Understanding Claude Code's Full Stack: MCP, Skills, Subagents, and Hooks Explained

Claude Code functions as an AI agent orchestration platform, not just a coding assistant. Simon Willison describes it as "a tool for general computer automation." This article dissects each extensibility layer and when to use it.

The Seven Layers

Model Context Protocol (MCP) connects external tools and systems. Add a Playwright server with claude mcp add playwright npx @playwright/mcp@latest and invoke it via /mcp__playwright__create-test.

CLAUDE.md files provide persistent project memory. They store coding standards, architecture decisions, and context that survives across sessions. See writing-a-good-claude-md for best practices.

Slash commands are user-triggered workflows. They require explicit invocation (/command) and support argument passing via $ARGUMENTS or $1, $2. Use them for repetitive multi-step tasks.

Subagents are specialized AI personalities that run in parallel. They prevent "context poisoning" by handling detailed implementation work in separate conversations. Each subagent gets its own system prompt, tools, and model selection.

Hooks automate event-driven actions. Configure them to trigger on events like PostToolUse to auto-lint after edits. They run shell commands without user confirmation.

Skills activate automatically based on task context. Unlike slash commands, they don't need explicit invocation—Claude detects when a skill applies and uses it.

When to Use What

Need	Solution
External tool integration	MCP servers
Persistent project knowledge	CLAUDE.md
Explicit reusable workflow	Slash command
Parallel specialized work	Subagent
Automatic post-action behavior	Hook
Context-aware capabilities	Skill

Code Snippets

Hook Configuration for Auto-Linting

This triggers automatic linting after any file edit:

"hooks": {
  "PostToolUse": [{
    "matcher": "Edit|Write",
    "type": "command",
    "command": "pnpm lint:fix"
  }]
}

Subagent System Prompt Pattern

Define a specialized agent with scoped tools:

name: security-auditor
tools: Read, Grep, Bash
model: sonnet

Connections

This builds on claude-code-is-a-platform-not-an-app, which introduces the platform mindset. The subagent patterns align with building-effective-agents principles. For practical tips, see claude-code-best-practices and context-engineering-guide.