recursive-synthesis

license	Apache-2.0
name	recursive-synthesis
description	Orchestrate multi-agent collaborative document synthesis through 6 phases - Divergence, Synthesis, Commentary, Consolidation, Reality Check, Final Merge. Produces authoritative founding documents from complex multi-perspective inputs. Use for constitutional documents, architecture decisions, organizational charters, or any document requiring rigorous multi-perspective synthesis. Activates on "synthesize document", "multi-agent authorship", "collaborative synthesis", "founding document", "architecture document", "recursive synthesis", "constitutional document", "multi-perspective document". NOT for simple document writing, single-author tasks, quick summaries, or documents that don't require adversarial review.
allowed-tools	["Read","Write","Edit","Bash","Grep","Glob","Task"]
category	Cognitive Science & Decision Making
tags	["recursive-synthesis","analysis","integration","deep-thinking","methodology"]
pairs-with	[{"skill":"team-builder","reason":"Design agent teams for each phase"},{"skill":"dag-planner","reason":"Plan execution waves for parallel phases"},{"skill":"orchestrator","reason":"Coordinate multi-phase execution"}]

Recursive Synthesis

Orchestrate multi-agent collaborative document synthesis through 6 phases to produce authoritative founding documents that require rigorous multi-perspective analysis.

Decision Points

Primary Execution Tree

Is this a founding document (constitution, charter, ADR)?
├── NO → Use single-author writing skills instead
└── YES → Continue to phase selection

Are 10+ perspectives genuinely needed?
├── NO → Use standard collaborative writing
└── YES → Continue to agent selection

Are there irreconcilable tensions expected?
├── NO → Use simplified consensus process
└── YES → Full 6-phase recursive synthesis
    │
    ├── Phase 0: Setup (always required)
    │   ├── Problem scope clear? → NO: Clarify before proceeding
    │   ├── Agents selected for max diversity? → NO: Redesign roster
    │   └── Ground rules established? → NO: Define steel-man rules
    │
    ├── Phase 1: Divergence
    │   ├── All 10 papers received? → NO: Wait/chase missing
    │   ├── Genuine diversity achieved? → NO: Replace echo agents
    │   └── Quality threshold met? → NO: Request rewrites
    │
    ├── Phase 2: Synthesis
    │   ├── Convergence > 50%? → YES: Proceed to commentary
    │   ├── Convergence 25-50%? → Flag high-tension synthesis
    │   └── Convergence < 25%? → Abort: Unsynthesizable problem
    │
    ├── Phase 3: Commentary
    │   ├── All steel-man sections present? → NO: Enforce requirement
    │   ├── Irreconcilable tensions flagged? → YES: Prepare dissenting appendix
    │   └── Constructive amendments provided? → NO: Request specifics
    │
    ├── Phase 4: Consolidation
    │   ├── Soul Document coherent? → NO: Lead Architect revision
    │   ├── Dissenting Appendix honest? → NO: Add minority voices
    │   └── Scope clearly documented? → NO: Define boundaries
    │
    ├── Phase 5: Reality Check
    │   ├── Any "COMPLEX" verdicts? → YES: Major rework required
    │   ├── Multiple "BUILD" verdicts? → YES: Address P0/P1 demands
    │   └── Majority "SHIP" verdicts? → YES: Proceed to final merge
    │
    └── Phase 6: Final Merge
        ├── Constitution authoritative? → NO: Strengthen principles
        ├── Practitioner's Guide actionable? → NO: Add concrete steps
        └── Editorial Notes complete? → NO: Document all decisions

Model Selection Decision Tree

Which agent type?
├── Philosophical/Complex reasoning → Use Opus
├── Domain-specific technical → Use Sonnet  
└── Senior practitioner judgment → Use Opus

Which phase?
├── Phase 1 (Divergence) → Mix: 4 Opus (complex), 6 Sonnet (technical)
├── Phase 2 (Synthesis) → Opus (long context required)
├── Phase 3 (Commentary) → Match Phase 1 model per agent
├── Phase 4 (Consolidation) → Opus (highest judgment)
├── Phase 5 (Reality Check) → Opus (senior practitioner)
└── Phase 6 (Final Merge) → Opus (best writing quality)

Failure Modes

Echo Chamber Collapse

Symptoms: All agents reach similar conclusions despite different prompts; position papers sound alike; no genuine tensions emerge : Agent selection lacks intellectual diversity OR prompts insufficiently enforce perspective differences : Replace 3-4 agents with genuinely opposing viewpoints; strengthen agent persona prompts with specific belief systems; enforce no-cross-talk rules