Quick reference for the concepts this manual uses most: Skills, the four memory layers, progressive disclosure, Hooks, subagents, and MCP — one sentence each.
Skill — a self-contained folder that teaches Claude how to do a class of task, centered on a SKILL.md; loads automatically on task match, or can be invoked by name in conversation.
SKILL.md — the skill’s entry file. In frontmatter, name sets the command name and description decides when it triggers; the body holds steps and hard-won gotchas.
Progressive disclosure — the loading strategy: ~100 tokens of metadata at startup, the body on task match, references / scripts / assets only when needed.
Marketplace — a skill source added via /plugin marketplace add (like the official anthropics/skills); install from it with /plugin install.
Slash command — a /-prefixed in-session command. Custom slash commands merged into Skills: .claude/skills/deploy/SKILL.md automatically becomes /deploy.
CLAUDE.md — the memory file injected at the start of every session, in four layers: global (~/.claude/CLAUDE.md), project (./CLAUDE.md), directory (src/CLAUDE.md), personal (CLAUDE.local.md).
CLAUDE.local.md — the personal layer at the repo root, gitignored; loads on startup but never ships with the repo.
Lazy loading — how directory-level CLAUDE.md files load: injected only when Claude reads files in that directory, invisible to sibling directories.
@ references — point at other files from CLAUDE.md (@README.md, @docs/architecture.md) instead of restating them; relative/absolute paths, recursive up to 5 levels.
/init · /memory · /clear · /compact — the four memory commands: draft a project CLAUDE.md / inspect and edit all layers / cut off unrelated context / compress a long session to its essentials.
Hook — an action enforced at lifecycle points (like PreToolUse); hard constraints belong in hooks, not in prose.
Subagent — an independent executor defined in .claude/agents/<name>.md; a skills: field in its frontmatter preloads skills for it.
Plugin — a distribution unit installed via /plugin; can bundle Skills, Hooks, subagents, and MCP.
MCP — Model Context Protocol, the standard for connecting external data and systems; wire tools in with MCP, then teach Claude to use them well with a Skill.