📖 Every AI Word That's Been Confusing You, Explained

AI has about six real ideas and thirty words for them. Most of the words exist to make simple things sound expensive.

This page is alphabetical, so find your word, read one line, get on with your day. Nothing here needs learning, and the prompt at the bottom handles anything that isn’t on the list.

The A to Z

Agent, agentic: AI that takes several steps on its own instead of just replying. Most things sold as agentic are, on inspection, a chatbot.

AGI: a hypothetical future AI as capable as a person at everything. Not a product you can buy. Podcast fuel.

Algorithm: a recipe a computer follows. Now mostly used to mean “computer thing I don’t trust”.

Artifact: the side panel where Claude builds you an actual thing, a document or a working tool, instead of just talking about it.

Chatbot, assistant: the chat window you type into. Same product, different marketing meeting.

Connector: a plug between Claude and another tool you use, like Gmail or your calendar, that only does what you approve.

Context: everything Claude can see right now: this chat, what you pasted, what it remembers about you. If it isn’t in the context, it doesn’t exist.

Context window: how much fits in the context. Very long chats push the early bits off the end, which is why they go vague and a fresh chat perks things up.

Custom instructions: your standing orders, written once in settings and applied to every chat. In Claude these live in your profile.

Effort: a setting in the model menu for how hard Claude thinks before answering. Lower effort is faster and stretches your allowance further.

Fine-tuning: retraining a model on specific material so it’s sharper at one job. Factory work. What you actually want is instructions, which take a minute.

Generative AI: AI that makes new things rather than sorting existing ones. Describes everything in these guides. Mainly a word for conference banners.

GPT: an acronym that escaped and became a brand name, like Hoover. Nobody will test you on it.

Hallucination: a confident answer that happens to be wrong. It can’t always tell the difference. Check anything that matters.

LLM: large language model. Means: the thing you’re already using.

Memory: notes Claude keeps about you between chats, so Thursday knows what Monday learned. On by default for most accounts, and you’ll find it in Settings, then the memory option.

Model: the brain behind the chat window. Claude comes in a few, from quick and light to slow and heavyweight.

Model picker: the little menu next to the send button where you choose the model, the effort and whether it thinks first.

Open source: models anyone can download and run, as opposed to the big commercial ones. Genuinely important in the industry, changes nothing about your Tuesday.

Project: a folder for one area of your work, with its own files and instructions that every chat inside it knows about. One of the things the paid plan (Pro, around £20 a month) is for.

Prompt: your message. Any message. The most over-mystified word in the building.

Prompt engineering: asking well. For normal people it means: be specific, give context, say what you want back. You need practice, not a course.

Reasoning, thinking: the model working through steps before it answers. Slower, better on tangled problems, wasted on rewording an email.

Skill: your own instructions installed into Claude once, so it follows your process without being told again.

Style: a saved tone (short, plain, formal) you pick in a chat instead of typing “less waffle” for the ninth time this week.

Token: the chunks text gets chopped into for reading and writing, roughly syllable-sized. Explains the typewriter effect. Otherwise ignorable.

Training data: what the model read before you ever met it. Which is why it knows a bit about nearly everything and not this morning’s news.

Usage limits: the cap on a free plan. It resets after a few hours. You aren’t banned, you were busy.

Voice mode: talking to Claude out loud on the phone app. Odd for ninety seconds, then obvious.

Wrapper: an app that’s one of the big models in a costume, usually with a subscription on top. Worth knowing before you pay for one.

When the word isn’t on the list

New jargon breeds faster than any glossary, so here’s the bit that doesn’t go out of date. Save this one.

Nick this one:

You're my jargon translator. I've hit some tech language I don't understand and I'd rather not nod along again.

Here it is: [paste it: for example "our agentic RAG pipeline delivers enterprise-grade LLM orchestration for your team"]
Where it came from: [for example "the pricing page of a tool a client wants me to use"]

Give me:
1. The whole thing rewritten the way a friend would say it out loud, with no jargon left in it.
2. Each bit of jargon, one line each, and whether it's a word I need or decoration.
3. The one thing in there, if any, that changes what I'd actually do.
4. A short version of this prompt I can save and re-run next time something like this lands.

Rules: plain English, under 150 words before the saved prompt, and don't pretend to know a term you're unsure of, say so instead. Ask me up to three questions before you start.

Sales page, client email, the bit of the invoice software you’ve never understood. Same prompt every time.

The honest bit

These are working definitions, not textbook ones, and three new words will be invented by Friday. That’s what the prompt is for.

Today

Save the prompt in your notes app. Owning a translator beats owning a dictionary, and the translator doesn’t date.

If the confusing thing is a whole document rather than a single word, that job has its own guide: Paste Any Techy Thing In Here And Get English Back.

That’s the vocabulary sorted, which is the easy half. The Claude Kickstart is the doing half: Claude set up around your actual work, five real jobs off your plate, and a playbook for your trade. £47, one weekend.