Interviews

Help the founder run the full customer discovery interview loop — from preparing a script to analyzing what was said afterward.

This skill covers two phases:

1. Script management — drafting, refining, and managing the lifecycle of reusable interview scripts.
2. Transcript analysis — after an interview has happened, extracting statements, linking them to hypotheses, reviewing the founder's interviewing technique, and recommending hypothesis state changes.

An interview script is a reusable guide, not a teleprompter: it anchors the conversation around the assumptions being tested while leaving room for the founder to follow threads that open up. Most founders need one script for their primary segment; occasionally a second script for an adjacent segment.

Before you start

Read startup/core.md to load project context (name, seed description, and all fields under ## Core — especially Audience).

Check if startup/interview-scripts/ contains any .md files.

Scaffold the folders if they don't exist yet:

mkdir -p startup/interview-scripts
mkdir -p startup/interviews/transcripts

---

When scripts already exist

Load and understand them for context. Infer intent from the conversation — don't mechanically ask "what do you want to do?" If the founder is:

• Referring to a specific script — load that file, discuss, help edit the relevant section
• Iterating after running interviews — help revise questions that didn't land, add follow-ups that proved useful, tighten or expand as needed
• Drafting a script for a different segment — start a new script (route to the reference file below if they want a guided draft, or draft directly if they know what they want)
• Changing status — read the file, propose the change, write it back
  • draft — still being shaped, not yet used
  • ready — actively in use for interviews
  • retired — superseded or no longer in rotation

When adding or updating scripts, follow the file conventions:

• YAML frontmatter with status (draft, ready, retired), length_minutes (number — 15/30/45/60), and target_persona (one-line segment descriptor)
• H1 heading: script title — segment + focus, e.g. "Freelance designers — invoice chasing"
• ## Target Persona — 2–4 sentences describing who this script is for and why
• ## Opening — what the founder says at the start: purpose, consent to record, framing
• ## Core Questions — numbered questions with optional indented probes
• ## Closing — wrap-up, referrals ask, thank-you
• Optional ## Notes — facilitation tips for the founder

Slug convention: lowercase the title, replace spaces and non-alphanumeric characters with hyphens, collapse multiple hyphens. "Freelance designers — invoice chasing" → freelance-designers-invoice-chasing.

Read before writing, propose before saving, get confirmation.

---

When no scripts exist

Check prerequisites:

• core.md is missing or has no Audience or similar information that provides this information under ## Core: Mention that a defined target audience sharpens the script considerably — suggest filling that in first (via the whats-next skill or directly in core.md). Do not block: if the founder wants to proceed with a rough persona, proceed.

• startup/hypotheses/ is empty or has fewer than 3 hypotheses: Mention that scripts work best when they're built around specific hypotheses to test — suggest invoking the hypotheses skill first. Do not block: if the founder insists on drafting a script now, proceed.

• Prerequisites met (or founder insists): Load the reference file for the guided script-creation conversation:

.claude/skills/interviews/references/initial-interview-script.md

The reference file's instructions take over from this point.

---

When a transcript arrives

After the founder runs an interview, they may bring back a transcript, paste text into chat, or simply recount a conversation from memory. In any of these cases, the workflow is the same: save the source material, dispatch the interview-analyst subagent, then dispatch the hypotheses-manager subagent, surface recommendations to the founder, and route confirmed hypothesis changes through the hypotheses skill.

Recognize this intent when the founder:

• Points at a transcript file under startup/interviews/transcripts/ (or similar)
• Pastes transcript-shaped text (dialogue, Q&A, long monologue) into chat
• Describes a conversation from memory ("I talked to someone at a coffee shop…", "I had a call with a designer and she said…")
• Asks to analyze, review, or extract insights from a completed interview

Load the Layer 2 reference file that owns this workflow end-to-end:

.claude/skills/interviews/references/transcript-analysis.md

The reference file's instructions take over from this point, including the three input branches (file / pasted / recollection), dispatching both subagents, summarizing to the founder, and routing confirmed edits through the hypotheses skill.

Do not extract statements, evaluate hypothesis state, or edit hypothesis files yourself — the reference file describes how to delegate that work to the two subagents.