Loading interactive tool…
This page in plain text
AI glossary: 44 terms in plain words
The terms you will meet when you use AI, each in one or two sentences, with the page of this site to read next. Six to know first: Brief, Token, Context window, Hallucination, Agent, Skill.
The basics
- AI model. The part that does the work: a program trained on a very large amount of text to predict what comes next. Everything else is built around one. Which Claude model.
- Large language model (LLM). A model trained on text, and the kind behind ChatGPT, Claude and Gemini. When people say “the AI”, this is usually what they mean. The tools we use.
- Chat assistant. The app you type into. It wraps a model in a conversation and adds extras such as file uploads, search and memory. The tools we use.
- Prompt. What you send to the model. It can be one line or a page, and the answer is only as good as what went in. The prompt library.
- Brief. Our word for a prompt with all its parts: who the AI is, the goal, the facts it cannot guess, the shape of the answer and how to tell it worked. How to write a brief.
- Token. The unit a model reads and writes in: a short word, or a piece of a longer one. Limits and prices are counted in tokens, not words. Which Claude model.
- Context window. How much the model can keep in view at once: your messages, its replies and any documents you gave it. In a very long conversation the earliest parts can fall out of view. Turn a long document into a short one.
- Hallucination. An answer that is wrong or made up, given with complete confidence. The model is not lying; it is filling a gap with something that sounds right. Research a question and see the sources.
- Training data. The text a model learned from before you ever used it. It shapes what the model knows and what it gets wrong. Free AI courses.
- Knowledge cutoff. The date a model’s training stopped. It knows nothing later than that unless it can search the web or you tell it. The AI timeline.
Asking well
- Role. Who you tell the AI to be for a job, such as “an editor who cuts hard”. It sets the standard the answer is held to. How to write a brief.
- Context. In a brief, the facts the AI cannot guess: who the reader is, what you already know and what you have tried. How to write a brief.
- Output format. The shape you ask for: how long, what layout, what tone and what to leave out. Unsaid, you get the default. The brief builder.
- Example (few-shot). Showing the model one or two samples of what you want, where describing it would fall short. Two things you wrote teach it your voice faster than any description. Write faster and still sound like yourself.
- Follow-up. A second message that says exactly what is off in the first answer. Two or three rounds beat one long prompt. How to write a brief.
- System prompt. Standing instructions a model is given before the conversation starts. The person chatting usually never sees them. Answer customers on your own website.
- Custom instructions. A system prompt you write yourself in a chat assistant’s settings, so it knows your job and preferences in every conversation. The AI timeline.
- Prompt library. A collection of finished prompts to copy and change. A good one saves you writing the brief from nothing. The prompt library.
Models
- Model size. Makers sell small, medium and large models. Small is fast and cheap; large is slower, costs more and handles harder work. Which Claude model.
- Reasoning (thinking). A mode where the model works through a problem before it answers. It is slower and better on hard questions, and wasted on easy ones. The AI timeline.
- Multimodal. Able to take in more than text: images, audio or files. You can show it a photo or a chart and ask about it. The AI timeline.
- Voice mode. Talking to the assistant and hearing it answer, with no typing. Get ready for a job interview.
- Image generation. Making a picture from a written description. Expect to make several and keep one. Make an image for a post or a deck.
- Text to speech. Turning a written script into a spoken voice. Changing a line means retyping it, not recording it again. Record a voiceover without a studio.
- Deep research. A mode where the assistant searches and reads for several minutes, then writes a report with its sources. Research a question and see the sources.
Building
- API. A way for your own software to use a model directly, paid by use, with no chat window. It is how a chat gets onto a website. The Claude API.
- Agent. An AI that takes steps by itself to finish a job: it reads, acts, checks the result and goes again, where a chat answers once. Build a site or a tool without a dev team.
- Coding agent. An agent that works in your code: it reads the files, runs commands and makes the changes. You say what it is for and look at what comes back. Claude Code.
- Tool. Something a model can call to act and not only talk: search the web, read a file, run code. The model decides when to use one. Connect your first MCP server.
- MCP (Model Context Protocol). A shared standard for connecting tools and data to an AI assistant. One connection works with any assistant that speaks it. MCP servers.
- MCP server. A small program that hands an assistant a set of tools for one thing: a folder, an account, a database. Connect your first MCP server.
- Skill. A way of doing one job, written down once in a text file, that Claude picks up by itself when the job comes along. Write your first Claude skill.
- Instruction file (CLAUDE.md). A text file of standing rules that a coding agent reads at the start of every session in a project. It is where lessons get written down. Claude Code.
- Retrieval (RAG). Looking things up first, then answering from what was found. It is how a chat answers from your pages and not from memory. Answer customers on your own website.
- Embedding. Text turned into numbers so that similar meanings sit close together. It is what makes search by meaning work. Free AI courses.
- Fine-tuning. Training a model further on your own examples. It is rarely the first thing to try; a better brief usually is. How to write a brief.
- Rate limit. A cap on how much you can ask for in a period of time. Hit it and you wait, or you move to a bigger plan. The Claude API.
Trust
- Source (citation). A link to where a claim came from. It is not proof until you open it and it says what the answer says. Research a question and see the sources.
- Check. The part of a brief that says how either of you can tell it worked. Without one, the first draft that sounds fine is the one you get. How to write a brief.
- Human in the loop. A person looks at the work before it is used, sent or published. The AI drafts; someone answerable decides. Build a site or a tool without a dev team.
- Personal data. Names, emails, account numbers and anything else that points to a real person. Take it out before you paste or upload. Make sense of a spreadsheet.
- Guardrail. A rule that limits what an AI will do: a subject it stays off, a promise it must not make. Answer customers on your own website.
- Prompt injection. Text hidden in a page, file or email that tries to give the AI instructions. It is a risk whenever an AI reads something you did not write. MCP servers.
- Bias. A slant a model picked up from the text it learned from. It can show in who it assumes you are and what it treats as normal. Free AI courses.
Ask Another AI Guy
I'm just Another AI Guy.
You're chatting with our fast model. Create a free account for the more thorough one, and more messages.