Customer Interviews skill for Codex

Interviews are done well when every question asks about something the person actually did, the guide serves a written research question tied to a decision, nobody pitched the product during the call, and each interview produces a snapshot within 24 hours that quotes the person rather than paraphrasing them into agreement.

Start from the decision

Write this before any questions:

Decision: <what we will decide with this research, and by when>
Research questions (2-4): <what we need to learn, not what we will ask>
Participants: <segment, behaviour criteria, count>
Out of scope: <topics we will not explore>

A research question (“How do teams decide which tool to adopt for sprint planning?”) is not an interview question. Never read research questions aloud to participants.

Recruiting

• Screen on behaviour, not demographics or self-description: “has run sprint planning in the last 2 weeks” beats “is a project manager”.
• 5-8 interviews per distinct segment usually surface the main patterns; stop when the last two interviews produced no new themes.
• Include people who churned, never activated, or chose a competitor. Only interviewing happy customers produces a love letter, not research.
• Screener: 4-6 questions, at least one that disqualifies, with answer options that do not reveal the “right” answer.

Sending recruiting emails or scheduling messages on the user’s behalf is outward communication: draft them and get approval before anything is sent.

Question rules

Avoid	Why	Ask instead
“Would you use a feature that…?”	Hypothetical; people are poor predictors of their behaviour	“Tell me about the last time you needed to…”
“Don’t you find X frustrating?”	Leading	“How did X go last time?”
“Do you like our dashboard?”	Invites politeness	“Walk me through the last time you opened the dashboard. What were you looking for?”
“How often do you…?”	Invites idealized averages	“When did you last…? And before that?”
“What features do you want?”	Customers design poorly	“What did you try? What did it cost you when it failed?”
Two questions in one	Only one gets answered	Split them

Follow-up probes do most of the work: “Tell me more about that”, “What happened next?”, “Why was that important?”, “What did you do instead?”, “Can you show me?” Silence after an answer is a probe too; wait three seconds before speaking.

For a bank of questions by interview type (problem, switching, churn, usability follow-up, pricing), read references/question-bank.md when drafting a new guide.

Discussion guide structure (45-60 min)

1. Intro (3 min): thanks, purpose in neutral terms, no right answers, consent to record
2. Context (7 min): role, team, how their week works around <topic>
3. Last-time story (20 min): reconstruct the most recent instance step by step
   - trigger, who was involved, tools, workarounds, what went wrong, time and cost
4. Deep dive on research question 2 (10 min)
5. Reactions to concept or prototype, only if needed (10 min) - always last
6. Wrap (3 min): "Anything I should have asked?", who else to talk to, thanks/incentive

Showing a concept earlier contaminates everything after it.

Consent and data handling

• Get explicit consent to record before recording starts; state who will see the recording and how long it is kept.
• Store recordings and transcripts where the organization’s policy says; never paste identifiable transcripts into public tools or repo files that are not access-controlled.
• When summarizing, refer to participants by code (P1, P2) with segment, not by name.
• Respect requests to delete data; deleting files is irreversible, so confirm with the user before doing it.

Interview snapshot (fill within 24 hours)

## P4 - Ops lead, 40-person agency, churned in August
Date: 2026-10-14 | Interviewer: <name> | Recording: <link, access-controlled>

Context: runs weekly resourcing; team uses spreadsheets plus our tool for 5 months
Last-time story: <5-8 bullets of what happened, in order>
Pains (with intensity 1-3 and evidence):
- Rebuilding the plan when someone is sick (3) - "I spent my whole Monday redoing it"
Workarounds: duplicate spreadsheet, Slack thread per project
Goals / jobs: <what they were trying to achieve>
Surprises: <anything that contradicts our assumptions>
Quotes (verbatim):
- "..."
Follow-ups: <questions to test in next interviews>

Summarizing many transcripts

When given transcripts:

1. Produce one snapshot per transcript first; do not jump to cross-interview themes.
2. Keep quotes verbatim with participant code and timestamp if available.
3. Mark anything inferred as inference.
4. Count how many participants support each emerging theme and how many contradict it.
5. Hand the snapshots to $user-research-synthesis for affinity mapping and insights.

Workflow

- [ ] 1. Write decision, research questions, participants
- [ ] 2. Draft screener; get approval before any outreach
- [ ] 3. Draft guide; rewrite every hypothetical or leading question
- [ ] 4. Pilot the guide with one colleague; cut to fit the time
- [ ] 5. Run interviews; snapshot each within 24 hours
- [ ] 6. After every 3 interviews, revise the guide for gaps
- [ ] 7. Stop at saturation; hand off to synthesis

Gotchas

• Interviewing to validate a solution you already love produces confirmation; ask about the problem, and let people describe their own workarounds.
• Compliments and “I would definitely use that” are not data. Commitments (time, money, intro to a decision-maker, sharing real files) are.
• Two interviewers maximum; one leads, one takes notes. More feels like an audit.
• Sales and support calls are not research interviews; the incentives differ.
• Do not offer incentives that bias answers (discounts on the product being discussed); use neutral gift cards where policy allows.

Verification

Scan the final guide: every question in sections 2-4 should be answerable only with a story about something that already happened. Count hypotheticals (“would”, “could you imagine”): zero outside the concept section. Confirm each research question maps to at least one guide section, and every snapshot contains at least two verbatim quotes.