AI glossary

An AI model is the trained system behind Claude, ChatGPT, Gemini and Copilot that reads what you give it and writes an answer.

The models behind these tools are large language models, or LLMs. They predict the next piece of text, one small step at a time, which is why the same question can get two different answers.

A prompt is what you write to the AI, with the task and the context it needs to do the job.

A good prompt says what done looks like. For longer work, the prompt often moves into a file the agent reads every time, such as CLAUDE.md.

Tokens are the small pieces of text a model reads and writes, and the unit that usage and limits are counted in.

A token is often part of a word. Everything in the context window costs tokens, your files and instructions as well as the answer.

The context window is everything the model can hold in view at once, including your instructions, files and the conversation so far.

When the window fills up, older parts get shortened or left out. That is why a new conversation often works better for a new task.

A hallucination is when the model states something with confidence that is not true, such as a figure or a source that does not exist.

It happens because the model writes what is likely, not what it has looked up. Ask for sources, and let a script do the counting when the number must be right.

Deterministic means the same input always gives the same output, like a calculator, and non-deterministic means the output can vary, like an AI model.

Let the model write a script for the parts that must be exact, and use the model for the parts that need judgement.

Effort is a setting for how much the model thinks before it answers.

Higher effort helps on hard tasks and uses more of your plan. In a project in Claude Code, the conversation starts at low effort and the threads at high.

RAG, or retrieval-augmented generation, is when the AI first searches a set of documents and then answers from the parts it found.

It is how many company chatbots answer from internal documents. An agent in a folder does something similar by reading the files it needs.

An AI agent is an AI that works towards a goal in steps, gathering context, acting and checking its own work, instead of answering once.

Agentic AI is the wider word for tools that work this way. The difference between a chat and an agent is mostly how much of the loop the AI runs on its own.

The agentic loop is the cycle an agent repeats until the goal is met, where it gathers context, acts and checks the result.

An agent in a folder is an AI agent that works inside one folder with the instructions, examples and files for a job, so it never starts from zero.

The folder is the agent's memory. What you correct today goes into the folder, and the next run starts from it.

A shared agent folder is one agent folder a whole team works from, with the jobs, skills and feedback in one place.

Human in the loop means you approve each step, while human on the loop means you set the goal and the limits and review what comes back.

Start with yourself in the loop, and move outwards as you come to trust the result.

Subagents are helpers an agent starts for one part of a job, each with a fresh context window, that report back when they are done.

Claude Code has a few general helpers built in. A helper with a set role, such as one that checks the costs, is a file you add to the folder.

Graph engineering is designing a team of AI agents with set roles and handovers, instead of one agent doing everything.

The term spread on X in July 2026, after Peter Steinberger asked whether people had moved on from loops to graphs.

Computer use is when an AI agent works a screen or a browser like a person, reading, clicking and typing.

It is slower than a connector, so it is best for tools that have no connector.

A skill is a saved way of doing a task, instructions with examples and sometimes a script, that the agent uses when the task comes up.

Anthropic introduced Agent Skills as an open way to package instructions, and skills work in Claude, Claude Code and Cowork.

Connectors are links that let the AI read from and act in your other tools, such as email, calendar or Slack.

MCP, the Model Context Protocol, is the open standard most connectors are built on, published by Anthropic in 2024.

It means a tool can be built for MCP once and then used by any AI app that supports the standard.

A plugin is a bundle of skills, connectors and other tools you install in one go, often for one role or team.

CLAUDE.md is a text file in the folder that Claude reads first every time, with the rules, context and lessons for that work.

Other agents read a similar file called AGENTS.md, and Claude Code can read it too.

Memory is what the AI carries from one conversation to the next, in its own memory feature or in files you keep.

Files in the folder are the memory you control. You can read them, edit them and share them.

A scheduled task is a task Cowork runs on its own at a set time, like a Monday brief from your inbox and calendar.

Routines are saved tasks in Claude Code that run in the cloud on a schedule, or when another tool such as GitHub sets them off.

Each run works on its own from a saved prompt, without a conversation around it. Routines are in research preview.

Markdown is plain text with a few marks for headings, lists and bold, the format AI agents read and write most easily.

A script is a small program the agent writes and runs when the answer must be exact, such as summing a column.

Python is the language it uses most often.

GitHub is a service that stores folders with their full history, so a team can share the same files and roll back a change.

Vibe coding is building software by describing what you want and judging the result, without writing the code yourself.

Andrej Karpathy, one of the founders of OpenAI, gave it the name in 2025.

Compounding engineering is building software with AI agents so that each piece of work makes the next one easier.

Kieran Klaassen at Every calls it compound engineering, a loop of plan, work, review and compound. In the last step, what the agent learned goes into CLAUDE.md, so the next feature starts from it.

A compounding workflow is an agent task that gets better every time it runs, because corrections, examples and lessons are saved in the agent's folder.

It is the same idea as compounding engineering, used for everyday work such as a monthly report or a reply to a customer. You correct the result once, and the next run starts from that correction.

AI capital is what a company builds up when AI work leaves better instructions, skills and examples behind for the next round.

Satya Nadella, the head of Microsoft, writes about a similar idea he calls token capital, the AI capability a company builds and owns.

An HTML page is a single web page file the AI makes in place of a document or a deck, with room for charts, sliders and live numbers.

Claude is Anthropic's AI, and the app where you work with it on the web, desktop and phone.

Claude Cowork is the part of Claude where you hand over a task and Claude works through it with your files and apps.

It works in the desktop app, on the web and on the phone, and is built for work other than coding. Cowork tasks can also run in the cloud, on Anthropic's servers.

Cloud means the AI works on Anthropic's servers and keeps going when your laptop is closed, while local means it works on your own computer.

Cloud tasks need sources they can reach without your laptop. Local work can use your own files and programs, but only while the computer is on.

Claude Code is Anthropic's AI agent that works in a folder, built for software and also used for analysis, documents and web pages.

It runs in the terminal, in code editors, in the desktop app and in the cloud. The folder holds the instructions, so the work can be repeated.

A project in Claude chat or Cowork is a workspace with its own instructions, files and memory, shared across the chats in it.

On Team and Enterprise you can share a project with colleagues. Projects in Claude Code are a different thing, with threads that do the work.

A project in Claude Code is one conversation where Claude splits the work into threads that run in parallel.

The conversation works like a manager and the threads like workers. Projects are in public beta for Pro and Max.

Threads are the separate pieces of work in a project in Claude Code, each a full session with its own task.

Most run in the cloud. A thread that needs an approval waits for you inside the thread.

Claude Tag is Claude in Slack, for work that belongs to the team rather than one person.

Anyone in the channel can give Claude a task, and the whole channel sees the result. It is in beta for Team and Enterprise.

Opus, Sonnet and Haiku are Anthropic's main Claude models, from the most capable to the fastest and cheapest.

Anthropic also has Fable, a step above Opus, for demanding reasoning and long agent work. A bigger model is better at hard, long tasks and uses more of your plan.

Artifacts are documents, pages or small apps Claude makes beside the chat, which you can open, change and share.

Plan limits are the cap on how much you can use Claude in a set period, and the cap depends on your plan.

Long agent runs, many threads and high effort use more of your plan.

ChatGPT is OpenAI's AI assistant.

ChatGPT Work handles longer tasks with many steps, and Codex writes code.

Codex is OpenAI's coding agent, the closest match to Claude Code.

Gemini is Google's AI, in its own app and inside Gmail, Docs and the rest of Google Workspace.

Microsoft Copilot is Microsoft's AI inside Word, Excel, Outlook and Teams.

A basic version, Copilot Chat, comes with Microsoft 365 for work, while the full Copilot is licensed separately. Copilot Cowork is Microsoft's agent that carries out tasks across Microsoft 365, such as emails, meetings and documents, and asks before each important step.