Glossary: Claude Code Building Blocks

All the pieces that make up Claude Code's customization system. Easy to get confused by the terminology — here's a clear breakdown of every piece and how they relate to each other.

CLAUDE.md

A markdown file with instructions that load into every single conversation. Always present, always using context. Use for global project rules that should never be forgotten. Think of it as the constitution of your project.

Skills

Markdown files with specialized knowledge that load on demand. Claude reads the description, decides if it's relevant, and only then loads the full content. Use for task-specific expertise. Think of it as a reference book Claude pulls from the shelf when needed.

Slash commands

Shortcuts you type manually (like /review or /commit). They trigger predefined prompts. The difference with skills is that commands require you to type them — skills activate automatically.

Hooks

Scripts that run automatically on system events (file saved, tool called, session started). They don't add knowledge — they execute code. Use for validation, linting, blocking dangerous actions. Think of them as security guards that watch every action.

Subagents

Specialized AI assistants that Claude creates within your session. Each has its own isolated context, system prompt, and tool permissions. Claude delegates a task, the subagent works independently, and returns the result. Think of them as temporary contractors hired for a specific job.

Agent Teams

Multiple Claude Code sessions running in parallel, coordinating across separate terminals. One level above subagents. Think of it as a whole department working together.

MCP servers

External connections that give Claude new capabilities (Figma, Notion, Playwright, GitHub). They don't add knowledge — they add tools. Think of them as plugins that connect Claude to the outside world.

Plugins

Bundled packages that can contain skills, subagents, and commands together. Distributable via marketplaces. Think of them as an app you install that brings multiple features at once.

How they work together

CLAUDE.md sets the global rules. Skills provide on-demand expertise. Hooks enforce automated checks. Subagents handle delegated work. MCP connects external tools. Plugins package and distribute all of the above. Each solves a different problem — use them in combination, not as alternatives to each other.