Performance Diagnosis

Find the real cause of underperformance across the growth supply chain: audience, media, creative, CRM, funnel, offer, data, tracking, operations, or market conditions. Use root-cause logic, not superficial reporting.

How This Skill Orchestrates

This is a 3-script diagnostic pipeline — run them in sequence:

1. Script 1 (scripts/funnel_gap_analyzer.py): Trace the conversion funnel stage by stage, detect where the largest drop-off occurs vs benchmarks, identify the primary leakage point, and map initial root causes per stage
2. Script 2 (scripts/root_cause_ranker.py): Take the leakage point + user-provided evidence signals and rank 24 root causes by evidence-weighted likelihood score (prior × evidence multiplier). Covers 8 domains: Media, Creative, Conversion, CRM, Strategy, Data, Operations, Market
3. Script 3 (scripts/corrective_action_prioritizer.py): Generate prioritized corrective actions scored by Impact × Urgency × Feasibility. Classifies as immediate (24-72h), short-term (1-2 weeks), or structural (2-8 weeks)
4. Reference lookup (references/diagnostic_framework.md): Root-cause taxonomy (8 domains), funnel leakage map, 3 diagnostic decision trees (media/CRM/creative), evidence signal catalog, benchmark search protocol
5. Reference lookup (references/corrective_action_playbook.md): Corrective actions by domain, escalation rules (by severity), war room protocol (Hour 0 → Day 14), recovery measurement expectations
6. Web search (Claude — MANDATORY): Search benchmarks before diagnosing — "is this us-specific or market-wide?"
7. Upstream context (Claude): Pull control tower health from weekly-control-tower, media allocation from media-routing-planner, CRM metrics from crm-journey-architect, margin from margin-simulation
8. Visual output (Visualizer): Funnel leakage waterfall + root cause ranking + corrective action priority matrix

Model decision: Anomaly detection (funnel gap vs benchmark) + rule-based classification (root cause ranking) + prioritization scoring (Governance layer). Not ML — we're detecting where the funnel deviates from expected and scoring root causes by evidence weight. The root cause ranker uses a Bayesian-inspired prior × evidence multiplier, but it's a scoring heuristic, not a trained model.