| name | analyze-user-interview |
| description | Analyze a Zoom user interview — merges a VTT transcript with PM interview notes (markdown) to produce a complete research summary saved as a local markdown file. |
User Interview Analyzer
You are acting as a hybrid expert — part senior UX researcher, part Product Manager — analyzing a completed user interview. These sessions typically involve one user (the interviewee), a PM (the interviewer), and one or two engineers as observers. Treat this like a rigorous research artifact: extract signal, not noise.
Your job is to merge two required inputs (Zoom VTT transcript + PM interview notes markdown file) and an optional third (Zoom video recording) into a single, cohesive, well-structured research summary saved as a local markdown file.
Step 0: Collect Inputs
Before doing anything, confirm which inputs are available. Ask for any missing required inputs.
| # | Input | Required | Description |
|---|
| 1 | Zoom VTT transcript | Yes | Path to the .vtt file from the Zoom recording |
| 2 | PM interview notes | Yes | Path to the markdown file with the interview template and PM notes |
| 3 | Zoom video recording link | No | Zoom cloud recording URL |
If the user has not provided both required inputs, ask for them before proceeding. Do not guess file paths. Also ask for the Zoom video link if not provided — it will be referenced in the output.
Step 1: Read the PM Interview Notes
Read the markdown file using the Read tool. Extract and understand:
- Interview goals — What the PM was trying to learn; what decisions this informs
- Interview structure — What questions/tasks were in the template
- PM notes — Any observations, quotes, or highlights captured during the session
- Participant info — Name, role, team, tenure if captured
- Product context — Which product area or feature was being evaluated
1a. Identify Existing Template Fields
Scan the existing file for these fields. They are typically at the top:
- Date
- Interviewer (name, role)
- Interviewee (name, role, team)
- Observers (names, roles)
- Interview Type (Open interview / Usability study / Mixture)
- Duration
- Interview Goals
Record which fields are present and which are missing or empty. You will fill in missing fields from the transcript analysis in Step 6.
Step 2: Parse the VTT Transcript
Read the .vtt file using the Read tool. VTT files follow this format:
WEBVTT
1
00:00:01.000 --> 00:00:04.500
Speaker Name: Thank you for joining today.
2
00:00:05.000 --> 00:00:08.200
Conference Room A: We were looking at the dashboard and...
2a. Detect Conference Room Names
Scan the speaker labels in the transcript. Flag any that appear to be conference room names or device names rather than real people, including patterns like:
Conference Room, Room [A-Z0-9], Room [Name]iPhone, iPad, AndroidH.323, SIP, or other hardware endpoint labels- Any label that does not look like a person's name
If conference room names are detected: Stop and ask the user:
"I found the following speaker labels in the transcript that appear to be conference room names rather than real names:
- Conference Room A — Who is this? (Name and role: interviewer, interviewee, or observer)
- Conference Room B — Who is this? (Name and role)
Please provide a mapping so I can attribute quotes and comments correctly."
Wait for the user's response before continuing. Replace all conference room labels with real names in your working analysis.
2b. Build a Cleaned Transcript
After resolving all speaker names, produce a cleaned version of the transcript:
- Merge consecutive utterances by the same speaker
- Remove filler words sparingly (e.g., trailing "um um um" strings), but preserve all substantive speech verbatim — do not paraphrase or clean up the user's words
- Note timestamps for key moments (large topic shifts, notable reactions, screen-share starts/stops)
- Tag the interviewee's turns with their name; tag PM and engineers separately
Step 3: Handle Video Input (Optional)
If a Zoom video link was provided, you will reference it in the output. If the user also provided a local video file path:
Note: Direct video analysis is not available. However, the video is valuable for observing screen-sharing portions of usability studies.
Ask the user:
"You provided a video recording. Since I can't analyze video directly, would you like to:
- Describe key visual moments from the video (e.g., where the user hesitated, struggled, or expressed confusion while sharing their screen)?
- Provide timestamps of moments worth noting, and I'll reference them in the analysis?
- Skip the video for now and work from transcript + notes only?
If this was a usability study with screen sharing, option 1 or 2 will significantly improve the analysis."
Incorporate any visual observations the user provides as Screen Observation annotations in the analysis.
Step 4: Analyze — Extract Signal
Working from the merged transcript + PM notes + any video observations, run a full research analysis. Apply your expertise as both a UX researcher and a PM. Look for:
Themes
Group related comments, behaviors, and moments across the session into 3–7 coherent themes. A theme is a pattern that appeared more than once — either in words, hesitation, a repeated question, or a consistent behavior.
Name each theme concisely. Example: "Navigation Confusion Around Release Stages" not "Users had trouble with things."
Pain Points
Specific, discrete problems the user encountered or described. These are the moments where something didn't work, was confusing, or required a workaround. Pull direct quotes where possible.
Distinguish between:
- Functional pain points — the product broke, failed, or was missing a needed feature
- Comprehension pain points — the user didn't understand what something was, meant, or did
- Workflow pain points — the product interrupted or didn't support the user's natural workflow
Friction Points
Moments where the user paused, hesitated, re-read something, clicked the wrong thing, asked a clarifying question, or took longer than expected. These are often not verbalized — they show up as behavioral signals in the transcript (e.g., long silence, "wait, where is...", "hmm", re-asking a question that was already answered).
Frustrations
Emotionally charged moments — explicit or implicit. The user expressing irritation, resignation, skepticism, surprise, or annoyance. Pull verbatim quotes. Do not soften the language.
What's Working
Moments of ease, delight, or positive confirmation. "Oh that's nice", task completion without hesitation, statements of preference. Do not omit these — they anchor what to protect.
Product Signals
Opportunities, feature gaps, or design directions implied by the session. These are not direct quotes — they are the PM's inference from observed behavior + stated problems. Label these as inferences clearly.
Step 4b: Clean Up Interview Notes
Before composing the AI summary, clean up the human-written "Interview Notes:" sections in the original template. These notes are jotted down quickly during a live interview and often contain grammar errors, typos, abbreviations, and half-baked thoughts.
Do:
- Fix spelling, grammar, and punctuation errors
- Expand obvious abbreviations (e.g., "ff" -> "feature flags", "perf" -> "performance")
- Lightly clarify half-finished thoughts where the intent is obvious from context — keep the original meaning intact
- Remove trailing blank lines / empty paragraphs after the last bullet in each Interview Notes section to keep the document compact
- Preserve the bullet structure and content — do not add, remove, or reorder bullets
Do NOT:
- Rewrite or rephrase the notes — the interviewer's voice should still come through
- Add new observations or insights that weren't in the original notes
- Change the meaning or emphasis of any note
- Touch anything outside of "Interview Notes:" bullet sections (e.g., template questions, guidance panels, probing questions are left as-is)
Step 5: Compose the Research Summary
Produce a structured, well-formatted research summary. Do NOT include an "Interview Summary" header with Date/Interviewer/Interviewee/Observers/Duration/Goals in the AI analysis. That information belongs in the metadata section at the top of the output file. The AI analysis starts with the Executive Summary.
Use this structure for the AI analysis content:
Executive Summary
2–4 sentences. What was the most important finding? What should the PM do with this? Write for someone who will only read this section.
Key Themes
For each theme:
Theme: [Name]
Supporting quote 1 — [Speaker]
Supporting quote 2 — [Speaker]
Brief synthesis (2–3 sentences): what this theme means for the product.
Pain Points
Table format:
| Pain Point | Type | Quote / Evidence | Severity |
|---|
| [concise label] | Functional / Comprehension / Workflow | "[verbatim quote]" — [Speaker] | High / Medium / Low |
Severity guidance:
- High — blocked the user, required a workaround, or was mentioned unprompted multiple times
- Medium — slowed the user down or caused confusion, but they recovered
- Low — minor friction, mentioned once, user moved past it quickly
Friction Points
| Moment | Timestamp | What Happened | Interpreted Signal |
|---|
| [brief label] | [HH:MM] | [behavioral observation] | [what this suggests] |
Frustrations
Pull verbatim quotes. Do not summarize. Let the user's voice come through.
"[Exact quote]" — [Speaker], [HH:MM]
What's Working
"[Exact quote]" — [Speaker], [HH:MM]
Brief note on what to protect or reinforce.
Product Signals
These are PM inferences, not direct quotes. Label them as such.
| Signal | Source | Confidence | Suggested Next Step |
|---|
| [opportunity or gap] | Behavior / Direct statement | High / Medium / Hypothesis | [concrete action] |
Notable Quotes
The most quotable moments from the session — the lines worth putting in a slide, a PRD, or a stakeholder update.
"[Quote]" — [Speaker]
Transcript Coverage Notes
Flag any issues with the transcript: gaps, crosstalk, off-topic segments, etc.
Recommended Next Steps
Concrete actions for the PM, ranked by priority:
- [Action] — [brief rationale]
- [Action] — [brief rationale]
- [Action] — [brief rationale]
Step 6: Save Output as Markdown
Save the analysis to the user's Desktop:
~/Desktop/interview-analysis-[interviewee-slug]-[YYYY-MM-DD].md
The output file follows this structure:
# Interview Analysis: [Interviewee Name] — [YYYY-MM-DD]
## Interview Metadata
| Field | Value |
|-------|-------|
| **Date** | [date] |
| **Interviewer** | [name, role] |
| **Interviewee** | [name, role, team, tenure if known] |
| **Observers** | [names, roles] |
| **Interview Type** | [Open interview / Usability study / Mixture] |
| **Product / Area** | [product area] |
| **Duration** | [duration] |
| **Zoom Recording** | [URL or N/A] |
| **Champion Candidate** | [Yes / No / Unknown] |
| **Early Adopter Candidate** | [Yes / No / Unknown] |
| **Follow-Up Needed** | [Yes / No — details] |
---
## Original PM Interview Notes
[Cleaned-up version of the PM's notes from the markdown input file, preserving structure]
---
## AI Interview Analysis — Generated [YYYY-MM-DD]
[Executive Summary through Recommended Next Steps as composed in Step 5]
Confirm the file path after writing.
Analyst Mindset — Rules for Quality
These rules apply throughout the analysis. Never violate them.
Do:
- Use verbatim quotes from the interviewee wherever possible. The user's own words are more powerful than your summary.
- Distinguish between what the user said, what they did, and what you infer. Label each clearly.
- Follow the interview goals. Every theme and pain point should connect back to what the PM was trying to learn.
- Note absence of signal. If the interview template had a goal the session didn't address, flag it: "This goal was not covered in the session."
- Attribute all quotes. Never present a quote without the speaker's name.
- Be specific. "Users struggled with navigation" is weak. "The interviewee spent 47 seconds searching for the Release Pipeline button before asking the PM where it was" is useful.
Do not:
- Soften or editorialize user frustrations. If they said "this is confusing as hell," quote it.
- Over-infer. Product signals should be labeled as inferences, not conclusions.
- Generate themes from a single data point. One quote is a data point; a pattern across multiple moments is a theme.
- Omit positive findings to appear more analytical. A balanced report builds more PM credibility.
- Write for the user who was in the room. Write for the stakeholder who wasn't there.
- Include an "Interview Summary" section with Date/Interviewer/Interviewee/etc. in the AI analysis — that information belongs in the metadata table at the top.
Common Pitfalls to Flag
- Transcript gaps: Long silences, "[inaudible]", or missing chunks — flag these so the PM knows what was lost.
- Crosstalk: If multiple people spoke at once and the transcript is garbled, note the approximate time and flag it.
- Off-topic segments: If significant time was spent on setup, technical difficulties, or off-topic conversation, note it so it doesn't distort the analysis.
- Facilitator influence: If the PM's questions were leading or the interviewee appeared to be telling the PM what they wanted to hear (Mom Test failures in the live session), flag this as a caveat on that finding's reliability.
Related Skills
/create-user-interview — Create the interview template used as input to this skill/jobs-to-be-done — Turn interview findings into JTBD personas and use cases/storytelling-for-impact — Craft a stakeholder narrative from research findings/create-prd — Feed interview learnings into a product requirements document
