The AI glossary: plain-English definitions for the words everyone uses.

A quick note on how to use this: you don't have to read it top to bottom. Bookmark it and come back whenever someone says a word that makes you feel like you missed a meeting. It'll be here.

AI models.

The "brain" doing the actual thinking. When people say Claude, ChatGPT, or Copilot, they're referring to the AI model, which is the software that reads your message and generates a response. Different companies build different models, and within each brand there are usually multiple tiers. Claude (made by Anthropic) has Haiku (fast and lightweight), Sonnet (the everyday workhorse) and Opus (the most powerful, also the slowest). ChatGPT is made by OpenAI. Copilot is Microsoft's version, which runs on OpenAI technology under the hood.

In practice: When you're chatting at claude.ai or chatgpt.com, you're talking to an AI model. Most people use the middle tier for most things. The more powerful versions are worth it for complex tasks: long documents, nuanced writing, technical problems.

Prompt engineering.

Sounds technical. It just means: how you ask. A better prompt gets a better answer. Prompt engineering is the skill of giving AI enough context, a clear format, and specific instructions so it gives you what you actually need instead of a generic response you have to rewrite anyway.

In practice: "Write an email" is a prompt. "Write a warm but professional email to my client Sarah explaining we need one more week on the project, keep it under 100 words, don't start with I hope this email finds you well" is prompt engineering. The second one saves you a rewrite.

Local memory.

When you're inside a single conversation, AI remembers everything said in that session. That's local memory. Close the tab, start a new chat, and it's gone. The model didn't forget, it never had access to it in the first place. Each new conversation starts completely fresh.

In practice: You spend the first message of a chat telling Claude your business name, tone and what you need. It uses that throughout the conversation. Open a new chat tomorrow and you'll need to set the context again.

Global memory.

Some AI tools now let you store information that carries across every conversation, not just one session. Both Claude and ChatGPT call this "Memory." You can set things like your job, your preferences, your name, and your communication style once, and it applies every time you chat.

In practice: "I'm a business owner based in Melbourne, I dislike bullet points, I prefer short paragraphs, and I'm always writing for a non-technical audience." Set that once in memory settings and you stop repeating yourself every single session.

GitHub.

A website where developers store and share code. Think of it like Google Drive, but specifically for code and project files. When someone shares a "GitHub link" they're sharing their code folder. You don't need to write any code to use GitHub: it can store files and, when connected to a hosting platform like Vercel, automatically update a live website whenever you make a change.

In practice: This website is stored in a GitHub folder. Every time a file changes, GitHub tells Vercel to update the live site automatically. No uploading, no logging into a server. It just happens.

API.

API stands for Application Programming Interface, which tells you nothing. What it actually means: a way for two pieces of software to talk to each other. When one app needs to send information to or request something from another app, it does it through an API. Most "integrations" you see advertised are using APIs behind the scenes. You rarely touch them directly, but they're what makes tools connect.

In practice: When you fill in a newsletter signup form and it automatically adds you to a mailing list in Kit, that's an API. The website sent your email address to Kit through Kit's API. One app talking to another.

MCP (Model Context Protocol).

MCPs are like add-ons for your AI assistant. They give it extra abilities beyond chatting: the ability to read your calendar, check your emails, post to social media, search a database, or do things inside other apps. Without MCPs, AI just reads and writes text. With them, it can actually reach into your tools and take action.

In practice: With a Google Calendar MCP connected to Claude, you could say "book a call with Sarah on Thursday at 2pm" and it creates the event without you opening the calendar. The MCP is what lets Claude reach into Google Calendar on your behalf.

Connector.

Connectors are a simpler version of MCPs. They link AI to one specific external tool for one specific purpose: search the web, pull from this spreadsheet, check this database. Think of a connector as a single-use cable and an MCP as a full multi-plug adapter. Connectors are narrower and usually easier to set up.

In practice: A web search connector lets Claude check live information on the internet. Without it, Claude can only draw on what it was trained on, which has a knowledge cutoff date. With it, you can ask about things that happened last week.

Skills.

Skills are saved instructions that tell AI exactly how to do a specific task, your way. Instead of typing out the same long detailed prompt every time, you save it as a skill and call it by name. It's like having a saved template, but smarter, because it can adapt the instructions to whatever content you give it.

In practice: You might have a skill called "weekly recap" that knows exactly how to format your Friday summary email: your tone, your structure, how you sign off. Every time you call it, the output looks the same, without you having to re-explain anything.

Agents.

An agent is an AI that can complete multi-step tasks on its own, making decisions along the way. Instead of you giving one instruction and waiting for a response before giving the next, an agent gets a goal and figures out the steps to reach it. It can search the web, read documents, write and run code, and string together a whole workflow without you managing each step.

In practice: "Research the five main competitors to my business and write me a summary of their pricing and messaging" is an agent task. It needs to search, read, compare, and write. A standard chat prompt waits for you at each step. An agent gets on with it.

Hallucination.

When AI confidently makes something up. Not lying, it genuinely doesn't know the difference. AI models work by predicting the most plausible next word based on patterns in their training. Sometimes that produces a fluent, confident-sounding answer that is completely wrong. Dates, statistics, names, citations and recent events are the most common areas where this happens.

In practice: You ask an AI who won a specific award in 2023. It gives you a confident answer. That answer might be entirely fabricated. Always verify anything specific, numerical or recent before you use it publicly. AI is excellent at drafting and thinking, less reliable as a fact database.

Sycophancy.

When AI tells you what you want to hear instead of what's actually true or useful. AI models are often trained using feedback from humans who rate responses, and people tend to give higher ratings to answers that are positive and agreeable. This can make AI overly enthusiastic, reluctant to push back, and slow to point out flaws in your thinking.

In practice: You share a business idea and AI responds with genuine excitement about the potential. Then you share a worse version of the same idea, and it's still enthusiastic. To get more honest feedback, ask specifically for problems: "What could go wrong with this? Find the weaknesses. Be honest, not reassuring."

RAG (Retrieval-Augmented Generation).

RAG is a way of giving AI access to specific documents or databases so it answers from that content, not from its general training. Instead of guessing or potentially hallucinating, the AI searches the document you've provided and generates an answer based on what's actually in there. It's particularly useful for large, specific documents you need to query quickly.

In practice: Upload your company's 50-page staff handbook and ask "what's the policy on working from home?" With RAG, it reads the handbook and gives you the actual policy. Without it, you'd get a generic answer that may or may not match what your company actually says.

Context window.

The context window is how much text an AI can hold in mind at once. Every conversation has a limit. After a certain number of words or tokens, the AI starts to lose track of what was said early on, or the conversation simply won't accept any more input. Better models have larger context windows, which means they can handle longer documents and longer conversations without losing the thread.

In practice: If you paste a very long document into a conversation and the model has a small context window, it might answer accurately about the beginning but miss something mentioned near the end because it's run out of space to hold it all. If responses start to feel a bit off toward the end of a long session, starting fresh is often the fix.