pov-expansion（视角扩展）

POV-Expansion: Cross-Domain Perspective Generation

Generate systematic cross-domain perspectives through boundary spanning research frameworks. Uses structure-mapping theory, convergence tracking, and transfer evaluation to identify analogous problems and transferable solutions.

When to Use

Use this skill when:

• Stuck on a problem - Need fresh perspectives from outside your domain
• Evaluating analogies - Want to assess if approach from field X applies to field Y
• Seeking breakthrough insights - Ready for 2-3 hour deep exploration
• Have structured problem - Can articulate problem clearly enough for abstraction

When NOT to Use

Do NOT use when:

• Exploratory browsing - No specific problem statement (use /research-pipeline instead)
• Time-constrained - Need answer in <1 hour (this workflow takes 2-3 hours minimum)
• Well-understood problem - Already know the solution domain (direct research faster)
• Vague problem - Cannot articulate problem clearly (refine first)

Architecture

Three-tier architecture:

1. Orchestrator: pov-expansion-pm (coordinates workflow, handles exceptions)
2. Specialist Agents: 5 new agents for abstraction, perspective generation, synthesis, transfer evaluation
3. Supporting Skills: Reuses requirements-analyst, fact-checker, devils-advocate, editor, strategist

Pre-flight Validation

Before invoking /pov-expansion, verify:

• Problem statement is specific (not "improve product" but "reduce onboarding time")
• Expected runtime: 2-3 hours (workflow cannot be rushed)
• Dependencies available: requirements-analyst, fact-checker, devils-advocate, editor, strategist skills exist

The 12-Stage Pipeline

Stage 0: Archival Guidelines Review

Owner: pov-expansion-pm (automatic) Duration: 2-5 min Checkpoint: Never (always runs automatically) Session Setup: Creates /tmp/pov-session-{YYYYMMDD-HHMMSS}-{PID}/

Initialize workflow session and extract archival guidelines from project CLAUDE.md.

Process:

1. Create session directory: /tmp/pov-session-$(date +%Y%m%d-%H%M%S)-$$/
2. Read project CLAUDE.md (if exists in working directory or parent)
3. Extract archival guidelines:

- Repository organization (directory structure for cross-domain analyses) - Naming conventions (analysis- prefix for POV expansion outputs) - Git rules (commit after edits, no version-numbered files) - Document structure requirements (Executive Summary, Key Parameters Table) - Citation format (Nature-style inline)

1. Write archival summary to session directory: archival-guidelines-summary.md
2. Store session path in workflow state for downstream agents

Output:

session_setup:
  session_dir: "/tmp/pov-session-{timestamp}-{pid}/"
  archival_summary_path: "{session_dir}/archival-guidelines-summary.md"
  guidelines_found: boolean
  guidelines_source: string  # Path to CLAUDE.md or "defaults"

Quality Gate: Session directory created, archival summary written.

Failure Handling:

• CLAUDE.md not found: Use sensible defaults, log warning
• Session directory creation fails: ABORT (cannot proceed without session isolation)

Stage 1: Problem Refinement

Owner: requirements-analyst (reused) Duration: 15-30 min Receives: Session directory path from Stage 0 Quality Gate: Specific problem statement with measurable criteria

Stage 2: Problem Abstraction

Owner: pov-abstractor-classifier Duration: 15-30 min Quality Gate: Rasmussen hierarchy populated (5/5 levels)

Stage 3: Domain Classification

Owner: pov-abstractor-classifier Duration: 20-45 min (includes user confirmation) Quality Gate: 4 domains proposed (1+ Near, 1+ Mid, 1+ Far), user confirmed

Stage 4: Parallel Perspective Generation

Owner: pov-perspective-analyst (x4 parallel) Duration: 30-60 min (parallel) Quality Gate: 3+ perspectives complete, each with structural analogies

Stage 5: Parallel Verification

Owner: fact-checker (x2 parallel, reused) Duration: 15-30 min (parallel) Quality Gate: No factual errors in central claims

Stage 6: Convergence Analysis

Owner: pov-synthesizer Duration: 20-40 min Quality Gate: Convergence floor met (1+ theme in 2+ reports)

Stage 7: Transfer Evaluation

Owner: pov-transfer-evaluator Duration: 20-40 min Quality Gate: Transfer feasibility scored, Far field warnings added

Stage 8: Master Synthesis

Owner: pov-synthesizer Duration: 45-90 min Quality Gate: All perspectives integrated, convergent insights highlighted

Stage 9: Strategic Assessment

Owner: strategist (reused skill) Duration: 20-40 min Quality Gate: 3+ actionable recommendations with implementation paths

Stage 10: Adversarial Review

Owner: devils-advocate (reused) Duration: 20-40 min Quality Gate: Critical issues identified, mitigations proposed

Stage 11: Editorial Polish

Owner: editor (reused) Duration: 15-30 min Quality Gate: Consistent voice, no substantive errors

Timeout Configuration

Global Workflow Timeout: 8 hours On Global Timeout: Save state, escalate to user

Session Cleanup:

• On successful completion (Stage 11 complete): Delete session directory
• On failure/abort: Retain session directory for debugging (log path to user)

Quality Gate Specifications

Gate 0: Session Initialization (Stage 0)

Pass Threshold: 3/3 checks pass On Failure: ABORT - cannot proceed without session isolation

Gate 1: Problem Abstraction (Stage 2)

Pass Threshold: 4/4 checks pass On Failure: Return to Stage 1 with feedback (max 2 iterations)

Gate 2: Domain Classification (Stage 3)

Pass Threshold: All checks pass On Failure: Re-run with user guidance

Gate 3: Perspective Generation (Stage 4)

Pass Threshold: All checks pass Marginal Pass: 3/4 checks (flag in synthesis) On Failure: Retry failed perspective or proceed with reduced coverage

Gate 4: Transfer Feasibility (Stage 7)

Pass Threshold: 4/4 checks pass Marginal Pass: 3/4 (proceed with caution flag) On Failure: Return to Stage 3 with broadened domains

Gate 5: Synthesis Coherence (Stage 8)

Pass Threshold: All checks pass On Failure: Return to synthesis with feedback

Gate 6: Strategic Actionability (Stage 9)

Pass Threshold: All checks pass On Failure: Add missing elements or skip stage

Session Directory Structure

Location: /tmp/pov-session-{YYYYMMDD-HHMMSS}-{PID}/

/tmp/pov-session-{YYYYMMDD-HHMMSS}-{PID}/
├── archival-guidelines-summary.md  # Stage 0 output
├── workflow-state.yaml             # Workflow state for resume
├── progress.md                     # Progress tracking
├── stage-1-refinement.md           # Stage 1 output
├── stage-2-abstraction.yaml        # Stage 2 output (structured)
├── stage-3-domains.yaml            # Stage 3 output (structured)
├── stage-4-perspectives/           # Parallel outputs
│   ├── near-field-1.md
│   ├── near-field-2.md
│   ├── mid-field.md
│   └── far-field.md
├── stage-5-verification/           # Parallel outputs
│   ├── verification-1.md
│   └── verification-2.md
├── stage-6-convergence.md
├── stage-7-transfer.md
├── stage-8-synthesis.md
├── stage-9-strategic.md
├── stage-10-review.md
├── stage-11-final.md               # Final deliverable
└── failure-log.md                  # Failure tracking

Workflow State Schema

workflow_state:
  workflow_id: string
  problem_statement: string
  started_at: ISO8601
  current_stage: integer  # 0-11
  stages_completed: [integer]

  session:                              # Added for session management
    session_dir: string                 # /tmp/pov-session-{...}/
    archival_guidelines_path: string    # Path to archival-guidelines-summary.md
    guidelines_found: boolean           # Whether CLAUDE.md was found
    guidelines_source: string           # Path to CLAUDE.md or "defaults"
    cleanup_on_complete: boolean        # Default true

  artifacts:
    stage_0: path | null                # archival-guidelines-summary.md
    stage_1: path | null
    stage_2: path | null
    # ... etc
    stage_11: path | null

  status: enum[in_progress, paused, completed, failed]
  last_checkpoint: ISO8601
  error_log: [string]

Session Handling on Resume:

• If session directory exists: Reuse existing session
• If session directory missing: Re-run Stage 0 to recreate (non-destructive)

Error Handling Protocols

Catastrophic Failure Protocol (Stage 4)

When ALL perspective agents fail:

1. Detection: 0 valid outputs from 4 parallel agents
2. Check for partial outputs: Any content > 500 chars is "partial"
3. Present user options:

PERSPECTIVE GENERATION FAILED: All 4 agents failed

Failures:
- Near Field 1: [status]
- Near Field 2: [status]
- Mid Field: [status]
- Far Field: [status]

Options:
(A) Retry SEQUENTIALLY (slower but more stable)
(B) Retry with 2 agents only (Near + Mid fields)
(C) Review partial outputs if available
(D) Abort workflow, preserve problem abstraction

Convergence Floor Protocol

Minimum Convergence Threshold:

• 1+ themes appearing in 2+ reports, OR
• 2+ strategies appearing in 2+ reports, OR
• 1+ structural analogies across 2+ domains

On Zero Convergence:

1. Re-analyze with relaxed matching (related concepts, not exact)
2. Look for "anti-convergence" (all agree something doesn't work)
3. If still zero: Offer portfolio framing (present as independent options, not validated patterns)

User Response Timeout Protocol

Parallel Execution

Stage 4: Perspective Generation (Fork-Join)

parallel_execution:
  stage: 4
  pattern: fork_join
  max_agents: 4

  agent_assignment:
    agent_1: "Near Field 1"
    agent_2: "Near Field 2"
    agent_3: "Mid Field"
    agent_4: "Far Field"

  output_paths:
    agent_1: "/tmp/pov-session-{id}/stage-4-perspectives/near-field-1.md"
    agent_2: "/tmp/pov-session-{id}/stage-4-perspectives/near-field-2.md"
    agent_3: "/tmp/pov-session-{id}/stage-4-perspectives/mid-field.md"
    agent_4: "/tmp/pov-session-{id}/stage-4-perspectives/far-field.md"

  completion:
    minimum_required: 2
    proceed_on_timeout: true
    log_missing: true

Stage 5: Verification (Parallel)

parallel_execution:
  stage: 5
  pattern: parallel
  max_agents: 2

  assignment:
    agent_1: "Near + Mid perspectives"
    agent_2: "Far perspective + cross-check"

Success Criteria

• Diversity: Perspectives span Near, Mid, and Far fields (4 domains minimum)
• Depth: Structural analogies via structure-mapping principles (>=2 per perspective)
• Convergence: Cross-perspective insights tracked with measurable algorithm
• Transferability: Rigorous transfer evaluation with Far field warnings
• Actionability: Strategic assessment with >=3 recommendations and implementation paths