| name | flywheel-metrics |
| description | Metrics and analytics for measuring the health and effectiveness of the data flywheel |
| version | 1.0.0 |
| triggers | ["analytics_review","flywheel_health_check"] |
Flywheel Metrics Skill
Purpose
Track the health and effectiveness of the data flywheel. Measure what matters, ignore vanity metrics.
North Star Metric
Appeal success rate improvement over time. If the flywheel works, appeals generated with Denali should have a higher success rate as more outcome data flows in.
Key Metrics
| Metric | Formula | Target (Month 3) | Target (Month 6) |
|---|
| Total outcomes collected | COUNT(appeal_outcomes) | 100 | 1,000 |
| Outcome response rate | responded / sent followups | 25% | 35% |
| Active SHIP counselors | Users WHERE role=counselor AND active < 30d | 10 | 30 |
| Provider practices | COUNT(provider_practices) | 1 | 3 |
| Success rate (overall) | approved / total outcomes | Baseline | +5% |
| Unique CPT+CARC combos | DISTINCT pairs in outcomes | 20 | 100 |
| Flywheel-informed appeals | Appeals where flywheel data was injected | 0% | 30% |
When Flywheel Data Is "Good Enough"
The system only injects flywheel data into Claude's prompt when:
- At least 3 cases exist for that CPT+CARC combination
- The data is from the last 12 months
Below these thresholds, the data is too sparse to be reliable.
Implementation Files
| File | Purpose |
|---|
flywheel_metrics materialized view | Aggregated outcome data |
get_flywheel_context() RPC | Query flywheel data for specific CPT+CARC |
refresh_flywheel_metrics() RPC | Nightly refresh |
supabase/functions/refresh-flywheel-metrics/ | Cron edge function |
app/src/lib/learning.ts | getFlywheelContext(), buildFlywheelPromptInjection() |
app/src/lib/skills-loader.ts | Injects flywheel data into system prompt |
