AI terms like token, context window, LLM, prompt and agent describe what an AI tool is made of and how it works. Here are the 40 most important ones for beginners, A to Z, each in a few sentences, with links to in-depth articles.
I work with Claude every day. My fourteen AI employees run on Claude Team, some also on open models or through the API. All AI basics are on my AI Basics page.
How the terms fit together
An example: you type a question into ChatGPT or Claude; that is your prompt. A tokenizer splits it into tokens. The LLM sees them with the system prompt in its context window and produces an answer token by token; that is inference. If it doesn't know the answer, it may guess, and then you get a hallucination. I explain what the umbrella term artificial intelligence covers in a separate article.
The terms from A to Z
Agent (AI agent)
An AI agent is an AI that gets a goal and finds its own way there, using tools such as web search, files or other programs. More in What Is an AI Agent? Explained Simply and AI Agent, Chatbot, or Assistant?.
AI disclosure
Since 2 August 2026, Article 50 of the EU AI Act applies. If your business uses AI, you mainly have to disclose deepfakes and AI texts on matters of public interest that no person has reviewed editorially; the machine-readable marking of AI content is the providers' job. More in AI Disclosure Rules Under the EU AI Act.
AI employee
That is what I call an AI role with a name, an area of responsibility, a memory and rules that works for you permanently. That is my own term, not a vendor's. More in What Is an AI Employee?.
AI Mode (Google)
AI Mode is a separate area of Google Search: you ask questions in full sentences and get a written answer with links to the sources instead of a list of links. A Gemini model puts the answer together.
AI model
An AI model is the computing core of an AI: the parameters learned in training and the blueprint for calculating with them. It is not the same as the app you chat in. More in What Is an AI Model?.
AI usage policy
An AI usage policy is the written house rule for how your team uses AI tools: which ones are allowed, which data may go in and who checks the results. For a small team, one or two pages are enough. A template is in AI Usage Policy for Small Teams.
API
API stands for application programming interface. Through it, two programs talk to each other directly, for example when software sends a request to an AI model without anyone opening a chat window. More in What Is an API?.
CLAUDE.md
A text file with standing instructions that Claude Code reads at the start of every session. Which parts also work in other tools is in CLAUDE.md vs AGENTS.md: What Survives.
Compaction
When the context window fills up, a tool like Claude Code automatically summarizes the conversation so far. Instructions given only in the chat can get lost.
Context and context window
Context is everything an AI has in front of it at one moment: your messages, its answers, files and instructions. The context window is the space for it, measured in tokens. Anthropic calls it the model's working memory. In detail in Context and Context Windows in AI.
Context engineering
Anthropic's term for the strategies that decide which information a model has in its context. It goes beyond the prompt to everything in the context window.
Embedding
An embedding turns text into a long list of numbers. Texts with similar meaning get similar numbers, so a program can search by meaning instead of exact words. RAG systems use it for search.
Fine-tuning
Fine-tuning trains a finished model further on additional data so it picks up its patterns and style. According to Anthropic, the Claude API doesn't currently offer it.
Generative AI
Generative AI creates new content: text, images, music or code. Large language models are one example, image generators another. For the difference to agentic AI, see Agentic AI Explained: What It Is and Isn't. Examples and limits are in Generative AI Explained.
Hallucination
A hallucination is an AI answer that sounds convincing but is wrong or made up, such as a source that doesn't exist. According to Anthropic, it happens even to the most advanced models. More in my article on AI hallucinations.
Hook
An action, usually a command, that Claude Code runs by itself at a fixed point in its workflow, for example before every file edit. More in Claude Code Hooks: Rules That Always Apply. Overview: Claude Code Extensions: Plugins, Mods, Skills.
Inference
Inference means using a trained model: you ask, the model computes the answer. Training happens beforehand at the vendor, inference with every request.
Knowledge cutoff
The knowledge cutoff is the point in time up to which a model knows things best. Anything later it only knows if it searches the web or you provide it.
LLM (large language model)
LLM stands for large language model. It has been trained on huge amounts of text and can write text, answer questions and summarize. Behind ChatGPT, Claude and Gemini is such a model. More in How Do LLMs Work?.
Local AI
Local AI means running a language model on your own hardware instead of through a cloud service. More in Local AI: What Actually Works in 2026.
Machine learning
Machine learning means software learns from examples instead of a person programming every rule by hand. The language models behind ChatGPT and Claude were built this way too. More in Machine Learning Explained Simply.
MCP (Model Context Protocol)
MCP is an open standard through which AI applications such as Claude or ChatGPT connect to data sources and tools. Its makers call it a USB-C port for AI. Such a program is an MCP server; see Set Up a Claude MCP Server.
Multimodal
A multimodal model understands or produces more than one kind of content, for example text and images. Claude, for instance, can read photos, screenshots and PDFs.
Open-weight model
A model whose weights, meaning the learned parameters, you can download and run yourself. Whether you may use it commercially is decided by the license. For an overview, see Open Source AI Models 2026 Compared. I explain the licenses in Open Weight vs Open Source: Licenses.
Parameter
Parameters are the numerical values a model learns during training. Their number gives the size of a model, today often many billions. The more parameters, the more memory the model usually needs.
Plugin
A package that installs several building blocks for Claude Code at once, such as skills, hooks and MCP servers. More in Claude Code Plugins: Install, Build, Trust. Overview: Claude Code Extensions: Plugins, Mods, Skills.
Prompt
A prompt is the text you give an AI to make it do something: a question, an instruction or a whole document. You'll find templates in Claude Prompts: Templates.
Prompt injection
In a prompt injection, hidden instructions sit in a file, web page or email and try to get the AI to do something you never asked for. How Claude Code guards against it: Claude Code Security: Permissions and Sandbox.
RAG (retrieval augmented generation)
With RAG, a system first finds matching passages in your documents and gives them to the model along with your question. That way the AI answers with knowledge from outside its training and can cite sources. A practical example: Build an AI Knowledge Base From Calls. More in RAG Systems Explained.
Reasoning model
A model that thinks in its own words before answering, trying approaches and checking intermediate results. In Claude this is called thinking. It takes longer but helps with calculations, code and tricky tasks.
Skill
A file called SKILL.md with instructions or a workflow that Claude loads when needed. More in Claude Subagents, Skills, Agents. Overview: Claude Code Extensions: Plugins, Mods, Skills.
Slash command
A command you type in Claude Code with a slash, such as /clear for a fresh conversation or /compact for a compaction. The most important ones are in my overview of Claude Code commands.
Subagent
A helper with its own context window that handles a subtask and only returns a summary. More in Claude Code Subagents: Best Practices. Overview: Claude Code Extensions: Plugins, Mods, Skills.
System prompt
The system prompt is the instructions before your conversation that set the AI's role, tone and rules. In chat apps the vendor writes it; through the API or in your own assistants you write it yourself.
Temperature
Temperature controls how randomly a model picks its words. Low means predictable and close to the most likely answer; high means more varied. According to Anthropic, even at temperature 0 the answers are not fully identical.
Token
A token is a piece of text, for example a word or part of a word, that an AI splits every text into. It has nothing to do with crypto tokens. Context windows and answer lengths are measured in tokens.
Training
In pretraining, a model reads huge amounts of text and learns to predict the next word. In post-training it learns to follow instructions and answer helpfully, partly with feedback from people. Only then is it an assistant like Claude or ChatGPT.
Transformer
The transformer is the blueprint on which today's language models are built. Researchers at Google introduced it in 2017. At its core is a mechanism that lets every token attend to the other tokens in the context.
Vibe coding
Vibe coding means letting an AI write software and accepting its code without reading it. That works for prototypes; for software with customer data, only with a review. More in What Is Vibe Coding? A Plain Guide.
Workflow
A workflow is a fixed sequence of work: steps that always happen in the same order when a certain event occurs. Unlike an AI agent, it follows a path that someone built in advance. More in What Is a Workflow?.
What to watch out for
- Same word, different meanings. Every vendor uses "agent" or "assistant" a little differently. Check what the product actually does.
- Terms age fast. Model names and features change every few months. This glossary reflects October 2026.
- Not every word is a technical term. Some are marketing or one person's labels, including mine, like "AI employee".
Frequently asked questions
What is the difference between AI, an LLM and ChatGPT?
AI is the umbrella term. An LLM is a particular kind of AI that works with language. ChatGPT is a product, the app in which you chat with an LLM from OpenAI.
Which terms should I learn first as a beginner?
Start with prompt, token, context window, LLM and hallucination. These five explain what happens when you chat and why an AI sometimes gets things wrong.
How up to date is this glossary?
The explanations reflect October 2026 and rely on the vendors' documentation. When a term changes significantly, I update the entry.
Why are there Claude Code terms here too?
Because many beginners try Claude Code after the chat and run into words like skill, hook or plugin there. These entries are short and link onward.
Where to go from here
If you are just starting with AI, first read what to know as a beginner. Then read how an LLM works: there you see many of these terms in context. We discuss how others use them day to day in my German-speaking community Claude Practitioners.
