layered-reasoning（分层推理）

Layered Reasoning

Common Patterns

Pattern 1: 30K → 3K → 300 ft Decomposition (Top-Down)

When: Starting from vision/principles, deriving concrete actions

Structure:

• 30,000 ft (Strategic): Why? Core principles, invariants, constraints (e.g., "Customer privacy is non-negotiable")
• 3,000 ft (Tactical): What? Approaches, architectures, policies (e.g., "Zero-trust security model, end-to-end encryption")
• 300 ft (Operational): How? Specific actions, procedures, code (e.g., "Implement AES-256 encryption for data at rest")

Example: Product strategy

• 30K: "Become the most trusted platform" (principle)
• 3K: "Achieve SOC 2 compliance, publish security reports, 24/7 support" (tactics)
• 300 ft: "Implement MFA, conduct quarterly audits, hire 5 support engineers" (actions)

Process: (1) Define strategic layer invariants, (2) Derive tactical options that satisfy invariants, (3) Select tactics, (4) Design operational procedures implementing tactics, (5) Validate operational layer doesn't violate strategic constraints

Pattern 2: Bottom-Up Aggregation

When: Starting from observations/data, building up to principles

Structure:

• 300 ft: Specific observations, measurements, incidents (e.g., "User A clicked 5 times, User B abandoned")
• 3,000 ft: Patterns, trends, categories (e.g., "40% abandon at checkout, slow load times correlate with abandonment")
• 30,000 ft: Principles, theories, root causes (e.g., "Performance impacts conversion; every 100ms costs 1% conversion")

Example: Engineering postmortem

• 300 ft: "Service crashed at 3:42 PM, memory usage spiked to 32GB, 500 errors returned"
• 3K: "Memory leak in caching layer, triggered by specific API call pattern under load"
• 30K: "Our caching strategy lacks eviction policy; need TTL-based expiration for all caches"

Process: (1) Collect operational data, (2) Identify patterns and group, (3) Formulate hypotheses at tactical layer, (4) Validate with more data, (5) Distill strategic principles

Pattern 3: Layer Translation (Cross-Layer Communication)

When: Explaining same concept to different audiences (CEO, manager, engineer)

Technique: Translate preserving core meaning while adjusting abstraction

Example: Explaining tech debt

• CEO (30K): "We built quickly early on. Now growth slows 20% annually unless we invest $2M to modernize."
• Manager (3K): "Monolithic architecture prevents independent team velocity. Migrate to microservices over 6 months."
• Engineer (300 ft): "Extract user service from monolith. Create API layer, implement service mesh, migrate traffic."

Process: (1) Identify audience's layer, (2) Extract core message, (3) Translate using concepts/metrics relevant to that layer, (4) Maintain causal links across layers

Pattern 4: Constraint Propagation (Top-Down)

When: High-level constraints must guide low-level decisions

Mechanism: Strategic constraints flow down, narrowing options at each layer

Example: Healthcare app design

• 30K constraint: "HIPAA compliance is non-negotiable" (strategic)
• 3K derivation: "All PHI must be encrypted, audit logs required, access control mandatory" (tactical)
• 300 ft implementation: "Use AWS KMS for encryption, CloudTrail for audits, IAM for access" (operational)

Guardrail: Lower layers cannot violate upper constraints (e.g., operational decision to skip encryption violates strategic constraint)

Pattern 5: Emergent Property Recognition (Bottom-Up)

When: Lower-layer interactions create unexpected upper-layer behavior

Example: Team structure

• 300 ft: "Each team owns microservice, deploys independently, uses Slack for coordination"
• 3K emergence: "Conway's Law: architecture mirrors communication structure; slow cross-team features"
• 30K insight: "Org structure determines system architecture; realign teams to product lines, not services"

Process: (1) Observe operational behavior, (2) Identify emerging patterns at tactical layer, (3) Recognize strategic implications, (4) Adjust strategy if needed

Pattern 6: Consistency Checking Across Layers

When: Validating that all layers align (no contradictions)

Check types:

• Upward consistency: Do operations implement tactics? Do tactics achieve strategy?
• Downward consistency: Can strategy be executed with these tactics? Can tactics be implemented operationally?
• Lateral consistency: Do parallel tactical choices contradict? Do operational procedures conflict?

Example inconsistency: Strategy says "Move fast," tactics say "Extensive approval process," operations say "3-week release cycle" → Contradiction

Fix: Align layers. Either (1) change strategy ("Move carefully"), (2) change tactics ("Lightweight approvals"), or (3) change operations ("Daily releases")

Workflow

Use this structured approach when applying layered reasoning:

□ Step 1: Identify relevant layers and abstraction levels
□ Step 2: Define strategic layer (principles, invariants, constraints)
□ Step 3: Derive tactical layer (approaches that satisfy strategy)
□ Step 4: Design operational layer (concrete actions implementing tactics)
□ Step 5: Validate consistency across all layers
□ Step 6: Translate between layers for different audiences
□ Step 7: Iterate based on feedback from any layer
□ Step 8: Document reasoning at each layer

Step 1: Identify relevant layers and abstraction levels (details) Determine how many layers needed (typically 3-5). Map layers to domains: business (vision/strategy/execution), technical (architecture/design/code), organizational (mission/goals/tasks).

Step 2: Define strategic layer (details) Establish high-level principles, invariants, and constraints that must hold. These are non-negotiable and guide all lower layers.

Step 3: Derive tactical layer (details) Generate approaches/policies/architectures that satisfy strategic constraints. Multiple tactical options may exist; choose based on tradeoffs.

Step 4: Design operational layer (details) Create specific procedures, implementations, or actions that realize tactical choices. This is where execution happens.

Step 5: Validate consistency across all layers (details) Check upward (do ops implement tactics?), downward (can strategy be executed?), and lateral (do parallel choices conflict?) consistency.

Step 6: Translate between layers for different audiences (details) Communicate at appropriate abstraction level for each stakeholder. CEO needs strategic view, engineers need operational detail.

Step 7: Iterate based on feedback from any layer (details) If operational constraints make tactics infeasible, adjust tactics or strategy. If strategic shift occurs, propagate changes downward.

Step 8: Document reasoning at each layer (details) Write explicit rationale at each layer explaining how it relates to layers above/below. Makes assumptions visible and aids future iteration.

Guardrails

1. Maintain Consistency Across Layers

Danger: Strategic goals contradict operational reality, or implementation violates principles

Guardrail: Regularly check upward, downward, and lateral consistency. Propagate changes bidirectionally (strategy changes → update tactics/ops; operational constraints → update tactics/strategy).

Red flag: "Our strategy is X but we actually do Y" signals layer mismatch

2. Don't Skip Layers When Communicating

Danger: Jumping from 30K to 300 ft confuses audiences, loses context

Guardrail: Move through layers sequentially. If explaining to executive, start 30K → 3K (stop there unless asked). If explaining to engineer, provide 30K context first, then dive to 300 ft.

Test: Can listener answer "why does this matter?" (links to upper layer) and "how do we do this?" (links to lower layer)

3. Each Layer Should Be Independently Useful

Danger: Layers that only make sense when combined, not standalone

Guardrail: Strategic layer should guide decisions even without seeing operations. Tactical layer should be understandable without code. Operational layer should be executable without re-deriving strategy.

Principle: Good layers can be consumed independently by different audiences

4. Limit Layers to 3-5 Levels

Danger: Too many layers create overhead; too few lose nuance

Guardrail: For most domains, 3 layers sufficient (strategy/tactics/operations or architecture/design/code). Complex domains may need 4-5 but rarely more.

Rule of thumb: Can you name each layer clearly? If not, you have too many.

5. Upper Layers Constrain, Lower Layers Implement

Danger: Treating layers as independent rather than hierarchical

Guardrail: Strategic layer sets constraints ("must be HIPAA compliant"). Tactical layer chooses approaches within constraints ("encryption + audit logs"). Operational layer implements ("AES-256 + CloudTrail"). Cannot violate upward.

Anti-pattern: Operational decision ("skip encryption for speed") violating strategic constraint ("HIPAA compliance")

6. Propagate Changes Bidirectionally

Danger: Strategic shift without updating tactics/ops, or operational constraint discovered but strategy unchanged

Guardrail: Top-down: Strategy changes → re-evaluate tactics → adjust operations. Bottom-up: Operational constraint → re-evaluate tactics → potentially adjust strategy.

Example: Strategy shift to "privacy-first" → Update tactics (end-to-end encryption) → Update ops (implement encryption). Or: Operational constraint (performance) → Tactical adjustment (different approach) → Strategic clarification ("privacy-first within performance constraints")

7. Make Assumptions Explicit at Each Layer

Danger: Implicit assumptions lead to inconsistency when assumptions violated

Guardrail: Document assumptions at each layer. Strategic: "Assuming competitive market." Tactical: "Assuming cloud infrastructure." Operational: "Assuming Python 3.9+."

Benefit: When assumptions change, know which layers need updating

8. Recognize Emergent Properties

Danger: Focusing only on designed properties, missing unintended consequences

Guardrail: Regularly observe bottom layer, look for emerging patterns at middle layer, consider strategic implications. Emergent properties can invalidate strategic assumptions.

Example: Microservices (operational) → Coordination overhead (tactical emergence) → Slower feature delivery (strategic failure if goal was speed)

Quick Reference

Layer Mapping by Domain

Consistency Check Questions

Translation Hints by Audience

Resources

Navigation to Resources

• Templates: Layered reasoning document template, consistency check template, cross-layer communication template
• Methodology: Layer design principles, consistency validation techniques, emergence detection, bidirectional propagation
• Rubric: Evaluation criteria for layered reasoning quality (10 criteria)

Related Skills

• abstraction-concrete-examples: For moving between abstract and concrete (related but less structured than layers)
• decomposition-reconstruction: For breaking down complex systems (complements layered approach)
• communication-storytelling: For translating between audiences at different layers
• adr-architecture: For documenting architectural decisions across layers
• alignment-values-north-star: For strategic layer definition (values → strategy)

Examples in Context

Example 1: SaaS Product Strategy

30K (Strategic): "Become the easiest CRM for small businesses" (positioning)

3K (Tactical): "Simple UI, 5-minute setup, mobile-first, $20/user pricing, self-serve onboarding"

300 ft (Operational): "React app, OAuth for auth, Stripe for billing, onboarding flow: signup → import contacts → send first email"

Consistency check: Does $20 pricing support "easiest" (yes, low barrier)? Does 5-minute setup work with current implementation (measure in practice)? Does mobile-first align with React architecture (yes)?

Example 2: Technical Architecture

30K: "Highly available system with <1% downtime, supports 10× traffic growth"

3K: "Multi-region deployment, auto-scaling, circuit breakers, blue-green deployments"

300 ft: "AWS multi-AZ, ECS Fargate with target tracking, Istio circuit breakers, CodeDeploy blue-green"

Emergence: Observed: cross-region latency 200ms → Tactical adjustment: regional data replication → Strategic clarification: "High availability within regions, eventual consistency across regions"

Example 3: Organizational Change

30K: "Build customer-centric culture where customer feedback drives decisions"

3K: "Monthly customer advisory board, NPS surveys after each interaction, customer support KPIs in exec dashboards"

300 ft: "Schedule CAB meetings first Monday monthly, automated NPS via Delighted after ticket close, Looker dashboard with CS CSAT by rep"

Consistency: Does monthly CAB support "customer-centric" (or too infrequent)? Do support KPIs incentivize right behavior (check for gaming)? Does automation reduce personal touch (potential conflict)?