metrics-tree（指标树）

Metrics Tree

Workflow

Copy this checklist and track your progress:

Metrics Tree Progress:
- [ ] Step 1: Define North Star metric
- [ ] Step 2: Identify input metrics (L2)
- [ ] Step 3: Map action metrics (L3)
- [ ] Step 4: Select leading indicators
- [ ] Step 5: Prioritize and experiment
- [ ] Step 6: Validate and refine

Step 1: Define North Star metric

Ask user for context if not provided:

• Product/business: What are we measuring?
• Current metrics: Any existing key metrics?
• Goals: What does success look like?

Choose North Star using criteria:

• Captures value delivered to customers
• Reflects business model (how you make money)
• Measurable and trackable
• Actionable (teams can influence it)
• Not a vanity metric

See Common Patterns for North Star examples by type.

Step 2: Identify input metrics (L2)

Decompose North Star into 3-5 direct drivers:

• What directly causes North Star to increase?
• Use addition or multiplication decomposition
• Ensure components are mutually exclusive where possible
• Each input should be controllable by a team

See resources/template.md for decomposition frameworks.

Step 3: Map action metrics (L3)

For each input metric, identify specific user behaviors:

• What actions drive this input?
• Focus on measurable, observable behaviors
• Limit to 3-5 actions per input
• Actions should be within user control

If complex, see resources/methodology.md for multi-level hierarchies.

Step 4: Select leading indicators

Identify early signals that predict North Star movement:

• Which metrics change before North Star changes?
• Look for early-funnel behaviors (onboarding, activation)
• Find patterns in high-retention cohorts
• Test correlation with future North Star values

Step 5: Prioritize and experiment

Rank opportunities by:

• Impact: How much will moving this metric affect North Star?
• Confidence: How certain are we about the relationship?
• Ease: How hard is it to move this metric?

Select 1-3 experiments to test highest-priority hypotheses.

See resources/evaluators/rubric_metrics_tree.json for quality criteria.

Step 6: Validate and refine

Verify metric relationships:

• Check correlation strength between metrics
• Validate causal direction (does A cause B or vice versa?)
• Test leading indicator timing (how early does it predict?)
• Refine based on data and experiments

Common Patterns

North Star Metrics by Business Model:

Subscription/SaaS:

• Monthly Recurring Revenue (MRR)
• Weekly Active Users (WAU)
• Net Revenue Retention (NRR)
• Paid user growth

Marketplace:

• Gross Merchandise Value (GMV)
• Successful transactions
• Completed bookings
• Platform take rate × volume

E-commerce:

• Revenue per visitor
• Order frequency × AOV
• Customer lifetime value (LTV)

Social/Content:

• Time spent on platform
• Content created/consumed
• Engaged users (not just active)
• Network density

Decomposition Patterns:

Additive Decomposition:

North Star = Component A + Component B + Component C

Example: WAU = New Users + Retained Users + Resurrected Users

• Use when: Components are independent segments
• Benefit: Teams can own individual components

Multiplicative Decomposition:

North Star = Factor A × Factor B × Factor C

Example: Revenue = Users × Conversion Rate × Average Order Value

• Use when: Components multiply together
• Benefit: Shows leverage points clearly

Funnel Decomposition:

North Star = Step 1 → Step 2 → Step 3 → Final Conversion

Example: Paid Users = Signups × Activation × Trial Start × Trial Convert

• Use when: Sequential conversion process
• Benefit: Identifies bottlenecks

Cohort Decomposition:

North Star = Σ (Cohort Size × Retention Rate) across all cohorts

Example: MAU = Sum of retained users from each signup cohort

• Use when: Retention is key driver
• Benefit: Separates acquisition from retention

Guardrails

Avoid Vanity Metrics:

• ❌ Total registered users (doesn't reflect value)
• ❌ Page views (doesn't indicate engagement)
• ❌ App downloads (doesn't mean active usage)
• ✓ Active users, engagement time, completed transactions

Ensure Causal Clarity:

• Don't confuse correlation with causation
• Test whether A causes B or B causes A
• Consider confounding variables
• Validate relationships with experiments

Limit Tree Depth:

• Keep to 3-4 levels max (North Star → L2 → L3 → L4)
• Too deep = analysis paralysis
• Too shallow = not actionable
• Focus on highest-leverage levels

Balance Leading and Lagging:

• Need both for complete picture
• Leading indicators for early action
• Lagging indicators for validation
• Don't optimize leading indicators that hurt lagging ones

Avoid Gaming:

• Consider unintended consequences
• What behaviors might teams game?
• Add guardrail metrics (quality, trust, safety)
• Balance growth with retention/satisfaction

Quick Reference

Resources:

• resources/template.md - Metrics tree structure with decomposition frameworks
• resources/methodology.md - Advanced techniques for complex metric systems
• resources/evaluators/rubric_metrics_tree.json - Quality criteria for metric trees

Output:

• File: metrics-tree.md in current directory
• Contains: North Star definition, input metrics (L2), action metrics (L3), leading indicators, prioritized experiments, metric relationships diagram

Success Criteria:

• North Star clearly defined with rationale
• 3-5 input metrics that fully decompose North Star
• Action metrics are specific, measurable behaviors
• Leading indicators identified with timing estimates
• Top 1-3 experiments prioritized with ICE scores
• Validated against rubric (score ≥ 3.5)

Quick Decision Framework:

• Simple product? → Use template.md with 2-3 levels
• Complex multi-sided? → Use methodology.md for separate trees per side
• Unsure about North Star? → Review common patterns above, test with "captures value + predicts revenue" criteria
• Too many metrics? → Limit to 3-5 per level, focus on highest impact

Common Mistakes:

1. Choosing wrong North Star: Pick vanity metric or one team can't influence
2. Too many levels: Analysis paralysis, lose actionability
3. Weak causal links: Metrics correlated but not causally related
4. Ignoring tradeoffs: Optimizing one metric hurts another
5. No experiments: Build tree but don't test hypotheses