Token- A piece of a word, the unit of measurement and billing. You count what Claude reads (input) and what it writes (output).
Context- The working memory of a conversation: everything that has been said, read, and produced.
Context window- The maximum size of the context. When it fills up, Claude automatically summarizes the older content to keep going.
Compaction- The summary of the context (automatic or via /compact) that lets you continue a long session.
Prompt- The instruction you give Claude. Four components: context, a precise task, constraints, and the expected format.
Model- The engine that answers. Opus: the most powerful. Sonnet: the balance. Haiku: fast and economical.
Agentic- Able to act using tools (reading, writing, running commands) and not just to answer.
Subagent- An instance of Claude launched by Claude for a subtask, in an isolated context. It returns only its conclusion.
Orchestration- Coordination of several agents: in parallel (fan-out), in a chain (pipeline), or in cross-checking (adversarial verification).
Fan-out- Splitting a large task into pieces handled in parallel by several agents.
Pipeline- Chaining steps together: each item moves forward independently from step to step.
Adversarial verification- One agent produces, another tries to refute: the conclusions that hold up are reliable.
Checkpoint- A restore point created automatically before changes. It is the basis of /rewind.
Diff- The view of the proposed changes: green = added, red = removed. Always review it before accepting.
Skill- A folder that packages a piece of know-how (a SKILL.md file plus resources). Claude activates it on its own when the situation matches its description.
Plugin- A ready-to-use pack for a given role: connectors plus skills.
Artifact- Self-contained content created by Claude on claude.ai (a page, tool, or document), displayed in a dedicated panel and shareable by link.
Persistent storage- An artifact's ability to save data durably, with a separate space for each user.
Magic link- Passwordless sign-in by email: you click the link you receive and the session opens.
Prompt injection- Hidden instructions placed inside external content (an email, a web page, a document) to manipulate the AI. The countermeasure: review any action triggered by external content before approving it.