Jargon, decoded.

34 Claude Code and Claude API terms explained like you would to a smart friend.

Agentic coding: Working with an AI that takes actions (reading files, running tests, making edits) in a loop until a goal is met, rather than answering a single question.
Agentic loop: The cycle an agent repeats: ask the model, run any tools it requests, feed the results back, and stop when the model says it is finished (stop_reason "end_turn").
Claude Agent SDK: Libraries (TypeScript and Python) that give you the same agent loop that powers Claude Code, so you can build your own agents.
Claude Certified Architect (Foundations): Anthropic's certification for people who design production systems with Claude. Scenario-based multiple choice across five domains.
Claude Code: Anthropic's agentic coding tool. It runs in your terminal, IDE, desktop app or browser, reads your project, edits files, runs commands and works through multi-step tasks with you.
CLAUDE.md: A Markdown file Claude Code loads automatically as memory. Put project conventions, commands and rules here. Lives at user, project and sub-directory level.
/clear: Starts a fresh conversation, wiping the current context. Use it between unrelated tasks.
/compact: Summarises the conversation so far to free up context space while keeping the important points.
Context window: Everything Claude can "see" at once: your messages, files it has read, tool output. When it fills up, older detail gets squeezed out, so keep it focused.
Coordinator (orchestrator): In multi-agent systems, the agent that splits up the task, briefs subagents, gathers their results and decides what happens next.
Few-shot prompting: Including a handful of worked examples in your prompt so the model copies the pattern. The most reliable fix for inconsistent output.
Headless mode: Running Claude Code non-interactively with the -p (print) flag, so it can be used in scripts and CI pipelines.
Hook: A script that runs automatically at a lifecycle event, such as before a tool runs (PreToolUse) or after (PostToolUse). Hooks are deterministic: they run every time.
Hub-and-spoke: A multi-agent layout where every subagent talks only to the coordinator, never to each other. Makes the system observable and easier to debug.
Human-in-the-loop: Designing an AI workflow so a person approves risky or irreversible steps before they happen.
/init: A slash command that scans your project and drafts a starter CLAUDE.md for you.
MCP (Model Context Protocol): An open standard for connecting AI tools to outside systems such as GitHub, databases, Slack or Jira through "MCP servers".
.mcp.json: A project-level file listing MCP servers so the whole team shares the same integrations. Supports ${ENV_VAR} expansion to keep secrets out of git.
Message Batches API: Send many requests at once for asynchronous processing at a lower cost. Great for overnight jobs, not for anything a person is waiting on.
Messages API: The core Claude API. You send a list of messages and get the model's reply, plus a stop_reason explaining why it stopped.
Permission modes: Settings that control how much Claude can do without asking: ask before edits, auto-accept edits, plan mode and more. Cycle them with Shift+Tab.
Plan mode: A permission mode where Claude researches and proposes a plan without editing anything. Toggle it with Shift+Tab. Ideal for big or risky changes.
Plugin: A shareable bundle of Claude Code extras such as commands, skills, subagents, hooks and MCP servers, installed with /plugin.
Programmatic enforcement: Guaranteeing a rule in code (a hook or gate) rather than asking nicely in a prompt. Use it whenever one failure would cost money or break compliance.
Prompt: The instruction you give Claude. Good prompts state the goal, the context, the constraints and what "done" looks like.
Prompt caching: Reusing a long, unchanging prompt prefix between requests so it is cheaper and faster to process.
settings.json: Claude Code's configuration file. Lives in ~/.claude/ (you), .claude/settings.json (team, committed) and .claude/settings.local.json (you, this project, not committed).
Skill: A folder with a SKILL.md file that teaches Claude a repeatable workflow. Claude loads it on demand when the task matches its description.
Slash command: A shortcut typed with a leading slash, such as /clear or /compact. You can write your own as Markdown files in .claude/commands/.
stop_reason: A field on each API response. "tool_use" means Claude wants a tool run; "end_turn" means it has finished. Agent loops should rely on this, not on parsing text.
Subagent: A specialised helper Claude can hand work to, with its own context window, prompt and tool permissions. Defined as Markdown files in .claude/agents/.
Token: The unit models read and write in, roughly three quarters of a word. Context limits and API pricing are measured in tokens.
Tool use: Letting Claude call functions you define (look up an order, search docs). You describe each tool; Claude decides when to call it; your code runs it.
tool_choice: An API setting: "auto" (Claude decides), "any" (must call some tool) or a named tool (must call that one). Useful for forcing structured output.