Glossary

Every term used in this track, defined once.

Adaptive thinking — thinking: {type: "adaptive"}. Claude decides when and how much to reason, calibrated by task complexity and the effort setting. Replaces extended thinking.

Agent — a system where the model decides the control flow, as opposed to a workflow where you do.

Agent SDK — Claude Code as a library (Python, TypeScript). Gives you the agentic loop, tools, hooks, permissions, subagents and sessions with full orchestration control.

Agent team — multiple independent Claude Code sessions that message each other and share a task list. Experimental; disabled by default.

Agent view — the Claude Code dashboard (claude agents) for monitoring and dispatching parallel background sessions.

Agentic loop — model reasons → calls a tool → harness executes → result returns to context → repeat. The mechanism behind every agent.

Artifact — (1) in chat, substantial content rendered in a side panel; (2) in Claude Code, session output published as a live shareable page.

Auto memory — notes Claude writes for itself, stored per repository at ~/.claude/projects/<project>/memory/. Only the first 200 lines / 25KB of MEMORY.md loads at session start.

Auto mode — a Claude Code / Cowork permission mode where a background classifier evaluates each action for safety and blocks risky ones, rather than prompting for each.

Batch API — asynchronous request submission at a discount. The right choice for evals and bulk processing.

Beta header — betas=["..."], required to enable pre-GA API features.

Cache breakpoint — a cache_control marker. Everything up to and including it is cached. Up to four per request.

CLAUDE.md — a markdown file Claude Code loads at session start, containing persistent project instructions. Target under 200 lines.

Channels — a mechanism for an MCP server to push events (webhooks, alerts, chat) into a running Claude Code session.

Checkpointing — tracking file changes so they can be rewound.

Compaction — summarising a conversation to free context space. Project-root CLAUDE.md is re-injected afterwards; nested ones and conversation-only instructions are not.

Connector — the consumer-app name for an MCP server integration.

Constrained decoding — the mechanism behind structured outputs: the model literally cannot produce schema-violating output.

Context editing — programmatically removing content from a conversation, typically stale tool results.

Context engineering — managing what occupies the context window: adding, ordering, compressing, removing.

Context awareness — some models (Sonnet 5, 4.6, 4.5, Haiku 4.5) track their own remaining token budget during a conversation.

Context window — everything Claude can see at once. 1M tokens on current frontier models, 200k on Haiku 4.5.

Cowork — Claude’s agentic interface for non-developers. Multi-step tasks over local files, with cloud-run sessions. Paid plans only.

Dynamic workflow — a script Claude writes that orchestrates many subagents; inspectable and rerunnable.

Effort — output_config.effort. How much internal work Claude does before responding. Defaults to high on the API for Opus 5 and Sonnet 5.

Extended thinking — the legacy thinking: {type: "enabled", budget_tokens: N}. Deprecated; returns a 400 on Claude 4.7+.

Few-shot / multishot prompting — giving 3–5 examples to steer format, tone and structure.

Files API — upload a file once, reference it by ID. Required for getting files in and out of the code execution container.

Fork — branching a session so you can explore alternatives from the same state without altering the original.

Grounding — requiring answers to come only from supplied sources, ideally with quotes extracted first.

Hook — a command, HTTP request, LLM prompt, or subagent that fires at a Claude Code lifecycle event. Deterministic, unlike instructions.

LLM-as-judge — using a model to grade outputs against a rubric. Requires calibration against human scores.

Managed Agents — Anthropic-hosted agents running in a managed cloud sandbox, with sessions, memory stores, scheduling and webhooks handled for you.

Managed settings — organisation-level configuration enforced by the client regardless of user settings.

Marketplace — a hosted collection of plugins.

MCP (Model Context Protocol) — an open standard for connecting AI tools to external systems. Servers expose tools; clients call them.

MCP tunnel — enterprise connectivity for MCP servers inside your network, reachable by Anthropic-hosted surfaces.

Memory tool — a server tool giving Claude persistent memory across sessions.

Multishot — see few-shot.

Namespacing — plugin skills are invoked as /plugin:skill so multiple plugins coexist.

Orchestrator–worker — the dominant multi-agent pattern: an orchestrator plans and synthesises, workers do isolated subtasks.

Output style — a Claude Code mechanism for adapting the assistant’s persona and defaults for non-coding uses.

Path-scoped rule — a .claude/rules/*.md file with paths frontmatter, loading only when Claude touches matching files.

Permission mode — default/manual, acceptEdits, plan, auto, dontAsk, bypassPermissions. Cycle with Shift+Tab.

Plan mode — Claude proposes without editing. The highest-value habit in Claude Code.

Plugin — an installable bundle of skills, subagents, hooks, MCP servers, commands and themes.

Prefill — supplying a partial assistant message to continue from. No longer supported on Claude 4.6+; returns a 400.

Programmatic tool calling — Claude writes code that calls your tools, rather than individual tool_use round trips.

Progressive disclosure — the Skills loading model: metadata always (~100 tokens), instructions when triggered, resources only when read.

Prompt caching — caching a stable prefix server-side so repeat requests read it at a fraction of the cost.

Prompt injection — an attack where content Claude reads contains instructions aimed at redirecting it.

RAG (retrieval-augmented generation) — retrieving relevant chunks and putting them in context, rather than loading everything. How Project knowledge bases work above a size threshold.

Routine — a Claude Code automation running on Anthropic-managed infrastructure, triggerable by schedule, API call, or GitHub event.

Rules — .claude/rules/*.md. Modular instructions, optionally path-scoped.

Sandbox — an isolated execution environment. Claude Code’s sandboxed Bash tool provides filesystem and network isolation.

Server tool — a tool Anthropic executes for you (web search, code execution, computer use, memory, advisor), as opposed to a client tool you implement.

Service tier — an API setting trading latency against throughput and priority.

Session — a persisted conversation. Continue, resume, or fork it.

Skill — a SKILL.md file with frontmatter plus optional bundled files and scripts, loaded on demand.

Stop reason — why generation ended: end_turn, max_tokens, tool_use, stop_sequence, refusal.

Strict tool use — strict: true. Guarantees tool names and inputs validate against their schemas.

Structured outputs — output_config.format. Guarantees the response matches a JSON Schema, via constrained decoding.

Subagent — an isolated worker with its own context window. Only its summary returns to the caller.

Surface — a place Claude Code runs: terminal, VS Code, JetBrains, desktop, web, mobile, Slack, CI.

System prompt — a top-level system parameter (not a message) setting role and standing instructions.

Token — the unit Claude reads and is billed in. Roughly 3–4 characters in English; ~750 words ≈ 1,000 tokens.

Tool — a function description Claude can request to call. Claude never executes; the harness does.

Tool search — loading only the tools needed on demand, so large tool sets don’t fill context. On by default in Claude Code.

Trajectory evaluation — grading an agent’s path, not just its final answer.

Ultraplan / ultrareview — cloud-based deep planning and deep multi-agent code review in Claude Code.

WIF (Workload Identity Federation) — API access without long-lived keys, using your existing identity provider.

Workspace — a Console-level container with its own API keys, budget and rate limits.

Worktree — an isolated git working directory. claude -w <name> runs a session in one.

ZDR (Zero Data Retention) — available to qualified Enterprise accounts, with documented feature trade-offs. Agent Skills are not covered.

Anthropic’s own glossaries: platform · Claude Code