. Ask about what happened, not

what would you do if. Bad: "Would you use feature X?" Good: "Walk me through the last time you tried to do X — what did you actually do?"

• Open, not leading. Bad: "Don't you find the import frustrating?"

Good: "Tell me about importing your previous data — what was the experience like?"

• JTBD-anchored. Frame around the user's job, not the product.

"When you were trying to set up the team, what did you need to figure out?" over "What do you think of our onboarding?"

• One question at a time. No compound questions.

1. Add a short warm-up section (2–3 questions to set context and

relax the interviewee) and a short close section (a single "is there anything I didn't ask that I should have?" closer).

1. Emit the output contract.

Output contract

Emit exactly these four sections, in this order.

What this guide is for

One short paragraph: the segment/persona this guide is targeted at (or "any user" if no specific one was named), the load-bearing gaps it aims to close, and an honest note if the corpus is thin.

Warm-up

2–3 short, low-friction questions that set context (their role, the work they were doing, the period the interview covers).

Targeted questions

The main body. For each chosen gap, a block like:

> Gap: [one-line gap, drawn from the graph] > > 1. [question one] > 2. [question two]

Each question follows the rules in step 5. The block makes the gap-to-question linkage explicit, so the interviewer can drop a question if the conversation already covered the gap, or follow up deeper if it landed.

Close

A single closer question (typically "what didn't I ask that I should have?"), plus a one-line reminder to leave space and not pitch.

End with one line: "ask, then listen. the guide points you at the gaps; the conversation is yours to run."

Guardrails

• NEVER include leading questions, sales questions, or product-pitch

questions. If the interviewee should leave the conversation knowing what the product does, this guide has failed. The interview is for learning, not promotion.

• NEVER ask a hypothetical "would you use" or "would you pay for"

question. The Mom Test: hypotheticals get nice-sounding answers that do not predict behaviour. Past behaviour predicts.

• Every targeted question MUST be tied to a named gap in the graph.

If you cannot name the gap, the question is generic and does not belong here — drop it or tighten the gap statement.

• On a thin corpus, do NOT produce a long guide with many questions

per gap. A thin corpus's leverage is in one well-aimed interview; pick the single highest-leverage gap and three sharp questions around it. Quality over quantity, especially when there is little to triangulate against.

• Do NOT script the interviewee's likely answers, prepare counters, or

anticipate objections. Those framings turn an interview into a negotiation.

• Do NOT recommend follow-up product actions in this output. The

guide's job ends when the interviewer has it in hand.

Source & license

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

• Author: Waddling-Penguin
• Source: Waddling-Penguin/mogkit
• License: MIT
• Homepage: https://mogkit.com

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