Work Decomposer

Core Principle

Any intellectual work = Input + Context + Process → Output

If you can formalize these components, you can prompt it. This skill helps decompose complex intellectual work into multi-agent AI systems.

When to Use

Use this skill when:

• User wants to automate recurring analytical or strategic work
• Task involves creating artifacts (reports, strategies, analyses, frameworks)
• Work follows a pattern but requires intelligence (not simple templates)
• Multiple decision points or steps exist
• Examples: competitive analysis, BMC creation, OKR planning, market research, GTM strategy, customer segmentation

Decomposition Workflow

Step 1: Identify Work Components

Ask user to provide or help identify:

1. Input: What information is needed to start?

   • Documents, data, requirements, constraints
   • Example: "List of competitors, their websites"

2. Context: What knowledge enables good execution?

   • Domain knowledge, frameworks, best practices, company specifics
   • Example: "Understanding of ICP, positioning frameworks"

3. Process: What steps are performed?

   • Decision points, analysis phases, synthesis methods
   • Example: "Analyze landing pages → Extract positioning → Compare offers"

4. Output: What artifact is produced?

   • Format, structure, quality criteria
   • Example: "Excel with columns: Competitor, ICP, UVP, Pricing, Positioning"

Step 2: Design Agent Architecture

Based on complexity, choose architecture:

Simple (1-2 agents):

• Single input → Single analysis → Structured output
• Example: Landing page analysis → Extract key elements

Orchestrated (3-5 agents):

• Orchestrator assigns tasks to specialized sub-agents
• Example: Main agent → Discovery agent + Analysis agent + Synthesis agent

Complex (5+ agents):

• Multiple orchestration levels, iterative refinement
• Example: Strategy → Research agents → Analysis agents → Critique agent → Synthesis

Decision heuristic:

• 1 clear step = Simple
• 3-5 distinct subtasks = Orchestrated
• Multiple phases or iterations = Complex

Step 3: Define Agent Roles

For each agent, specify:

Role name: What it does (e.g., "Competitive Positioning Analyzer")

Input: What it receives

• From user, from other agents, or from external sources

Task: Specific instructions

• Be concrete: "Extract ICP indicators from landing page: job titles mentioned, company size signals, pain points addressed"

Output: What it produces

• Format: JSON, table, text, structured list
• Required fields
• Quality criteria

Context/Constraints:

• What it should know or follow
• Examples of good output
• Common pitfalls to avoid

Step 4: Define Data Flow

Map how information moves:

User Input → Agent 1 (discovers competitors) → List of URLs
           → Agent 2 (analyzes each URL) → Raw analysis per competitor
           → Agent 3 (synthesizes) → Structured table
           → Human review → Corrections
           → Agent 4 (refines) → Final output

Specify:

• Where human review/input is needed
• What gets stored/cached vs. regenerated
• Error handling (what if URL is broken, data is missing)

Step 5: Implement Prompt Templates

For each agent, create prompt template:

Orchestrator template:

You are [role]. Your goal: [goal].

Available agents:
1. [Agent name]: [What it does]
2. [Agent name]: [What it does]

User input: {user_input}

Steps:
1. [What to do first]
2. [What to do next]
3. [How to synthesize]

Output format: [Specification]

Sub-agent template:

You are [specific role]. 

Input: {input_from_orchestrator}

Task: [Concrete instruction]

Context: [Domain knowledge, examples, constraints]

Output: [Exact format specification]

Step 6: Specify Quality Gates

Define how to validate:

Self-critique prompts:

• "Review your output against these criteria: [criteria]"
• "What assumptions did you make? Are they justified?"

Validation checks:

• Required fields present
• Data format correct
• Logical consistency
• Example: "All competitors must have at least one pricing tier identified"

Human review points:

• Where expertise matters
• Where errors are costly
• Where creative input is needed

Implementation Patterns

Pattern 1: Sequential Analysis

For work with clear linear steps.

See references/sequential-pattern.md for competitive analysis, market research examples.

Pattern 2: Framework Completion

For work filling structured frameworks (BMC, OKRs, SWOT).

See references/framework-pattern.md for BMC, GTM strategy, OKR examples.

Pattern 3: Iterative Refinement

For work needing multiple passes (strategy, positioning, messaging).

See references/iterative-pattern.md for strategy development, messaging examples.

Pattern 4: Parallel Research

For work with independent research threads.

See references/parallel-pattern.md for multi-source research, due diligence examples.

Output Delivery

Provide user with:

1. Architecture diagram (text-based):

Orchestrator
├── Agent 1: [Role]
├── Agent 2: [Role]
└── Agent 3: [Role]

2. Prompt templates for each agent (ready to use)

3. Implementation guide:

   • Recommended tools (n8n, Python, API calls)
   • Data storage approach
   • How to iterate and improve

4. Test scenario to validate system works

Resources

references/

Pattern libraries with concrete examples:

• sequential-pattern.md - Linear analysis workflows (competitive intel, market research)
• framework-pattern.md - Structured frameworks (BMC, OKRs, GTM)
• iterative-pattern.md - Multi-pass refinement (strategy, positioning)
• parallel-pattern.md - Independent research threads (due diligence, multi-source analysis)

Load specific pattern file based on user's work type.