Skip to main content

Glossary

All levels

Plain-English definitions, no prior knowledge assumed. Skim it, or search for a term you hit elsewhere.

Review the essentials as flashcards โ€” tap to flip:

Core terms
Press Enter or Space to flip the card. Use the left and right arrow keys to move between cards.Term shown.
1 / 9

Agentโ€‹

An AI set up to take actions in a loop โ€” it reads a goal, calls tools, sees the results, and repeats until done. Claude Code is an agent for your codebase.

Anthropicโ€‹

The company that makes Claude.

API (Application Programming Interface)โ€‹

A way for your software to talk to Claude programmatically, instead of typing in a chat window. See Your First API Call.

Artifactโ€‹

A live, runnable output (an app, chart, or document) that Claude renders in a side panel on claude.ai, separate from the chat.

Chain-of-thought (CoT)โ€‹

Asking the model to reason step by step before answering, which improves accuracy on hard problems.

Claudeโ€‹

The family of AI models (and apps) made by Anthropic. "Claude" can mean the chat app, the models, or the brand depending on context.

Claude Codeโ€‹

Anthropic's agentic command-line/IDE tool that can read and edit your codebase, run commands, and use tools. See What Claude Code Is.

CLAUDE.mdโ€‹

A plain-text file where you give Claude Code persistent instructions about your project โ€” its highest-leverage customisation.

Compactionโ€‹

Automatically condensing older conversation turns into a short summary when the context window fills, so a long session can keep going without losing the thread. See Context Management.

Context engineeringโ€‹

Deliberately deciding what goes into the model's context window โ€” instructions, examples, retrieved data โ€” to get the best result. The discipline behind prompting, RAG, and memory.

Context windowโ€‹

The maximum amount of text (measured in tokens) a model can consider at once โ€” your prompt plus its reply. Exceed it and the oldest content drops off.

Effortโ€‹

A control on newer Claude models for roughly how much the model thinks before answering. Match the tier to the task instead of maxing it out โ€” you pay for thinking in latency and cost. See Extended Thinking & Effort.

Embeddingโ€‹

A list of numbers representing the meaning of a piece of text, so software can find things that are similar in meaning (not just matching keywords). Powers semantic search and RAG.

Eval (evaluation)โ€‹

A repeatable test of AI output quality, so you can tell whether a change made things better or worse.

Few-shotโ€‹

Including a few worked examples in your prompt to show the model the pattern you want.

Fine-tuningโ€‹

Further-training a model on your own examples to specialise it. Usually a last resort โ€” try prompting and RAG first.

Groundingโ€‹

Instructing the model to answer only from the context you provide โ€” and to say "I don't know" when the answer isn't there. The main prompt-level defense against hallucination. See the Pattern Library.

Guardrailsโ€‹

Rules and limits that keep an AI's behaviour safe and on-task โ€” e.g. permission rules, "run tests before done", or human approval for risky actions. See Securing Agents.

Hallucinationโ€‹

When a model states something false but confident. The core skill is verification. See Hallucinations.

Hookโ€‹

In Claude Code, a command that runs automatically at a lifecycle event (e.g. after every file edit) to enforce formatting, gates, or safety.

Inferenceโ€‹

The act of running a model to get an output (as opposed to training it).

Knowledge cutoffโ€‹

The date after which a model has no training data โ€” it genuinely doesn't know what happened later unless you supply the facts in your prompt. Claude's cutoff is published in each model's release notes. This is why asking about very recent events without context can produce outdated or invented answers; the fix is to paste in the relevant facts or use a tool that retrieves live information.

Latencyโ€‹

How long you wait for a response. Bigger models are usually slower; streaming improves perceived latency by showing output as it's generated.

LLM (Large Language Model)โ€‹

The kind of AI behind Claude: trained to predict the next chunk of text, which turns out to be enough to write, code, summarise, and reason. See What Is an LLM?.

MCP (Model Context Protocol)โ€‹

An open standard for connecting AI to external tools and data (databases, GitHub, browsers). An "MCP server" exposes a tool; Claude can call it. See MCP.

Multimodalโ€‹

A model that handles more than text โ€” e.g. images and PDFs as input.

Output styleโ€‹

In Claude Code, a preset that changes how Claude communicates (e.g. concise, explanatory, learning) without changing what it can do. See Output Styles.

Permission modeโ€‹

In Claude Code, how much it can do without asking โ€” from prompting on every action to auto-accepting edits. The main speed-vs-safety dial. See Permissions.

Plan modeโ€‹

A read-only Claude Code mode that researches and proposes a plan before touching any files โ€” you approve before it acts. See Plan Mode.

Promptโ€‹

What you send the model: your instructions, question, and any context.

Prompt cachingโ€‹

Reusing the unchanged front part of a prompt across calls to cut cost and latency. See Prompt Caching.

Prompt injectionโ€‹

An attack where malicious instructions hidden in content the AI reads (a web page, a document) try to hijack it. See Prompt Injection.

RAG (Retrieval-Augmented Generation)โ€‹

Fetching relevant snippets from your data and adding them to the prompt so the model answers about them. See RAG.

Reasoning / extended thinkingโ€‹

A mode where the model spends extra compute "thinking" before answering, for harder tasks. On newer models you tune it with effort.

Skillโ€‹

A packaged capability (instructions + optional scripts) that Claude loads on demand when relevant. See Skills.

Slash commandโ€‹

A /-prefixed shortcut in Claude Code โ€” built-in (like /clear) or your own reusable prompt โ€” that triggers a defined action. See Slash Commands.

Streamingโ€‹

Sending the model's output token by token as it's generated, instead of waiting for the whole response. To a user it looks like the model is typing in real time; to a developer it means processing a stream of chunks instead of one response object. It dramatically improves perceived latency on long outputs and lets you start rendering before the model finishes. See Streaming.

Structured outputโ€‹

Getting results in a strict, machine-readable shape (e.g. JSON matching a schema) instead of free text โ€” so other software can rely on them. See Structured Output.

Subagentโ€‹

A separate Claude instance with its own context and tools, used to delegate part of a task (e.g. research) so the main session stays focused.

System promptโ€‹

The high-level instructions that set the model's role and rules for a whole conversation โ€” your most powerful steering lever.

Temperatureโ€‹

A setting (on some models) controlling randomness: low = focused and repeatable, high = varied and creative.

Tokenโ€‹

The unit models read and count in โ€” roughly ยพ of a word in English. Pricing and context limits are measured in tokens. See Tokens & Context.

Tool use / function callingโ€‹

Letting the model call functions you define (search, calculator, your API) and use the results. See Tool Use.

Zero-shotโ€‹

Asking the model to do a task with no examples โ€” just instructions. Most everyday prompts are zero-shot. Contrast with few-shot, where you include worked examples to guide format or style. Zero-shot is fine for straightforward tasks; add examples when the model keeps missing the output shape you want.


Missing a term? Add it โ€” glossary entries are a perfect first contribution.