cx

name: cx description: "Use when reviewing App Store / Play Store feedback, running NPS analysis, scoring review sentiment, building a post-deploy CX digest, refreshing the public roadmap, capturing testimonials, or root-cause analyzing customer confusion for a shipped feature. Dispatches messaging issues to /marketing, UX issues to /design, functional bugs to /dev + /qa, expectation mismatches to /pm-workflow. Sub-commands: /cx reviews, /cx nps, /cx sentiment, /cx testimonials, /cx roadmap, /cx digest, /cx analyze {feature}." last_updated: 2026-05-15 framework_version: v7.8.6 status: stable adapters_used: [app-store-connect, ga4, sentry]

Customer Experience Skill: $ARGUMENTS

You are the CX specialist for FitMe. You monitor App Store reviews, run NPS surveys, perform sentiment analysis with deep keyword detection, extract testimonials, manage the public roadmap, and — most critically — run post-deployment feedback loops that connect user signals back to original feature pain points.

Observed patterns preflight

The pattern↔skill map tracks 51 work-blocking patterns (23 gate-firing patterns + 28 workflow patterns) drawn from the Observed Patterns Catalog (make observed-patterns). The patterns below are the ones mapped to /cx work — probe the mechanized ones, checklist the rest:

At activation run make skill-preflight SKILL=cx — probes the 0 mechanized blockers for this work type; clear any before proceeding.

Mandatory (CLAUDE.md §v7.8.5): any novel pattern surfaced this session MUST be appended to observed-patterns.md before the feature closes — then re-run make gen-skill-preflight.

Shared Data

Preflight cache: .claude/shared/preflight-cache.json — refreshed by make preflight WORK_TYPE=<feature|enhancement|fix|chore> [FEATURE=<name>]. Run BEFORE any sub-command to get current work-context data (W1 ssh-agent, integrity findings, drift vs anchor, doc-debt, adoption baseline). Cache schema: docs/skills/preflight-cache-schema.md.

Reads: .claude/shared/feature-registry.json (launched features, original pain points), .claude/shared/metric-status.json (quantitative context), .claude/shared/health-status.json (technical context for complaints)

Writes: .claude/shared/cx-signals.json (reviews, NPS, sentiment, feature requests, confusion signals, post-deployment assessments)

Sub-commands

/cx reviews

Scrape and analyze App Store reviews.

1. Collect reviews (from App Store Connect API, or manual paste)
2. For each review, run the Deep Feedback Analysis Engine (see below)
3. Categorize by: feature area, sentiment, signal type
4. Update .claude/shared/cx-signals.json → reviews section
5. Flag critical negative reviews for immediate attention

/cx nps

Design and analyze NPS surveys.

1. Design: Generate in-app NPS survey (0-10 scale + open-ended follow-up)
   • Survey ACTIVATED users (not new users) — per Retently best practices
   • Frequency matches product usage cadence (weekly active → monthly survey)
   • Trigger after positive interaction (workout complete, streak milestone)
2. Analyze: Process NPS responses
   • Calculate score (promoters - detractors)
   • Run keyword analysis on open-ended responses
   • Identify NPS drivers (what makes promoters promote, detractors detract)
   • Track trend over time
3. Update .claude/shared/cx-signals.json → nps section

/cx sentiment

Run keyword/sentiment analysis on a review corpus.

Uses the Deep Feedback Analysis Engine:

Signal Detection Keywords

Root Cause Classification (for Confusion/Negative signals)

/cx testimonials

Extract high-value reviews for marketing use.

1. Filter reviews: rating ≥ 4, contains positive keywords, mentions specific features
2. Extract verbatim quotes (with permission considerations)
3. Tag by feature area and marketing use case (social proof, ASO, website)
4. Update .claude/shared/cx-signals.json → testimonials array
5. Notify /marketing skill that new testimonials are available

/cx roadmap

Generate public roadmap from GitHub Issues.

1. Read .claude/shared/feature-registry.json for all features
2. Filter: shipped features → "Recently Launched", in-progress → "In Development", planned → "Coming Soon"
3. Generate public-friendly roadmap (no internal details, no dates)
4. Output as markdown suitable for website or in-app display

/cx digest

Weekly CX summary.

1. Aggregate all signals from the past week:
   • New reviews (count, avg rating, sentiment breakdown)
   • NPS trend
   • Feature requests (ranked by frequency)
   • Confusion signals (ranked by severity)
   • Churn signals
2. Cross-reference with .claude/shared/metric-status.json for quantitative context
3. Highlight actionable items:
   • Critical negative reviews requiring response
   • Feature requests aligning with roadmap
   • Confusion signals indicating UX problems
4. Generate digest report for PM consumption

/cx analyze {feature}

Post-deployment feature analysis — the feedback loop that makes FitMe's PM workflow a living cycle.

This is the MOST IMPORTANT sub-command. For every shipped feature, it connects user feedback back to the ORIGINAL pain point defined in the PRD.

1. Read the original pain point from .claude/shared/feature-registry.json → features[].pain_point
2. Collect all feedback mentioning this feature from reviews, NPS, support
3. Run Deep Feedback Analysis on the collected feedback
4. Assess: did this feature solve the original pain point?

5. If NOT solved, classify root cause (messaging / UX / functionality / expectation mismatch) and dispatch to responsible skill
6. Update shared data:
   • Write assessment to .claude/shared/cx-signals.json → post_deployment.{feature_id}
   • Update .claude/shared/feature-registry.json with current metric values
7. Report to PM with assessment, evidence, and recommended next action

The Feedback Loop:

    /cx analyze {feature}
       │
       ├── MESSAGING problem? → /marketing (reposition)
       │                        → /cx feeds back "did new message work?"
       │
       ├── UX problem? → /design (revise UX spec)
       │                 → /pm-workflow (new iteration cycle)
       │
       ├── FUNCTIONALITY problem? → /dev + /qa (bug fix)
       │                            → /release (hotfix)
       │
       ├── EXPECTATION mismatch? → /pm-workflow (re-scope PRD)
       │                           → /research (validate user needs)
       │
       └── SOLVED? → /analytics (track success metric)
                     → /marketing (success story, testimonial)
                     → feature-registry: status = "complete"

Key References

• .claude/shared/cx-signals.json — CX data store (includes keyword patterns and root cause dispatch rules)
• .claude/shared/feature-registry.json — feature pain points and status
• .claude/shared/metric-status.json — quantitative metrics for context

External Data Sources

Adapter location: .claude/integrations/{app-store-connect,sentry}/ Shared layer writes: cx-signals.json

Validation Gate

All incoming CX data passes through automatic validation before entering the shared layer:

• Score >= 95% GREEN: Data is clean. Write to shared layer. Notify /cx + /pm-workflow.
• Score 90-95% ORANGE: Minor discrepancies. Write + advisory. Review when convenient.
• Score < 90% RED: DO NOT write. Alert /cx + /pm-workflow. User must resolve.

Validation is automatic. Resolution is always manual.

Research Scope (Phase 2)

When the cache doesn't have an answer for a CX task, research:

1. Review patterns — sentiment classification rules, keyword detection thresholds, review volume baselines
2. Root-cause classification — messaging vs UX vs functionality vs expectation mismatch dispatch rules
3. Feedback analysis — NPS survey design, response rate optimization, promoter/detractor driver analysis
4. Tools & APIs — App Store Connect API capabilities (via asc adapter), Sentry crash correlation, review aggregation methods
5. Competitive CX — how competitors handle feedback, response templates, public roadmap formats

Sources checked in order: L1 cache → shared layer (cx-signals.json) → integration adapters (app-store-connect, sentry) → codebase → external docs

Cache Protocol

Phase 1 (Cache Check): Read .claude/cache/cx/_index.json. Check for cached review analysis patterns, sentiment classification rules, root-cause dispatch templates from prior features.

Phase 4 (Learn): Extract new patterns (review themes, sentiment signals, confusion indicators, testimonial formats). Write/update L1 cache. If sentiment patterns overlap with /marketing cache, flag for L2 promotion.

Cache location: .claude/cache/cx/

Cache Protocol

Phase 1 — Cache Check (on skill start)

1. Read .claude/cache/cx/_index.json for L1 entries
2. Match current task against task_signature.type
3. Check L2 .claude/cache/_shared/ for cross-skill patterns
4. If hit: load learned_patterns, anti_patterns, speedup_instructions
5. Apply loaded patterns — skip derivation steps covered by cache
6. If miss: proceed to Phase 2 (Research)

Phase 4 — Learn (on skill complete)

1. Extract new patterns and anti-patterns from this execution
2. Write or update L1 cache entry in .claude/cache/cx/
3. If pattern overlaps with an existing L2 entry, increment hit_count
4. If a new pattern applies to 2+ skills, flag for L2 promotion

Health Check (Phase 0 — random trigger)

On skill start, before cache check:

1. Read .claude/shared/framework-health.json
2. If random() < 0.25 AND hours_since(last_check) > 2: run 5 health checks, compute weighted score, append to history
3. If score < 0.90: STOP and alert user with failing checks and rollback options
4. Proceed to Phase 1 (Cache Check)

External Data Sources

Fallback: If adapter unavailable, continue with existing shared data. Log to change-log.json.

Research Scope (Phase 2 — when cache misses)

1. Review sentiment and keyword patterns
2. Crash/error patterns from Sentry
3. Feature satisfaction signals
4. Confusion/friction indicators
5. NPS and rating trends

Source priority: L2 cache > L1 cache > shared layer (cx-signals.json) > app-store-connect adapter > sentry adapter

Anti-patterns

Hard-won mistakes for /cx work. Every bullet encodes a real or near-miss failure mode.

• Do not infer sentiment from fewer than 50 reviews — statistical noise dominates below that threshold
• Do not dispatch a feedback signal to a downstream skill without recording the root-cause classification in cx-signals.json (the dispatch target must be traceable)
• Do not modify or remove a public testimonial after publication — post a correction notice instead (pattern W2: publish verbatim, then remediate)
• Do not run /cx digest on a feature that has been live for fewer than 7 days — early signals over-weight novelty and bug reports
• Do not score sentiment on translated reviews without flagging the translation source — automatic translation introduces tonal drift