Claude AI Kitchen Metaphor Guide

Overview

Understanding Claude’s architecture of Skills, Sub-Agents, and MCP (Model Context Protocol) is easier when viewed through the lens of a professional kitchen. Each component serves a specific role, just like in culinary operations.


The Kitchen Metaphor

🍳 Skills = Kitchen Appliances & Tools

What they are:
Specialized capabilities that Claude can use to perform specific, well-defined tasks. Skills are like kitchen appliances—each designed for a particular purpose.

Examples:

  • Blender (Skill) = “Code Review” skill
    • Designed to blend multiple code files and produce coherent feedback
    • Knows exactly how to handle its specific domain
    • Works efficiently when given the right ingredients (code)
  • Oven (Skill) = “Test Generation” skill
    • Takes raw requirements (ingredients)
    • Transforms them through a defined process
    • Produces consistent, reliable output (tests)
  • Knife (Skill) = “Code Refactoring” skill
    • Precise, surgical cutting of code into better shapes
    • Works best with clear intention and target
    • Produces clean, organized results

Key Insight: Skills don’t think about why they exist—they just do their one thing exceptionally well. A blender doesn’t question whether blending is useful; it blends.


👨‍🍳 Sub-Agents = Specialized Chefs

What they are:
Autonomous agents that can handle complex, multi-step tasks independently. Like specialized chefs in a professional kitchen, each brings expertise in their domain.

Types of Chefs (Sub-Agents):

  • Pastry Chef → “UI Design Agent”
    • Handles frontend development
    • Knows about aesthetics, user experience, responsiveness
    • Works autonomously on design systems
    • Coordinates with other chefs but owns their domain
  • Sous Chef → “Code Quality Agent”
    • Manages code standards and conventions
    • Reviews work from other agents
    • Ensures consistency across the kitchen
    • Makes decisions when guidelines apply
  • Expediter → “Task Orchestration Agent”
    • Coordinates between chefs
    • Ensures timing and order of operations
    • Manages dependencies
    • Communicates when something is ready for next steps

Key Insight: Unlike skills (which are tools), sub-agents are thinking entities. They understand context, can make decisions, and know how to sequence complex operations.


🚚 MCP = Ingredient Suppliers & External Resources

What they are:
The Model Context Protocol (MCP) is how Claude connects to external tools, databases, and data sources. Like a kitchen’s supply chain, MCP brings in raw materials and external expertise.

MCP as Supply Chain:

  • Produce Supplier (MCP Server) → “GitHub Connector”
    • Brings fresh code repositories and issue data
    • Updates in real-time (like daily produce deliveries)
    • Claude queries what it needs without leaving the kitchen
  • Spice Merchant (MCP Server) → “Documentation Lookup”
    • Provides specialized knowledge when needed
    • API documentation, library references
    • Extends Claude’s knowledge beyond training data
  • Equipment Rental Service (MCP Server) → “External Tool Access”
    • Access to tools Claude doesn’t have built-in
    • Database connections
    • Cloud services and APIs
    • Like renting specialized equipment for complex dishes

Key Insight: Without suppliers (MCP), the kitchen is limited to what’s already in the pantry (training data). With MCP, Claude can reach out for real-time information and external capabilities.


How They Work Together

Daily Kitchen Operations (Real-World Example)

Scenario: “Build a complete authentication system”

1. EXPEDITER (Sub-Agent) receives order  
   "Build login flow with JWT tokens and refresh logic"  
   └─ Breaks down task, assigns to specialists  
  
2. BACKEND CHEF (Sub-Agent) takes database design  
   └─ Uses KNIFE SKILL (refactoring existing schema)  
   └─ Uses APPLIANCE SKILL (code generation)  
   └─ Queries MCP for PostgreSQL documentation  
  
3. FRONTEND CHEF (Sub-Agent) takes UI design  
   └─ Uses UI DESIGN SKILL (component patterns)  
   └─ Coordinates with BACKEND CHEF through EXPEDITER  
  
4. CODE QUALITY CHEF (Sub-Agent) reviews both  
   └─ Uses CODE REVIEW SKILL  
   └─ Uses TEST GENERATION SKILL  
   └─ Requests fixes from relevant chefs  
  
5. RESULT  
   Complete authentication system:  
   - Database schema (Backend)  
   - API endpoints (Backend)  
   - Login/register UI (Frontend)  
   - Comprehensive tests (QA)  
   - Documentation (from all chefs)  

What Made This Work:

  • Skills: Provided specialized execution capabilities
  • Sub-Agents: Brought independent thinking and coordination
  • MCP: Accessed real documentation and external resources
  • Together: Created sophisticated outcome > sum of parts

Why This Metaphor Works

Skills (Appliances)

  • Focused: Do one thing well
  • Reliable: Consistent output
  • Fast: No deliberation needed
  • Limited: Can’t handle surprises
  • Rigid: Can’t adapt to context

Sub-Agents (Chefs)

  • Intelligent: Understand context
  • Adaptive: Handle complex situations
  • Coordinating: Work together
  • Slower: Require thinking/planning
  • Expensive: More compute intensive

MCP (Suppliers)

  • Extended reach: Access external info
  • Real-time: Current data, not training data
  • Specialized: Deep integrations
  • Latency: Network calls required
  • Overhead: Setup and configuration

Best Practice

Use Skills for predictable, performance-critical operations. Use Sub-Agents for complex reasoning. Use MCP to extend beyond Claude’s knowledge.


In Claude Code

How You Experience This

Scenario: You ask Claude Code to “Build a REST API with authentication”

1. Interactive Mode (You: Chef Owner)  
   └─ You ask Claude (Head Chef)  
   └─ Claude uses relevant SKILLS immediately  
   └─ Produces code while you watch  
   └─ You can interrupt and redirect  
  
2. Background Mode (Delegated Tasks)  
   └─ You spawn a SUB-AGENT  
   └─ It works independently with SKILLS  
   └─ Queries MCP for documentation  
   └─ Reports back when done  
  
3. Long-Running Tasks (Complex Projects)  
   └─ Multiple SUB-AGENTS coordinate  
   └─ Each uses SKILLS for their domain  
   └─ MCP provides context and data  
   └─ Final result is orchestrated and reviewed  

Configuration Points

AGENTS.md defines:

  • Which SKILLS are available
  • How SUB-AGENTS should coordinate
  • Team preferences and standards

MCP Servers connected:

  • GitHub (code access)
  • Documentation (knowledge)
  • APIs (external capabilities)

Common Questions Answered by This Model

”When should I use a Skill?”

When you have a well-defined, single-purpose task:

  • Code formatting
  • Test generation
  • Documentation writing
  • Code review

In kitchen terms: Use an appliance (blender) when you know exactly what you need (blended smoothie).

”When should I use a Sub-Agent?”

When you need independent thinking and multi-step reasoning:

  • Feature implementation
  • Architecture decisions
  • Complex debugging
  • Project planning

In kitchen terms: Hire a specialized chef when you need someone who can think, adapt, and coordinate.

”When should I use MCP?”

When you need information or capabilities outside Claude’s training data:

  • Real-time code repositories
  • Current API documentation
  • Live database queries
  • External tool integration

In kitchen terms: Call the supplier when you need fresh ingredients that aren’t in your pantry.


The Philosophy Behind the Model

Claude was built with this insight: The best results come from specialization + coordination + external knowledge.

  • Specialization = Skills (focused excellence)
  • Coordination = Sub-Agents (intelligent orchestration)
  • External Knowledge = MCP (real-world data)

This mirrors how professional organizations work:

  • Individual experts with specialized skills
  • Team leaders who coordinate efforts
  • Supply chains that bring in external resources

The result: System intelligence > Individual intelligence



Last updated: February 3, 2026