Customer research

Customer Research

What this skill does

PMMs are buried in customer signal — sales calls, reviews, support tickets, interview transcripts, survey responses, community threads — and most of it sits unread. This skill turns that pile into structured, themed insight using a three-stage workflow:

1. Inputs — gather multi-source data
2. AI processing + structure — transcribe, tag, cluster, pattern-detect
3. Human synthesis + decision — themes, implications, what changes downstream

The goal is to compress weeks of synthesis into hours, without losing the qualitative texture that makes customer research valuable.

When to invoke

• The user has a pile of unprocessed research (interview transcripts, sales call recordings, review exports) and needs synthesis.
• The user is starting a positioning, messaging, persona, or launch project and needs grounding evidence.
• The user wants to set up ongoing VoC (voice-of-customer) infrastructure.
• The user is conducting win/loss interviews and needs an analysis framework.
• The user wants to mine sales calls for objection patterns or buyer language.

Prerequisites

1. Read pmm-context.md first when available. If it is missing, offer to run pmm-context. If unavailable or declined, collect a minimum brief covering product, audience, research goal, decision to inform, available sources, and known sample limitations. Mark assumptions and continue.
2. Ask the user for the research goal. Synthesis is goal-shaped — themes that matter for positioning are different from themes that matter for product roadmap.
3. Ask what inputs are available and in what form (transcripts, audio, raw exports, structured data).

The three-stage workflow

Stage 1 — Inputs (multi-source data)

Common input types and how to handle each:

Source	What it tells you	Pre-processing needed
Customer interviews	Motivations, language, JTBD	Transcribe (Otter, Rev, Whisper); anonymize
Sales call recordings	Objections, triggers, buying process	Export from Gong/Chorus with timestamps
Win/loss interviews	Decision criteria, alternatives, what tipped it	Structured interview guide, separate won vs. lost
Support tickets	Friction patterns, feature confusion	Export with category tags, dedupe
Reviews (G2/TR/etc)	Public-facing satisfaction patterns	Export with star ratings + verbatims
Community threads	Unprompted language, peer dynamics	Filter for relevance, capture full thread context
NPS / surveys	Quantified satisfaction + open-ended quotes	Separate quant from qual; treat free-text as interview data
Product usage data	What they actually do vs. what they say	Behavioral cohort by segment

Coverage rule: Aim for at least 2 sources per theme you want to validate. A theme that shows up in interviews AND reviews AND support tickets is real. A theme that shows up only in one source is hypothesis-grade.

Stage 2 — AI processing + structure

This is where AI does the work that used to take weeks. Concrete operations:

• Transcribe audio at scale (Otter, Whisper, Rev). Tag by source, persona, date, deal stage if applicable.
• Code/tag by theme, sentiment, persona, JTBD, stage. Use a consistent taxonomy — see INTERVIEW_GUIDE.md for a starting set.
• Cluster related quotes. Pattern detection across sources.
• Surface anomalies — what's said by a single passionate customer but not in the average? Outliers often signal emerging trends.

Prompt patterns for the AI processing layer:

"Read these N interview transcripts. Extract every mention of [topic]. Cluster by sub-theme. For each cluster, give me 3-5 verbatim quotes with speaker context."

"Across these sales call transcripts, identify the top 10 objections by frequency. For each, give me the actual language used (not a summary) and tag by deal outcome."

"Read these G2 reviews. Cluster by theme. For each cluster, give me sentiment trend over time and 3 representative quotes."

"Compare won-deal call transcripts to lost-deal call transcripts. What language patterns differ? Especially in the first 5 minutes of the call."

Stage 3 — Human synthesis + decision

AI produces themed clusters. The PMM produces meaning. This is where the work matters most.

Synthesis questions:

• What do we now know that we didn't know? (true new insight, not confirmation)
• What changed? (compared to last synthesis cycle)
• What's the implication for positioning, messaging, product, GTM?
• What's our confidence? Multi-source corroboration vs. single-source signal
• What do we need to validate further?

Decisions to produce:

• Do we need messaging updates? Where?
• Should we inform roadmap decisions? Which?
• Is re-segmentation needed? On what basis?
• Are personas drifting? Trigger persona refresh?

Workflow

Step 1 — Define research goal

"What decision will this research inform?" If the user can't answer this, narrow scope before continuing. Generic "let's understand our customers" produces generic output.

Step 2 — Inventory inputs

List available sources, volume per source, age, quality. Identify gaps.

Step 3 — Build the coding taxonomy

Co-create with the user a taxonomy of themes to look for. See INTERVIEW_GUIDE.md for starter set. Keep it short — 8-15 top-level codes plus sub-codes. Too many codes makes synthesis impossible.

Step 4 — AI processing pass

Run the prompts above against the inputs. Produce structured output: themed clusters with quotes.

Step 5 — Human synthesis

Walk through the clusters with the user. For each, ask:

• Is this real? (multi-source corroboration check)
• Is this new? (vs. our prior understanding)
• What does it imply?
• What's the implied action?

Step 6 — Synthesis brief

Produce the output artifact (see below).

Step 7 — Set up ongoing infrastructure (optional)

If the user wants this to be a continuous practice rather than one-time, define:

• Source ingestion automation (Gong → tag → cluster → digest)
• Review cadence (monthly synthesis ritual)
• Distribution (who sees synthesis briefs, when)
• Refresh of the coding taxonomy (quarterly)

Outputs

A. Themed insight clusters

For each theme: name, summary (2-3 sentences), 3-5 verbatim supporting quotes with attribution, cross-source corroboration tag, confidence level (H/M/L).

B. Synthesis brief

A 2-3 page document covering:

• Research goal
• Inputs analyzed (counts, sources, dates)
• Top 5 insights (clusters that matter most for the decision at hand)
• What changed (vs. prior understanding)
• Implications (for positioning, messaging, product, GTM)
• Recommended actions (specific, with owners)
• Open questions (what we don't yet know; what to test next)
• Methodology note (sample size, limitations, biases)

C. Quote bank

Searchable archive of tagged verbatim quotes, with source attribution. Becomes raw material for messaging, content, sales enablement.

D. Monitoring setup (if ongoing)

• Ingestion pipeline
• Review cadence
• Distribution list
• Taxonomy doc

Quality bar

• Quote, don't paraphrase. Voice is the value. "They want better integration" loses everything. "I'm spending 6 hours a week stitching reports together" is the insight.
• Multi-source corroboration tag on every claim. One source = hypothesis. Two = signal. Three+ = pattern.
• Bias check. AI synthesis amplifies whatever you fed it. If your interviews are all closed-won customers, your insights skew positive. Always state sample biases.
• No averages over distinct sub-segments. If two sub-segments emerge, split them. Averages describe nobody.
• Don't trust the model's summary uncritically. Spot-check 10% of quote-to-source mappings. AI hallucinates attributions.

Anti-patterns to refuse

• Synthesizing without a defined research goal
• Treating AI cluster output as the final synthesis (it's the input to synthesis, not the synthesis itself)
• Conflating won-deal and lost-deal sources without segmenting
• Removing source attribution from quotes ("anonymized" should mean "identifier removed", not "source removed")
• Skipping the methodology note on biases