["Unify all feedback channels before any analysis (prevents blind spots)","Detect mismatch between numerical score and text sentiment","Use hierarchical taxonomy for ticket triage (30-50 tags max)","Correlate feature requests with churn risk for roadmap prioritization","Always produce an actionable output — analysis without action has no value"]
error_handling
graceful
streaming
supported
source
builtin
trust_score
100
provenance_sha
db198bf01e37b2d9
Feedback Analysis
Overview
Customer feedback analysis transforms raw feedback into actionable intelligence across six interconnected capability areas. All capabilities share a common data pipeline: unified multi-channel feedback collection feeds sentiment detection, which powers NPS/CSAT scoring, feature clustering, ticket triage, churn signals, and ultimately roadmap prioritization.
A customer scoring 8 (NPS Passive) with deeply negative text is high churn risk.
A customer scoring 6 (NPS Detractor) with positive text is recoverable.
Mismatch = highest priority segment for intervention.
Mismatch Types:
- High score + negative text → At-risk, intervention needed
- Low score + positive text → Recoverable, reduce friction
- Neutral score + high emotion → Emerging issue, monitor closely
Output: Sentiment-tagged dataset with polarity, emotion, intensity, and mismatch flags.
Phase 3: NPS/CSAT Frameworks
Dual-Track Analysis: Process numerical scores AND open-text responses in parallel.
Rule-based: Fast pattern matching for obvious categories
AI-powered: NLP for ambiguous tickets (45% faster routing per Zendesk data)
Human review: Escalation queue for complex/sensitive cases
Priority Scoring:
Ticket Priority = Urgency (language cues) + Impact (user tier/revenue) + Sentiment (frustration level)
- P0: Critical + Enterprise user + High frustration
- P1: High urgency + Any paid user + Negative sentiment
- P2: Medium urgency + Any user + Neutral/negative
- P3: Low urgency + Any user + Neutral
Output: Categorized and prioritized ticket queue with taxonomy assignments and routing rules.
Phase 6: Churn Signal Detection
Behavioral Profile Clustering:
User Engagement Profiles:
- Power User: High session frequency, feature breadth, collaborative
- Dabbler: Irregular sessions, single workflow, no integrations
- One-Feature User: Deep single-feature use, no expansion
- Trial Tourist: Onboarding complete, then disengaged
Early Warning Signals (detect BEFORE explicit churn):
Session frequency drop (>50% decrease week-over-week)
Feature contraction (fewer features used than prior period)
Agents use first matching tag; root cause data is lost
Hierarchical taxonomy, max 50 leaf nodes
Clustering by frequency alone
Missing features that don't come up often but cause 80% of churn
Weight clusters by churn correlation (0.4 weight)
Waiting for explicit churn to detect it
Post-churn analysis doesn't save the customer
Behavioral early warning signals, 14-day detection horizon
Roadmap items without evidence count
Stakeholders can't evaluate priority or trade-offs
Every roadmap item needs: frequency, sentiment weight, churn %, quotes
Single-channel feedback collection
Blind spots by channel; social complaints ≠ support tickets
Unify all channels before analysis
Enforcement Hooks
Input validated against schemas/input.schema.json before execution.
Output contract defined in schemas/output.schema.json.
Pre-execution hook: hooks/pre-execute.cjs
Post-execution hook (observability): hooks/post-execute.cjs
Memory Protocol (MANDATORY)
Before starting:
Read .claude/context/memory/learnings.md
Check for:
Previous feedback analysis results
Known data quality issues in feedback channels
Prior roadmap decisions from feedback
After completing:
New pattern discovered → .claude/context/memory/learnings.md
