Claude Code Glossary: 20+ Terms in Plain Human English

How it thinks: the core mechanics

1. 1Token — the small chunk of text AI actually reads and counts; like syllables for a model, and everything is billed and measured in them.
2. 2Context window — the working memory of one session, holding your chat, files, and instructions; a desk that only fits so many papers.
3. 3Compaction — automatic summarizing when that desk gets full: old papers are condensed into a one-page brief so work can continue.
4. 4Extended thinking — visible step-by-step reasoning before the answer, shown in gray italics; the model showing its rough draft.
5. 5Effort level — a dial for how hard the model thinks on each reply; low is a quick glance, high is a deep stare at the whiteboard.
6. 6Agentic coding — a workflow where the AI acts (reads, runs, edits) while you supervise, instead of just advising you in chat.
7. 7Agentic loop — the repeated cycle behind every task: gather context, act, check the result, repeat until done. Like cooking by tasting.

How you extend it

1. 1Command — a reusable instruction you trigger by typing /name; a speed-dial button for a prompt.
2. 2Skill — a folder with instructions Claude loads when relevant or when you call /skill-name; a recipe card it pulls out at the right moment. Skills are the modern successor to custom commands.
3. 3Hook — a handler that fires automatically at fixed moments, like after every file edit; a doorbell that always rings, no AI discretion involved.
4. 4MCP (Model Context Protocol) — an open standard plugging Claude into outside services like Slack, Jira, or databases; a universal power adapter for tools.
5. 5Plugin — a bundle of skills, hooks, agents, and MCP servers installed as one unit; an IKEA flat-pack of extensions.
6. 6Output style — a setting that changes Claude's tone and behavior by editing its system prompt; same brain, different uniform.

How the work is organized

1. 1Session — one conversation tied to one folder, with its own memory; a separate notebook per project.
2. 2Checkpoint — a restore point made at every prompt; press Esc twice or run /rewind to time-travel back before a bad change.
3. 3CLAUDE.md — a plain text file of standing instructions you write, read at the start of every session; the note you tape to a colleague's monitor.
4. 4Auto memory — notes Claude writes about your preferences in MEMORY.md and reloads each session; its private diary about working with you.
5. 5Subagent — a helper Claude spawns with its own separate memory that returns only a summary; an intern sent to the archive so your desk stays clean.
6. 6Permission mode — the session's baseline rule for what needs your approval; cycled with Shift+Tab.
7. 7Plan mode — a permission mode where Claude researches and proposes but edits nothing until you approve; an architect who must show blueprints first.
8. 8Worktree isolation — running Claude in a separate copy of the project so parallel sessions can't collide; giving each cook their own kitchen.
9. 9Sandboxing — an OS-level fence around what commands can touch on disk and network; a playpen with hard walls.
10. 10Non-interactive mode — running one prompt via `claude -p` and exiting, no conversation; a vending machine instead of a waiter.
11. 11Verification loop — giving Claude a check it can run (tests, a build) so it iterates until things actually pass; the taste test that decides dinner is ready, not the chef's confidence.