Claude Code & Desktop Glossary

Key terms used across Claude Code, Claude Desktop, and the Resource Hub -- defined concisely for quick reference. Terms link to related concepts where useful.

Agent

A subprocess spawned by the Task tool that runs in its own context window. Agents handle independent work -- reading files, running searches, generating content -- without consuming the main conversation's context budget. The main conversation is the orchestrator; agents are the workers it delegates to. See also: Orchestrator, Context Window.

Auto Memory

Notes Claude writes itself during sessions -- preferences, corrections, recurring context -- stored in ~/.claude/memory/ and loaded at session start. Distinct from CLAUDE.md (which you write). Review and edit via /memory. Auto Memory captures things Claude notices; CLAUDE.md captures things you deliberately want Claude to know.

Autonomy Levels

A graduated permission system that defines what Claude does independently versus what it pauses to ask about. Level 1 (Auto-Do): fix obvious errors silently. Level 2 (Auto-Add): apply professional standards and mention what was added. Level 3 (Auto-Fix): unblock progress and explain the fix. Level 4 (Stop): changes that affect scope, finances, or external commitments -- always ask first. Configured via a rules file in .claude/rules/.

CLAUDE.md

A markdown file placed at the workspace root (or in .claude/) that Claude Code loads automatically at every session start. Defines workspace identity, conventions, file paths, and references to rule files. Keep under 200 lines; delegate detail to rules files. Without it, Claude has no persistent context between sessions.

Claude Code

Anthropic's agentic coding tool. Available as a CLI, VS Code extension, JetBrains plugin, and embedded in Claude Desktop. Reads files, runs commands, edits code, and executes multi-step tasks directly inside your local workspace.

Claude Desktop

Anthropic's native desktop application for macOS and Windows. Full chat interface with Projects, persistent memory, file uploads, MCP server integration, and Cowork (scheduled task automation).

Command

A predefined operational instruction sequence stored in .claude/commands/. Handles system-state workflows -- committing changes, generating briefings, syncing state. Distinguished from skills by intent: commands change system state; skills produce content outputs.

Computer Use

A beta capability (available via the Anthropic API on Opus and Sonnet models) that lets Claude interact with on-screen software the same way a human would -- capturing screenshots, clicking, typing, navigating interfaces. Useful for legacy systems with no API.

Context Budget

How much of Claude's working memory (the context window) is currently being used. Every file read, tool result, and line of conversation consumes context. When the budget runs out, older content is compressed or lost. Managing the context budget -- by using agents, reading summaries instead of full documents, and clearing finished work -- is key to long productive sessions. You can check current usage with /context.

Context Window

The maximum amount of text Claude can hold in memory at once -- conversation history, file contents, tool results combined. Once exceeded, earlier content is compressed or lost. Agent delegation protects main context by offloading heavy work to subagents. See also: Context Budget.

Cowork

Claude Desktop's scheduled task feature (Pro plan and above). Define a task once -- what to do, when to run, where to put the result -- and it runs automatically on a recurring schedule. Requires the Desktop app to be open and the machine to be awake at the scheduled time.

Extended Thinking

A Claude feature that makes the model's step-by-step reasoning visible before it gives a final answer. Useful for complex analysis, multi-step problems, or decisions where you want to verify the logic rather than just the conclusion. Available on Opus and Sonnet models; adds to token usage.

Handoff File

A session summary Claude writes at the end of a working session -- what was completed, what was left unfinished, and the single most important next action. The next session reads the handoff file to resume without re-establishing context from scratch. Handoff files are distinct from state.json: state tracks project status; handoffs capture the narrative of where work stopped. See also: State File.

Hook

A shell script (or HTTP endpoint) that Claude Code runs automatically at defined session lifecycle points: SessionStart, PreToolUse, PostToolUse, PreCompact, Stop, SessionEnd. Used for logging, validation, checkpointing, and notifications. Hooks run outside the conversation -- they execute in the background without consuming context.

Lifecycle Event

A defined point in a Claude Code session where hooks can fire. The main events are SessionStart, PreToolUse, PostToolUse, PreCompact, Stop, and SessionEnd. Claude Code exposes 16+ lifecycle events in total.

MCP

Model Context Protocol -- an open standard developed by Anthropic for connecting AI applications to external tools. An MCP server is a local process that exposes callable tools (calendar queries, email sending, database reads) that Claude can use mid-conversation. Described by Anthropic as "USB-C for AI" -- a universal connector between AI and the tools it needs to work with.

Orchestrator

The main conversation context that coordinates agent work. The orchestrator delegates tasks to subagents, receives their results (file paths and summaries, not full content), and synthesises the final output. The orchestrator stays lean -- it directs and consolidates; it does not do the heavy lifting itself. See also: Agent.

Permission Mode

Controls how Claude Code handles tool call approval. In default mode, Claude prompts for confirmation on sensitive operations. In bypassPermissions mode, pre-approved operations run without prompting. Configure in .claude/settings.json. Most users set commonly used file operations to bypass permissions to avoid repeated confirmation prompts.

Pipeline

A sequence of processing steps where the output of one step feeds directly into the next. In content work: research leads to a brief, which leads to a draft, which leads to editing, which leads to publishing. In data work: extract leads to clean, which leads to transform, which leads to report. Pipelines are typically implemented as skills with sequential steps or as coordinated sequences of agent tasks.

PRIMA

A session memory MCP server that records every Claude Code conversation to a local database. Enables cross-session context retrieval, session resumption, and searchable history of past work. Keeps all data on your machine -- nothing is sent to a remote server.

Project (Claude Desktop)

A named container in Claude Desktop for related conversations. All conversations in a project share the same attached files and instructions. Distinct from a "project" in the CLAUDE.md sense (which refers to a folder of work inside a workspace).

Resource Pack

A downloadable bundle of workshop materials: prompt packs, scripts, reference cards, and setup checklists. Available through the Resource Hub after email verification.

Rule

A granular instruction file in .claude/rules/ that Claude Code loads alongside CLAUDE.md. One concern per file -- communication standards, blocked commands, autonomy levels. All files in the directory auto-load at session start. Rules keep CLAUDE.md short by moving detailed instructions into separate, purpose-named files.

Skill

A reusable instruction set stored as a SKILL.md file in .claude/skills/[name]/. Triggered by a slash command, a skill tells Claude how to perform a recurring task with consistent behaviour every time. Supporting files (templates, rubrics, examples) sit alongside the SKILL.md in the same subfolder. See also: Slash Command.

Slash Command

A trigger starting with / that invokes a skill, command, or built-in Claude Code feature. Built-in commands (/init, /memory, /mcp, /help, /clear, /compact) are reserved; custom skills use distinct names. Slash commands are the primary way to invoke consistent, repeatable processes in Claude Code.

State File

A JSON file (typically .claude/state/state.json) that tracks project status, milestones, priorities, and session history across all projects in a workspace. Acts like a project management dashboard that Claude reads on startup. Unlike a task manager you fill in manually, the state file is read and written by Claude as part of normal session commands (/day, /sync, /night). See also: Handoff File.

Template

A reusable starting point for common file types -- a project folder structure, a document layout, a skill scaffold, a CLAUDE.md skeleton. Templates reduce setup time and ensure consistency. In Claude Code, templates typically live in Documentation/Templates/ and are referenced by skills or commands that create new instances.

Token

The unit AI models use to measure and process text. One token is roughly three-quarters of a word in English (or about four characters). A sentence of 20 words uses approximately 27 tokens. Tokens matter because the context window has a token limit, and API usage is priced per token. In practice: shorter prompts, targeted file reads, and agent delegation all reduce token consumption. See also: Context Window, Context Budget.

Workspace

Any directory opened in Claude Code. Claude has read/write access to all files relative to the workspace root. A CLAUDE.md at the root defines workspace-level instructions. Not to be confused with a GitHub repository (the remote) or a project (a subfolder of work inside the workspace).

Workspace Identity

The metadata that tells Claude which workspace it is operating in -- its name, type (personal, client, or team), and GitHub remote. Stored in .claude/state/state.json under the workspace key. Claude Code uses workspace identity to display the correct session banner, validate that git pushes go to the right remote, and flag when a conversation references a different workspace's context. Without it, Claude cannot prevent cross-contamination between workspaces in a multi-workspace setup.

Worktree

An isolated git worktree used by agents running in isolation mode. The agent gets a temporary copy of the repository to work in without affecting the main working tree. Cleaned up automatically if no changes are made. Useful when you want an agent to experiment or make changes that need review before they land in the main workspace.