AI & Claude Glossary

"Token," "context window," "hallucination" — when reading about Claude or AI you keep running into unfamiliar words.

This page is a dictionary for looking up those terms. They are organized by category, so find the term you want to understand. If another article says "see the glossary for this term," that link will bring you here.

How to use this page: Select a term from the table of contents on the right, or use your browser's in-page search (Windows: Ctrl+F / Mac: Cmd+F) to find a term.

Basic Terminology

AI (Artificial Intelligence)

In one line: The umbrella term for technology that gives computers the ability to think and make decisions like humans.

This covers the entire field of research and technology aimed at giving computers the ability to learn and solve problems. "AI" is used very broadly — from programs that play chess to conversational assistants, to systems that recognize images, all of these fall under AI.

Conversational AIs like Claude and ChatGPT are a type called "Generative AI," which excel at producing new content that resembles text written by humans.

Related terms: Machine Learning / Large Language Model (LLM)

Machine Learning

In one line: A technique in which computers automatically learn "patterns" by being shown large amounts of data.

Instead of humans painstakingly programming rules into a computer, the computer is shown massive datasets and learns the patterns on its own.

For example, to build a spam filter you expose the system to large quantities of spam and legitimate email, and it learns to judge whether a new message is spam automatically. AIs like Claude similarly "learn" from enormous amounts of text on the internet, which is how they generate natural-sounding language.

Related terms: Deep Learning

Deep Learning

In one line: A machine learning method that uses a multi-layered structure called a "neural network," modeled on the neural circuits of the human brain.

One of the most powerful methods within machine learning. By stacking tens to hundreds of "layers," it achieves high accuracy across image recognition, speech recognition, natural language processing, and many other domains.

Modern large language models, including Claude, are built on this deep learning technology.

Related terms: Machine Learning / Large Language Model (LLM)

Large Language Model (LLM)

In one line: A large-scale AI model trained on vast amounts of text data that excels at generating and understanding language.

An AI model trained on enormous volumes of text — books, articles, web pages — from across the internet. By repeatedly predicting "which word comes next," it learns to generate natural language that resembles human writing.

Claude, ChatGPT (OpenAI), and Gemini (Google) are leading AI assistants powered by LLMs. The word "Large" refers to the fact that these models have billions to hundreds of billions of parameters (internal settings).

Common stumbling block: LLMs are often assumed to be all-knowing, infallible entities, but they have real limitations: they do not know recent information that was not in their training data, and they can confidently give wrong answers. See the Hallucination entry for more details.

Related terms: Token / Context Window / Hallucination

Natural Language Processing (NLP)

In one line: The field of technology that enables computers to understand, process, and generate human language (natural language).

"Natural language" here means the everyday languages people use — English, Japanese, Spanish, and so on — as opposed to artificial languages like programming languages. NLP is the umbrella term for the technology that lets computers understand human language, translate it, summarize it, and answer questions about it.

AI assistants like Claude represent the cutting edge of NLP technology.

Claude-Specific Terminology

Claude

In one line: The name of the AI assistant developed by Anthropic.

An AI assistant developed and provided by Anthropic. It handles a wide range of tasks including conversation, writing, analysis, and coding. It is available in many forms: a web browser, a desktop app, a smartphone app, and an API.

The name Claude is used not only for claude.ai but also as a brand name covering all of Anthropic's AI products, including Claude Desktop and Claude Code.

Related terms: Anthropic / Haiku / Sonnet / Opus

Anthropic

In one line: The AI company that develops and provides Claude.

An American AI company founded in 2021. Its central mission is "the responsible development of AI," and it is one of the major AI companies alongside OpenAI (maker of ChatGPT) and Google (maker of Gemini).

Anthropic's research outputs — Constitutional AI and RLHF — are key technologies that underpin Claude's safety and honesty.

Artifacts

In one line: A feature that displays documents, code, HTML, and other outputs Claude creates in a panel where you can preview, edit, and download them on the spot.

One of the features of Claude Web (claude.ai). When you ask Claude to "write a report," "give me a spreadsheet formula," or "build a simple webpage," the resulting output (artifact) appears in a panel on the right side of the chat screen.

You can look at the preview and ask for changes such as "make it shorter" or "add a heading," and you can also copy or download the content.

How it differs from a normal chat reply: A regular chat reply just shows text in the chat. Artifacts display the output in a separate panel and let you interact with it.

Projects

In one line: A feature that lets you store frequently used materials, rules, and background information so Claude automatically refers to them in every conversation.

Available in Claude Web and Claude Desktop. For example, if you register files and information such as "your company's product catalog," "your company's writing guidelines," or "a project specification" in a project, Claude automatically references that information every time you have a conversation within the project.

This eliminates the need to paste the same documents into every conversation, and lets you work continuously with shared context.

How it differs from a normal chat: A regular chat conversation resets its memory after each conversation. Projects let Claude reference the same information (project knowledge) across multiple conversations.

Deep Research

In one line: A feature in which Claude searches the web while conducting complex research and produces a detailed report.

Available in Claude Web. It lets you hand off complex research tasks to Claude — such as "research trends in the domestic SaaS market" or "compare and summarize our competitors' products." Claude visits multiple websites to gather information and then organizes everything into a report with citations.

A regular chat reply comes back in seconds; Deep Research can take several minutes, but the result covers the topic more comprehensively.

Memory

In one line: A feature that lets Claude retain memories across conversations, remembering your preferences and information in future sessions.

Normally, Claude's memory resets at the end of each conversation (as if it were meeting you for the first time every session). When you enable the Memory feature, Claude remembers important information (your name, occupation, preferred response style, etc.) and can use that information in future conversations.

You can enable, view, and edit Memory from Settings in Claude Web and Claude Desktop.

Note: You can view and delete what is stored in Memory from the settings screen. If you have privacy concerns, check Settings > Memory.

Cowork

In one line: A feature that lets Claude continue a long task in the background while you work on something else.

Available in Claude Desktop. You can give Claude a time-consuming task — such as "summarize this 50-page PDF" or "analyze these 100 records" — and then move on to other work. Claude continues working in the background and notifies you when it finishes.

A regular chat is an interactive format where you get a reply on the spot. Cowork is an asynchronous format where you assign the work and wait for it to be done.

Claude Code

In one line: An AI coding assistant for engineers that runs in the terminal (command line) or development environment.

A tool that lets software developers create, edit, debug, and refactor code using natural-language instructions. You use it via the claude command in a terminal or as an extension for IDEs (integrated development environments) like VS Code.

Beyond just writing code, Claude autonomously handles reading and writing files, running Git commands, executing tests, and other development tasks.

How it differs from Claude Web/Desktop/Mobile: Claude Web, Desktop, and Mobile are general-purpose AI assistants for everyday users. Claude Code is a tool specialized for software development work. If you do not have a programming background, Claude Web or Desktop is recommended.

MCP (Model Context Protocol)

In one line: A shared "standard (protocol)" for connecting Claude to external services and tools.

A standard connection specification for integrating Claude with external services such as Google Drive, Slack, and GitHub. By adding an MCP-compatible service or extension (MCP server) to Claude Desktop, Claude can directly use data and tools from those external sources.

For example, if you add the "Google Drive MCP server," you can give instructions like "read last month's sales data from My Drive and analyze it."

Analogy: Think of MCP like a USB port — a common connector that plugs Claude into external services.

Model-Related Terminology

Haiku

In one line: The lightest and fastest model in the Claude model family.

Named after the Japanese poetry form, Haiku is the simplest and fastest of the three model series (Haiku / Sonnet / Opus). It is suited for answering straightforward questions, quickly processing large volumes of text, and cost-sensitive use cases.

Everyday users of Claude Web will rarely encounter it directly, but developers often choose it via the API when processing a high volume of requests.

Comparison: Haiku < Sonnet < Opus in terms of capability, and cost increases in the same order.

Sonnet

In one line: Claude's standard model — the best balance of capability and cost.

Named after the 14-line poetic form, Sonnet handles a broad range of tasks, from everyday business work to advanced analysis and coding. In Claude Web, Sonnet is typically selected as the default, making it the model most users interact with most often.

As of 2025, the latest version is provided with a version number such as claude-sonnet-4-5.

Opus

In one line: The highest-performance model in the Claude family, designed for complex tasks.

Named after the Latin word for "work" or "masterpiece," Opus is the top-tier model. It excels at complex reasoning, advanced analysis, and tasks requiring sustained focus. On the other hand, it is slower than Sonnet and has a higher API cost.

In Claude Web, it can be selected on Pro plans and above. Sonnet is often sufficient for everyday tasks; Opus is recommended when you feel Sonnet is not powerful enough for a demanding task.

Token

In one line: The smallest unit AI uses to process text — slightly larger than a single character.

AI processes text not in "characters" but in "tokens." In English, roughly one word corresponds to one or two tokens; in Japanese, roughly one character corresponds to one or two tokens.

Why it matters: Claude's pricing and the maximum length of text it can process (the context window) are both calculated in tokens. For example, a context window of "200,000 tokens" can handle roughly 150,000 Japanese characters, equivalent to about two or three paperback novels, in a single session.

Common stumbling block: Japanese (and many non-English languages) tend to use more tokens than the same amount of English text, so Japanese content costs more to process.

Context Window

In one line: The upper limit of information Claude can hold in memory and reference within a single conversation.

Think of it as the "size of Claude's working memory" during a conversation. It is expressed in tokens: a 200,000-token context window can process roughly two to three paperback novels' worth of text in one session.

Concrete example: If you ask Claude to read a long PDF and answer questions, Claude can refer to the entire document as long as it fits within the context window. Content that exceeds the limit cannot be referenced.

What happens in a long conversation: The context window fills up as a conversation grows longer. If you notice Claude's replies becoming less relevant, starting a new conversation is a good idea.

Temperature

In one line: A parameter that controls how random (or creative) AI responses are — higher values produce more varied output.

It is set as a number between 0 and 1 (or 0 and 2).

• Low values (e.g., 0.1–0.3): Nearly identical, deterministic replies every time. Good for tasks requiring accuracy, such as fact-checking or translation.
• High values (e.g., 0.7–1.0): Responses vary more. Good for brainstorming or creative writing, where you want a range of ideas.

Everyday Claude Web users do not set this directly, but developers using the API can adjust it per request.

Cooking analogy: Imagine it as a seasoning dial. If you want precision, go light (low value); if you want to spark ideas, go bold (high value).

Prompt-Related Terminology

Prompt

In one line: The instruction, question, or request you give to AI.

The entire message you send to Claude is called a prompt. Phrases like "Rewrite this email in a more polite tone" or "Suggest five agenda items for next week's meeting" are prompts.

Keys to a good prompt:

• State clearly what you want done (the task)
• Provide specific conditions or constraints ("within 100 words," "in bullet points," etc.)
• Share background information and purpose ("I want to use this in a presentation for X")

Related terms: System Prompt / Chain of Thought / Few-shot

System Prompt

In one line: Background instructions that set Claude's behavior and role, given before the conversation begins.

Separate from the user's regular messages (the user prompt), a system prompt is a special set of instructions that define Claude's "character" or "rules." For example, you might write: "You are a customer support agent. Always reply in a polite, professional tone and keep responses under 200 words."

Most Claude Web users never see this directly, but it is commonly used when companies embed Claude in their own products or when using the Projects feature in Claude Desktop. It is an important feature for developers who want detailed customization via the API.

Chain of Thought

In one line: A technique in which AI shows its reasoning step by step before arriving at a final answer.

When solving a complex problem, this approach has the model work through it step by step — "first I'll consider X, then I'll examine Y..." — before giving an answer. Simply adding "please think step by step" to your prompt gets Claude to show its reasoning process as it responds.

It is especially effective for math word problems, complex decisions, and situations that require logical reasoning.

Example: Rather than asking "Plan next month's budget allocation," try "Plan next month's budget allocation in this order: first review current income and expenses, then prioritize needs, then propose an allocation." Claude will think through it more carefully.

Few-shot

In one line: A technique where you show Claude two or three examples of the kind of reply you want before making your main request.

By including examples in your prompt — "given this kind of input, respond like this" — you help Claude understand the format or style you expect. As the name suggests, just a few examples (typically two to five) are enough to guide Claude's responses.

Example:

Create a one-line tagline for each product name using the pattern below.

Example) Smartwatch → "Time, made smarter."
Example) Electric bicycle → "Hills? No problem."

Now create a tagline for:
Noise-canceling earphones

Hallucination

In one line: The phenomenon where AI confidently states something that is not true.

One of the most well-known shortcomings of large language models. It shows up in forms such as Claude citing a paper that does not exist, giving the wrong founding year of a company, or describing a feature that does not actually exist. Because the language is fluent and the tone sounds confident, the answer can appear correct at first glance, which makes hallucinations especially tricky.

How to handle it:

• Always fact-check important information (numbers, proper nouns, quotations) against another source
• Asking "Is that really correct? Are you sure?" can sometimes prompt Claude to correct itself
• Request "Please provide your sources" as a follow-up

Common stumbling block: Treating Claude as an all-knowing encyclopedia is risky. Use it as an assistance tool and always have a human verify anything important.

API-Related Terminology

API

In one line: A specification that defines how software programs exchange data and functionality — a kind of "connector."

"API" stands for Application Programming Interface. It is the set of rules that lets different pieces of software communicate and work together.

An easy analogy: Think of dining at a restaurant. The customer (your application) gives an order (request) to the waiter (API), who carries it to the kitchen (Claude) and brings back the dish (response).

In Claude's case, developers use the "Anthropic Claude API" to embed Claude's capabilities into their own apps and services.

Relevance to everyday users: If you just use products like Claude Web or Claude Desktop, you never need to interact with the API directly. The API is primarily for developers.

Related terms: Endpoint / API Key

Endpoint

In one line: The specific URL (address) used to access an API.

An API has multiple functions, each with its own corresponding URL (endpoint). For example, the Claude API has an endpoint for sending messages (https://api.anthropic.com/v1/messages). Developers send data such as a question to this URL and receive Claude's response.

API Key

In one line: An "authentication password" for using an API — a string that identifies who is using the API and how much.

To use Anthropic's API, you need an "API key" — a string (e.g., sk-ant-api03-...) generated in the Anthropic console. By attaching this key to every request, Anthropic can track who is using the API and how much, and calculate the usage fees.

Security note: Treat your API key like a password. Embedding it directly in source code published on GitHub or similar platforms creates a risk of unauthorized use. Use environment variables or a secrets management tool to keep it safe.

Rate Limiting

In one line: A cap on the number of requests that can be sent within a given period, designed to prevent excessive use.

To protect servers from overload, limits are set such as "a maximum of N requests per minute" or "a maximum of X million tokens per day." If you exceed the limit, an error is returned and you must wait before sending more.

The "message limit" that Claude Web users occasionally encounter in everyday use is also a form of rate limiting, in the broad sense.

Streaming

In one line: A mode in which Claude displays text on screen in real time as it generates it, word by word.

With a standard API call, Claude finishes generating the entire response before delivering it all at once. With streaming, it delivers the text as it is generated, in real time. You have seen this in Claude Web — the response appears word by word as if being typed. That is streaming.

Because users do not have to wait for the full response to be complete, the perceived response speed feels faster.

Function Calling (Tool Use)

In one line: A feature that lets Claude call external functions or tools to retrieve real-time information or perform actual operations.

By default Claude responds from its trained knowledge, but with Function Calling it can call external capabilities — such as "check today's weather," "search a database," or "add an event to a calendar."

Developers define which functions (tools) Claude can use, and Claude calls those tools as needed when forming its responses. MCP (Model Context Protocol) is also built on this same mechanism.

Security and Safety Terminology

Constitutional AI

In one line: An Anthropic AI training method in which a set of "principles (a constitution)" is given to the AI in advance, and the AI uses those principles to evaluate and improve its own responses.

A proprietary technique developed by Anthropic. In standard AI training, humans evaluate enormous numbers of responses. In Constitutional AI, the AI is given a list of principles (a "constitution") — such as "respect human dignity" and "do not share dangerous information" — and evaluates and revises its own responses against those principles.

This training is what makes Claude inclined to avoid harmful content and give honest, helpful answers.

RLHF

In one line: A training method in which humans rate AI responses as "good" or "bad," and those ratings are used to make the AI better over time.

Humans rank multiple AI-generated responses, and the AI is trained to produce more of the "good" responses. Most major modern AI assistants, including ChatGPT and Claude, have been trained using this technique.

It is used to align the AI's responses with human values and preferences.

Safety

In one line: The collective term for efforts to design, train, and monitor AI so it does not cause harm.

The term covers the research, technology, and policies that ensure AI does not lie, does not share dangerous information, does not make discriminatory statements, and cannot be used for malicious purposes.

Anthropic lists "AI Safety" as its top priority mission, and safety is factored into every design decision made for Claude.

Alignment

In one line: The field of research that tries to align AI goals and behavior with human values and intentions.

This field focuses on designing AI so it acts in line with human intent and values — being helpful, harmless, and honest — and does not take unexpected actions or pursue goals that conflict with human intentions.

Constitutional AI and RLHF are both concrete technical methods for achieving this alignment.

Next Steps

Now that you understand the terminology, try deepening your knowledge by using Claude hands-on.

If you want to start using Claude:

• What is Claude? — The Big Picture — How to choose a product and get started
• Your First Claude Web Chat — Try it in your browser right now

If you want to learn about prompting (how to give instructions):

• Prompt Fundamentals — Learn how to write effective instructions

If you want to know how to use Claude safely:

• Safe Use & AI Limitations — Risks like hallucination and how to handle them

If you want to use the API or explore technical integrations:

• API Guide — Embedding Claude in Your App — API explained for developers
• MCP (Model Context Protocol) Guide — How to connect Claude to external services