feedback-analysis（反馈分析）

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.

Six Capability Areas:

1. Sentiment Analysis — Multi-channel emotion detection beyond positive/negative/neutral
2. NPS/CSAT Frameworks — Dual-track score + text analysis with mismatch detection
3. Feature Request Clustering — Group and prioritize by frequency, emotion, and churn correlation
4. Support Ticket Triage — Hierarchical taxonomy-based categorization and routing
5. Churn Signal Detection — Behavioral and textual early warning systems
6. Feedback-to-Roadmap Translation — Convert clusters to ranked product decisions

When to Invoke

Invoke Skill({skill: 'feedback-analysis'}) when:

• Analyzing customer feedback at scale (>50 items)
• Processing NPS/CSAT survey responses
• Triaging support tickets or classifying issues
• Detecting early churn signals in user behavior or feedback
• Prioritizing the product roadmap from customer requests
• Synthesizing qualitative feedback into actionable themes
• Running feedback campaigns and evaluating results

The Six-Phase Process

Phase 1: Multi-Channel Data Unification (MANDATORY FIRST STEP)

Iron Law: All analysis degrades without unified data. Single-channel view creates blind spots.

Channels to unify:
- In-app surveys (NPS, CSAT, CES)
- Support tickets (Zendesk, Intercom, Freshdesk)
- App store reviews (iOS, Android)
- Social mentions (Twitter/X, Reddit, LinkedIn)
- Chat transcripts (live chat, chatbot logs)
- Product reviews (G2, Capterra, Trustpilot)
- Email responses

Output: A unified feedback dataset with source, timestamp, channel, user tier, and raw text per item.

Phase 2: Sentiment Analysis

Method: NLP-based multi-dimensional sentiment classification.

Classifications:

• Polarity: Positive / Negative / Neutral
• Emotion: Frustration, Satisfaction, Confusion, Delight, Urgency
• Intensity: High / Medium / Low
• Topic: Product area, feature, support experience, pricing

Key Pattern — Mismatch Detection (Critical Insight):

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.

NPS Segments:

• Promoters (9-10): Surface testimonials, referral triggers
• Passives (7-8): Identify friction points preventing promotion
• Detractors (0-6): Root cause analysis, recovery playbooks

CSAT Layers:

• Score distribution by product area, agent, date range, cohort
• Text sentiment correlation with CSAT score
• Trend analysis: rolling averages, anomaly detection on score drops
• Causation analysis: what drove score changes?

Multi-Dimensional Segmentation:

Dimensions to segment by:
- User tier (free, pro, enterprise)
- Acquisition channel
- Product area (onboarding, core feature, billing, support)
- Agent/team (for support CSAT)
- Cohort (joined date, plan upgrade date)
- Region/language

Output: NPS/CSAT dashboard data with trend lines, mismatch segments, and causation narratives.

Phase 4: Feature Request Clustering

Method: Group verbatims into themes without manual tagging using pattern detection.

Clustering Dimensions:

• Frequency: How many users requested this?
• Emotional Weight: How strongly do they feel? (high emotion = high priority)
• Churn Correlation: Does absence of this feature correlate with churned users?
• Segment Impact: Which user tiers/cohorts are most affected?

Category Taxonomy:

Request Categories:
- Bug Report (broken functionality)
- Feature Gap (missing capability)
- UX Friction (confusing/slow workflow)
- Performance Issue (speed, reliability)
- Integration Request (connect to other tools)
- Pricing Feedback (too expensive, wrong tier)
- Documentation Gap (can't figure out how to use it)

Prioritization Formula:

Priority Score = (Frequency × 0.3) + (Emotion Weight × 0.3) + (Churn Correlation × 0.4)

Output: Ranked feature request list with evidence count, sentiment weight, and churn correlation per cluster.

Phase 5: Support Ticket Triage

Taxonomy Design: Hierarchical, max 30-50 tags to prevent tag bloat.

Recommended Taxonomy Structure:

Level 1 (Category):        Level 2 (Subcategory):       Level 3 (Root Cause):
- Technical Issue          - Login/Auth                  - Password reset broken
- Billing                  - Charge dispute              - Double-charged
- Feature Usage            - Onboarding                  - Setup wizard unclear
- Performance              - Slow response               - Database timeout
- Integration              - API error                   - Rate limit exceeded
- Account Management       - Team permissions            - Role not propagating

Triage Modes:

1. Rule-based: Fast pattern matching for obvious categories
2. AI-powered: NLP for ambiguous tickets (45% faster routing per Zendesk data)
3. 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)
• Support ticket surge (3+ tickets in 2 weeks)
• Negative sentiment spike in recent feedback
• Admin account downgrade events
• Billing inquiry (pricing-related tickets/queries)
• Export activity (data portability requests)

Churn Risk Scoring:

Churn Risk = (Behavioral signals × 0.4) + (Feedback sentiment × 0.3) + (Support ticket pattern × 0.3)
Risk Tiers:
- High (>0.7): Trigger immediate retention playbook
- Medium (0.4-0.7): Proactive outreach + success check-in
- Low (<0.4): Monitor, standard touchpoints

Reason Code Generation: Each high-risk user gets a human-readable reason code:

• "Feature-gap: Missing [X] blocking workflow"
• "Billing: Perceives pricing as misaligned with value"
• "Support: 3 unresolved critical tickets in 14 days"

Output: Churn risk cohort with risk scores, reason codes, and triggered playbook recommendations.

Phase 7: Feedback-to-Roadmap Translation

Input: Completed phases 1-6 (sentiment, NPS/CSAT, clusters, triage, churn signals)

Prioritization Matrix:

Roadmap Score = (Feature Request Frequency × 0.25)
              + (Churn Correlation × 0.35)
              + (NPS Impact × 0.25)
              + (Support Volume × 0.15)

Stakeholder Output Format:

## Roadmap Recommendation: [Feature/Fix Name]

**Evidence Summary**: [N] users requested this across [channels]
**Sentiment**: [Avg. emotional weight and polarity]
**Churn Correlation**: [% of churned users mentioned this]
**NPS Impact**: [Correlation to Detractor-to-Promoter potential]
**Support Impact**: [Ticket volume and priority distribution]

**Recommended Action**: [Implement / Investigate / Defer / Decline]
**Priority Tier**: P0 / P1 / P2 / P3
**Supporting Quotes**: [3-5 verbatim user quotes]

Continuous Loop: Feed roadmap decisions back into feedback collection ("Did we solve the problem?").

Output: Ranked roadmap items with quantitative evidence, stakeholder narrative, and action recommendations.

Iron Laws

1. ALWAYS unify channels first — single-channel analysis creates blind spots that lead to wrong prioritization decisions.
2. ALWAYS run mismatch detection — numerical scores alone miss high-churn-risk Passives and recoverable Detractors.
3. ALWAYS produce an actionable output — analysis that doesn't result in a decision, playbook trigger, or roadmap item has failed its purpose.
4. NEVER use flat taxonomy with >50 tags — tag bloat causes inconsistent categorization; use hierarchical taxonomy with 30-50 leaf nodes.
5. ALWAYS correlate feature requests with churn data — frequency alone is a poor proxy for priority; churn correlation is the real signal.

Anti-Patterns

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
• Issue found → .claude/context/memory/issues.md
• Roadmap decision → .claude/context/memory/decisions.md

ASSUME INTERRUPTION: If it's not in memory, it didn't happen.