Reference

Claude Code glossary

56 terms, in plain English, with links to the lessons that teach them.

A B C D E H L M O P R S T W

$ARGUMENTS: The placeholder in a skill for whatever the user typed after the command. Learn more →
acceptEdits: A permission mode where file edits and common file commands in your working directories run without asking. Learn more →
Agent SDK: The Claude Agent SDK: Claude Code's agent engine as a Python and TypeScript library for building your own agents. Learn more →
Agent teams: Multiple Claude instances working with shared tasks and messaging, coordinated centrally. Learn more →
Agentic loop: The cycle Claude Code runs every turn: gather context, decide, check permissions, act with a tool, observe the result, repeat. Learn more →
allowed-tools: Skill frontmatter that pre-approves specific tools while the skill runs. Learn more →
Auto memory: Notes Claude writes for itself per repository, with a MEMORY.md index loaded at session start. Learn more →
Auto mode: A permission mode where a background classifier reviews actions so Claude can work with few prompts. A safety net, not a guarantee. Learn more →
bypassPermissions: A permission mode that runs everything without asking. For isolated containers and VMs only. Learn more →
Checkpoint: An automatic snapshot of files Claude edits, which you can restore with rewind. It doesn't cover changes made by shell commands. Learn more →
CLAUDE.md: A markdown file of instructions Claude Code loads at the start of every session. Your project's memory. Learn more →
Code intelligence: Language-server plugins that give Claude type errors after edits and precise go-to-definition lookups. Learn more →
Compaction: Summarizing older conversation history to free context space, either automatically or with /compact. Learn more →
Context window: Claude's working memory for a session: instructions, messages, file reads and tool output. Learn more →
context: fork: Skill frontmatter that runs the skill in an isolated subagent context. Learn more →
Deny rule: A permission rule that blocks matching actions in every mode, including bypassPermissions. Learn more →
disable-model-invocation: Skill frontmatter that makes a skill run only when you invoke it with /name. Use it for anything with side effects. Learn more →
dontAsk: A permission mode that denies anything not pre-approved. Ideal for locked-down CI. Learn more →
Dynamic workflow: A script Claude writes that orchestrates many subagents and returns one result. Learn more →
Effort: How much reasoning Claude applies: low, medium (default), high, xhigh or max. Set it with /effort. Learn more →
Exit code 2: In a hook, means 'block': the action is stopped and stderr is sent to Claude as the reason. Learn more →
Explore subagent: A built-in, read-only subagent for searching the codebase at a chosen thoroughness. Learn more →
Headless mode: Running Claude Code non-interactively with claude -p for scripts and CI. Learn more →
Hook: A command (or HTTP call, prompt or agent) that runs automatically at a lifecycle event. Deterministic enforcement. Learn more →
Least privilege: Giving an agent or subagent only the tools and permissions its job needs. Learn more →
Managed settings: Organization-wide settings deployed by IT that can't be overridden. Learn more →
Marketplace: A catalog of plugins defined by .claude-plugin/marketplace.json, added with /plugin marketplace add. Learn more →
Matcher: The filter in a hook config that decides which occurrences fire, such as a tool name like Edit|Write. Learn more →
MCP: Model Context Protocol: an open standard for connecting AI apps to tools and data through servers. Learn more →
MCP Inspector: A browser-based tool for testing an MCP server's tools, resources and prompts. Learn more →
MCP prompt: A user-controlled template from an MCP server, run as a slash command. Learn more →
MCP resource: App-controlled data from an MCP server, attached with an @-mention. Learn more →
MCP tool: A model-controlled action from an MCP server, named mcp__server__tool. Learn more →
Output style: Instructions that change how Claude responds (tone, format, length), such as Explanatory. Learn more →
Permission mode: The overall autonomy setting: default (Manual), acceptEdits, plan, auto, dontAsk or bypassPermissions. Learn more →
Plan mode: A permission mode where Claude researches and proposes a plan but can't edit files. Learn more →
Plugin: An installable bundle of skills, subagents, hooks and MCP servers, with a manifest in .claude-plugin/plugin.json. Learn more →
PostToolUse: A hook event that fires after a tool succeeds. The usual home for formatters and linters. Learn more →
PreToolUse: A hook event that fires before a tool runs and can block it. Learn more →
Progressive disclosure: Keeping SKILL.md short and linking deeper files Claude reads only when needed. Learn more →
Prompt injection: Untrusted content that contains instructions aimed at the agent. Learn more →
Rewind: Restoring code, conversation or both to an earlier checkpoint (Esc Esc or /rewind). Learn more →
Routine: The former name for cloud scheduled tasks, which run Claude Code on a schedule or event in the cloud. Learn more →
Rules (.claude/rules): Markdown instruction files that load every session, or only for matching paths via paths frontmatter. Learn more →
Sandboxing: OS-level limits on what shell commands can write and which network hosts they can reach. Learn more →
Scope (MCP): Where an MCP server is configured: local (default), project (.mcp.json) or user. Learn more →
Scope (settings): Where a setting lives (managed, user, project or local), which determines who it applies to and its precedence. Learn more →
Session: A saved conversation you can name, resume, branch and export. Learn more →
Skill: A folder with a SKILL.md that packages instructions, files and scripts Claude loads on demand. Learn more →
SKILL.md: A skill's main file: YAML frontmatter plus markdown instructions. Learn more →
Slash command: A command typed with /, either built-in or a skill. Learn more →
Stop hook: A hook that runs when Claude tries to end its turn. It can block until checks pass. Learn more →
Subagent: An isolated worker with its own context window, tools and system prompt that returns a summary. Learn more →
Tool: A capability Claude uses to act: Read, Edit, Bash, Grep, WebFetch, Agent, Skill and MCP tools. Learn more →
Transport: How MCP messages travel: stdio for local servers, HTTP for remote ones (SSE is legacy). Learn more →
Worktree: An extra git working directory on its own branch, used to run parallel sessions without collisions. Learn more →