Session — One run of claude in a directory, from launch to exit. Each session has its own conversation history; project memory persists across sessions.
Context window — The total text the model can hold in mind at once. When it fills up, older content is compacted or dropped.
Token — The unit models use to measure text (~¾ of a word). Usage limits and billing are counted in tokens.
CLAUDE.md — A markdown file Claude reads at every session start. Holds your project’s conventions, commands, and constraints.
Plan Mode — Read-only mode: Claude explores and proposes but won’t edit files or run commands.
Accept Edits — Auto-approves file edits for the session; shell commands still prompt for approval.
Permissions — Rules governing which actions Claude can take without asking. By default it asks for everything. With “accept edits on” only the file can be changed by Claude without approbation, every tools use needs user approval (unless approved in /permissions).
Tool — A capability Claude can invoke: reading a file, editing, running bash, searching the web. Each call appears in the transcript.
MCP — Model Context Protocol — an open standard for plugging external systems (GitHub, Jira, databases) into Claude as tools.
Subagent — A secondary Claude instance spawned for a focused subtask (tests, search, review) in its own isolated context.
Hook — A shell command that runs automatically at a defined point in Claude’s lifecycle (before a tool, after an edit, on session start).
Skill — Packaged instructions that teach Claude a specific workflow. Loaded on demand, not always in context.
Slash command — Any input starting with /. Built-in ones control the session; custom ones are prompt templates in .claude/skills/.