Most AI terms describe one of a few things: what the AI is, how much it can hold in view at once, what it can reach beyond the chat, what it can do on its own, and how far to trust what it says. A token is a small piece of text, and the context window is how many tokens the AI can work with at once. A hallucination is a made-up answer said as if it were true. Knowing these words tells you why an assistant behaves the way it does, and what to do about it.

The basics

AI

Artificial intelligence: computers doing things that usually take human thinking. It is a family name. Chat assistants are one kind; face unlock and spam filters are others.

LLM and model

A large language model. A model is what an AI learned, packed into one program, and an LLM learned from a huge amount of writing. It writes a reply one small piece at a time, predicting what comes next. Every chat assistant you use at work runs on one.

Prompt

What you type to tell the AI what you want: a question, a request or a set of directions. A follow-up is a prompt too. In practice: the model only knows what you ask for, so say what you want, who it is for and how long it should be.

Output

Whatever the AI gives back: text, a table, a file. Your prompt shapes it. Name the format you need, then check it came back that way.

What it can hold at once

Token

A small piece of text: a short word, or part of a longer one. AI reads and writes in tokens, and more text means more tokens. A big file uses far more than a short message.

Context window

How much the AI can hold in view at once, counted in tokens. Think of a desk: everything it is working with has to fit on it. In practice: long chats and big files fill it, and what falls off the desk the AI can no longer see.

Compaction

Summing up the older part of a long chat so there is room to keep going. A summary is shorter than the real thing, so small details can drop out. If an exact figure from early on matters, repeat it.

Context

Everything in front of the AI right now: your messages, the chat so far, files and instructions. If it is not in the context, the AI cannot see it. A file you never shared is not there, however obvious it seems to you.

Instructions that are there before you type

System prompt

The app makers’ instructions, there before you type anything. They set how the AI behaves in that app. You do not write it, and you usually cannot see it.

Global instructions

Instructions you write once in settings, which the app adds to every chat. Some apps call them custom instructions or preferences. In practice: if you keep typing the same line (“use American spelling”, “keep it short”), put it here instead. The AI reads the system prompt first, then your global instructions, then the chat, then your newest message.

Memory

What an app keeps about you for next time, such as your job or how you like answers. It carries lasting facts from one chat to the next. Where an app has memory, look in its settings to see and delete what it saved.

Reaching beyond the chat

Tool

Something the AI can use to do a job, like searching, reading a calendar or drafting an email. The AI decides; the tool does the job.

API

Application programming interface: the set way one program asks another for something, like a weather app asking a weather service for the forecast.

MCP

Model Context Protocol, an open standard for connecting AI assistants to apps and information. Think of it as a standard plug, so an assistant connects to many tools the same way. In Claude and ChatGPT, the connectors and apps you add are often built on it.

RAG

Retrieval-augmented generation: the AI looks up the right pieces of a set of documents, adds them to its context, then answers with them. It is how an assistant answers from your own files rather than from memory.

Web search

The AI looking things up on the internet. It is how an assistant gets news and facts newer than what it learned. In practice: for anything that changes (prices, schedules, today’s news), ask it to search, and check the sources it cites.

Doing work for you

Skill

A saved set of instructions that teaches the AI how to do one job, so you do not explain it again. Think of a recipe card the AI keeps. See what a Claude skill is for how to write one.

Workflow

Steps in order: first this, then that, then the next thing. A workflow runs the same steps every time.

Trigger

What starts it: a time on the clock, something happening, or you asking. When the trigger goes off, the skill or workflow runs.

Agent

An AI that works toward a goal, choosing its next step and using tools until the job is done. Where a workflow follows fixed steps, an agent decides what to do next. In practice: the more an agent can do on its own, the more its permissions matter.

Subagent

A helper. The main agent hands it one part of the job, and it reports back.

Multi-agent

Several agents, each with a role, working on one job, like a team where one gathers facts, one writes and one checks.

How models are made and run

Training

How a model is made: it learns patterns from huge amounts of text, long before you use it. Nearly everything it knows comes from here, which is why it can be out of date.

Fine-tuning

Giving a trained model extra practice on one kind of job, using examples. It is a project, not a setting. For most work, a clear prompt or a skill gets you there more cheaply.

Inference

The model at work: reading a prompt and writing a reply. Every reply you get is one round of inference.

Latency

The wait between asking and getting the answer. Longer inputs and bigger jobs usually mean a longer wait.

Cost

Every run costs something, usually by the number of tokens going in and coming out. Long files in and long answers out run the meter up. On a chat subscription you rarely see it as money; it tends to show up as usage limits instead.

Access and safety

API key

A secret code, like a password for programs, that proves a program is allowed to use a service. Keep it private, never paste it into a chat or a shared document, and replace it if it leaks.

Permission

What the AI is allowed to do: read something, change it, or ask you first. In practice: when an assistant asks before it sends or changes something, that is a permission at work. Read what it is asking for before you allow it.

Guardrail

A built-in safety rule that blocks harmful or off-limits requests and answers. It is meant to hold even when someone tries to talk the AI past it.

How far to trust it

Hallucination

A made-up answer, said as if it were true: a citation to page 14 of a six-page file, a quote nobody said. It happens because the model predicts likely words, and likely-sounding is not always true. Hallucinations sound just as sure as true answers, which is why you check.

Eval

A test you build for your own task: real examples with known good answers. You run them, compare the AI’s answers with yours, and count. It tells you how well AI does on your work.

Benchmark

A shared, standard test, so different models can be scored on the same questions. A benchmark score says how a model did on that test, not on your work.

A quick way to use this list: when an assistant surprises you, ask which word explains it. It forgot something (context window, compaction). It does not know about last week (training; ask for a web search). It invented a source (hallucination). It cannot see your file (context). It asked before sending (permission).