Evidence synthesis

Customer Evidence Synthesis

This skill governs how you turn raw KG query results into evidence that PMs, executives, and engineers can act on. The output should feel like analysis from a skilled researcher, not a database export.

Core Principle

Synthesize, never dump. The value you provide is narrative and interpretation, not raw retrieval. Anyone can pull numbers — what PMs need is someone to tell them what the numbers mean, why they matter now, and what to do about it.

Quote Selection Rules

CRITICAL: Use actual customer words, not AI summaries

Quotes MUST come from nli.content (the customer's verbatim words), NOT fi.content (the AI-generated summary). If a query returns fi.content, it is a summary — do not present it as a customer quote. Always query nli.content AS verbatim for customer-facing quotes.

CRITICAL: Every quote must have a date

Every quote MUST include its nli.record_timestamp as a date. A quote without a date has no temporal context — the reader can't tell if it's from yesterday or 2 weeks ago. Never omit the date.

How many: 3-5 maximum per section

More than 5 quotes becomes a data dump. Fewer than 3 feels cherry-picked. The sweet spot is 3-5 carefully chosen quotes that together tell a complete story. For top themes in reports, aim for 3+ quotes minimum.

Diversity requirements:

• Different subthemes — don't pick 3 quotes about the same narrow issue
• Different time periods — show the problem exists across time, not just one spike
• Different sentiment angles — if possible, include one that shows the severity (frustrated) and one that shows the impact (workaround, churn mention)
• Recency bias toward recent — lead with the most recent quote, but include at least one older quote to show persistence

Selection criteria (prefer quotes that):

1. Illustrate with specific detail — "The CSV import fails every time I have more than 500 rows" > "Import doesn't work"
2. Show business impact — "We had to manually re-enter 200 records" > "This is annoying"
3. Represent the majority pattern — don't lead with an outlier
4. Include context — quotes with timestamps, account info, or channel context are more credible

What to avoid:

• Generic one-word complaints ("Bad", "Broken")
• Quotes that require extensive explanation
• Multiple quotes making the same point
• Quotes with PII or sensitive information — redact if present

Citation format:

Every quote MUST include date AND a clickable "View in Enterpret" link. Use this exact format consistently:

> "{Quote text}"
> — {YYYY-MM-DD} · [View in Enterpret]({citationBaseUrl}{record_id})

Never use "View record" or just "View" — always "View in Enterpret".

Problem Sizing (Required)

Every evidence section MUST include these six dimensions:

1. Volume — how many feedback items mention this? Absolute number + proportion of total feedback if available.

2. Breadth — how many distinct accounts/users? One account filing 50 tickets is different from 50 accounts filing one each. If account data isn't available, say so.

3. Trend (WoW per theme) — direction + percentage for EACH theme, not just overall. "↑ 45% week-over-week" not just "increasing." Include the comparison periods. Every theme table should have a WoW trend column.

4. Recency — when was the most recent mention? "Most recent: 2026-03-06" tells the reader whether this is active right now or tapering off. Include this for each top theme.

5. Duration — is this new (appeared this week) or chronic (6+ months)? New + rising = escalate. Old + stable = systemic.

6. Channel mix + severity implication — where is this feedback coming from? Tickets only vs. tickets + reviews + social carries different severity.

Limitations Section (Required)

Every output MUST include a "What this evidence does NOT tell us" section. This is non-negotiable. Omitting limitations destroys trust with sophisticated PMs who know data has gaps.

Standard limitations to consider:

Limitation	When to include
Sample size	Always if < 50 items. "N=23 — treat as directional."
Channel bias	Always. "This reflects [channels]. Users who don't contact support are invisible."
Segment gaps	When account data is unavailable or incomplete
Correlation ≠ causation	When suggesting impact or root cause
Survivorship bias	When relevant — "This only captures users who stayed long enough to complain."
Taxonomy coverage	When search returned few matches — the problem may exist under different labels
Recency bias	When most evidence is from the last few days of the window

Be specific. "Data has limitations" is useless. "All 23 data points come from support tickets filed in the last 14 days — no app review or survey signal" is useful.

Output Calibration by Context

Different commands need different evidence density:

Context	Evidence Depth	Quote Count	Limitations Depth
Quick lookup (/find)	Concise summary — volume, sentiment, top quotes	2-3	One-line scope
Deep analysis (/analyze)	Full block — this IS the deliverable	3-5	Full section
Root cause (/analyze --rootcause)	Minimal — action-oriented	1-2 most severe	Blast radius caveat
Report — weekly/account (/report)	Their pain vs. market pattern or headlines	2-3	Completeness caveat

Narrative Structure

When writing evidence narrative (not tables), follow this arc:

1. Lead with the finding — what you discovered. Not how you discovered it.
2. Size it — volume, trend, breadth.
3. Interpret it — what does this mean for the product?
4. Support it — quotes that prove the interpretation.
5. Bound it — what we don't know.

Bad: "I queried the KG for themes matching 'checkout' and found 340 results with 78% negative sentiment across 3 subthemes..."

Good: "Checkout friction is the fastest-rising customer pain point this month — 340 reports, up 45% WoW, with 78% negative sentiment. It's concentrated in mobile web where users report the payment form resets on back-navigation..."

JTBD Signal Extraction

When customer feedback contains signals about underlying jobs, extract them in this format:

"When [situation], I want to [action], so I can [outcome]"

How to identify JTBD signals:

• Workflow descriptions — customers explaining what they're trying to accomplish
• Workaround mentions — reveals the job they're trying to get done despite the tool
• Comparison language — "In [other tool], I could..." reveals the job being hired for
• Goal statements — "I need to be able to..." or "We're trying to..."
• Success criteria — "It would be great if I could just..."

JTBD extraction rules:

1. Extract no more than 2-3 JTBDs per evidence section — focus on the strongest signals
2. Ground each in a specific quote — don't fabricate JTBDs from thin evidence
3. Label inferred JTBDs as HYPOTHESIS — "Inferred JTBD (HYPOTHESIS): When [X]..."
4. Note how many customers expressed similar intent — a single customer's JTBD is an anecdote, 10 is a pattern

Insight Types

The KG categorizes feedback insights by type. Use this to interpret the nature of the signal:

Type	What It Means	How to Use
COMPLAINT	Active frustration with current state	Strongest signal for prioritization
IMPROVEMENT	Suggestion for enhancement	Feature request signal
QUESTION	Confusion or lack of understanding	Documentation/UX clarity signal
PRAISE	Positive reinforcement	Protect/amplify what's working

A theme that's 90% complaints is different from one that's 60% questions — the first needs a fix, the second needs better education.