Voice of Customer Synthesizer

---
name: voice-of-customer-synthesizer
description: Use when synthesizing customer feedback from calls, tickets, NPS, or churn interviews into themes. Triggers on "VoC synthesis", "voice of customer", "customer feedback themes", "synthesize NPS comments".
---

# Voice of Customer Synthesizer

Customer feedback is high-signal but noisy. The same complaint shows up in a dozen different framings. The same vague comment ("be better") could mean six different things. This skill finds the actual themes and the actual quotes that prove them.

## Required inputs

1. **The raw inputs** — call transcripts, NPS verbatims, ticket bodies, churn-reason notes
2. **Segment / persona / time window** — synthesizing across SMB and enterprise blends signal
3. **Source breakdown** — how many of each input type (10 NPS + 5 calls + 8 churn reasons is different from 23 NPS)
4. **What action will be taken with this** — informs whether to be roadmap-actionable or directional

## Synthesis framework

### Step 1: Tag every input

For each input, extract 1-3 specific claims. Tag each:
- **[FUNCTIONAL]** — the product doesn't do X / does X badly
- **[USABILITY]** — they can do it but it's hard
- **[RELIABILITY]** — outages, data accuracy, errors
- **[SUPPORT]** — response times, quality of help
- **[ONBOARDING]** — time to value, ramp, initial setup
- **[COMMERCIAL]** — pricing, contract terms, packaging
- **[RELATIONSHIP]** — CSM engagement, exec sponsorship, communication

If a comment is too vague to tag ("the platform should be better"), leave it OUT of the synthesis. Vague isn't signal.

### Step 2: Cluster

Group claims that point to the same underlying thing — even when worded differently.
- "I can't find the export button" + "Reports are hard to share" + "Why is exporting buried 4 clicks deep?" = ONE cluster (export discoverability)
- Don't cluster by feature name — cluster by user job or pain

### Step 3: Apply the signal threshold

A theme requires:
- **3+ distinct sources** (not the same person three times)
- **From different accounts** — five complaints from one customer is one signal, not five
- **Specific enough to act on** — "make it faster" is not a theme, "report builder is unusable for >100k row datasets" is

Single mentions go in **Isolated signals**, not themes. Exception: any mention of safety, compliance, security, or data integrity = always escalate, regardless of N.

### Step 4: Synthesis output

For EACH theme:

#### Theme name
Short, specific. "Export discoverability" not "UX is bad."

#### What customers said
Paraphrase the underlying job-to-be-done, not just the complaint.

#### Source count
"X out of N inputs (Y unique accounts)"

#### Representative quotes (2-3, anonymized)
Actual language from the inputs. This is what makes the synthesis credible.

#### Severity / actionability
- **High severity, high actionability**: clear pattern, fixable in product / process
- **High severity, low actionability**: real pain but ambiguous fix (e.g. "we want a different pricing model")
- **Medium**: theme exists but no urgency
- Flag if **conflicting signals** exist — sometimes 3 customers want the opposite of what 4 others want

#### Recommended action
- Product change → flag for product team with severity
- Process change → CS team owns
- Communication change → marketing / docs owns
- Watch only → not yet enough signal to act

### Step 5: Isolated signals
Outliers that didn't reach the theme threshold. Flag any:
- High-severity even at N=1 (compliance, security, data, regulated industries)
- Repeated by a strategic account (single account, but the account matters)
- Emerging pattern (1-2 mentions but novel — could grow)

### Step 6: What this synthesis CANNOT tell you
Be explicit about limits:
- Survivor bias (NPS skews toward respondents — silent majority unmeasured)
- Selection bias (call transcripts come from CSM-engaged accounts)
- Recency bias (last quarter's incident dominates verbatims)
- Persona blend (if SMB and enterprise are mixed, themes may be SMB-specific)

## Anti-patterns

- "Users want it to be faster, more reliable, and easier to use" — three useless buckets
- Treating one impassioned essay as a theme
- Quoting only the most colorful complaint when 5 calmer ones say the same thing — the calmer ones are MORE signal
- Synthesizing across SMB + enterprise + free tier — different products to different people
- Not naming what's NOT in the data ("nobody complained about pricing in this batch — note that")

## Tone

Honest about the data. If 4 themes are weak and 1 is strong, say so. Don't manufacture themes to fill the deck.

## Output format

Markdown. Sections: Executive summary (3-5 sentences), Methodology (sources, time window, sample size, segment), Themes (priority order), Isolated signals, Limits of this synthesis.