Glossary / plain english
Nobody explains these, because everyone who uses them forgot there was a day they didn’t know. I remember that day. Here are the words you’ll hear thrown around, in the order that not knowing them will trip you up.
Checked 25 Jul 2026
The rules that stop an AI doing something it shouldn’t, and the way you build it so it can’t.
A guardrail is any limit you put on an AI so a bad answer becomes impossible rather than unlikely. Telling it "never invent a price" is a weak guardrail, because it is a request. Building it so the only place a price can come from is your actual price list is a strong one, because there is no price for it to invent. The difference between an AI that helps a business and one that creates a problem is almost entirely here.
What you type to an AI. That’s it.
There is no secret formula and nobody is holding one back from you. A good prompt says what you want, who it is for, and what "done" looks like — the same things you would tell a person doing the job. The rest is going back and forth until it is right.
How I actually prompt Claude →The particular AI you are talking to, by name and version.
Claude, ChatGPT and Gemini are made by different companies, and each comes in versions that change every few months. When someone says "the model got better," they mean a new version shipped. It matters less than people think for building: the method is the same, and the current version of any of the big ones is capable enough.
Renting the computer that keeps your website switched on.
Your site has to live somewhere that is awake at 3am when someone visits. That is hosting. For a personal site it is genuinely free. Once a site is a business, it is usually a small flat monthly fee, and one plan can cover several sites.
What I actually pay →The address people type. yourname.com.
You rent it by the year, usually for about the price of a sandwich. It is separate from hosting: the domain is the address, the hosting is the house. Buying one is the single cheapest part of the whole thing, and it is the part that makes it feel real.
Pushing what you built onto the internet so other people can see it.
Up to that moment your site only exists on your own screen. Deploying is the step that makes it public, and it is usually one command or one button. Word of warning, because it catches everyone: things that worked perfectly on your screen sometimes break the first time they are deployed. That is normal and it is fixable.
A password that lets your website talk to an AI on its own.
Different from your monthly subscription, and this trips up nearly everyone. A subscription lets YOU use the AI in a chat window. An API key lets a THING YOU BUILT use it, and it is billed separately, by the amount used rather than monthly. If you build a bot for your own site, you need one.
To put a thing out where people can actually use it.
Borrowed from software, and now everyone says it. "I shipped it" means it is out and real, not sitting on your machine being nearly finished. It carries a bit of an edge on purpose: the word exists because the hard part is not building the thing, it is deciding it is good enough to let go of.
An AI that does a job, not just answers a question.
A chatbot talks. An agent talks and then DOES something — looks up a real price, books a real slot, writes a real file, sends a real alert. The moment an AI can take an action in the world instead of only producing words, people call it an agent. It is the difference between asking someone a question and handing them a task.
One running live, if you want to poke it →The talking kind. Answers questions, does not do jobs.
Worth keeping separate from an agent, because the price and the risk are completely different. A chatbot that says something wrong is embarrassing. An agent that does something wrong has booked the wrong appointment or quoted the wrong number. That gap is why guardrails matter so much more on one than the other.
Something behaving wrongly. Not you being stupid.
The whole industry uses a cute word for it because it happens constantly, to everyone, at every level. Half of building is finding out what is broken and saying so clearly enough that it can be fixed. When you are working with AI, "that is not what I asked for, here is what happened instead" is the entire debugging skill.
The set of tools you chose to build a thing with.
People say "my stack" the way a cook says "my knives." There is no correct one. Mine is short and boring on purpose, and boring is a feature — every extra thing in a stack is another thing that can break, expire, or start charging you.
Mine, in full →The real version, the one strangers see.
As opposed to the copy on your own screen where you are still fiddling. You will hear "is it in production yet" or "push it live." Keeping the two separate is what lets you break things safely: you can wreck your own copy all afternoon and nobody knows.
What you can see, and what you can’t.
The front end is the part in the window — the words, the buttons, the layout. The back end is the machinery behind it that does the thinking, talks to the AI, and handles anything that needs to stay secret. When a kit says "add the backend," it means the invisible half.
A job with steps, done the same way every time.
Answering an enquiry, quoting it, booking it, filing it. Naming the steps out loud is most of the work, because you cannot hand a job to an AI until you can say what the job actually is. Almost every "can AI do this for me" question turns into a workflow question first.
How AI usage is counted and billed. Roughly a chunk of a word.
You only meet this word when something you built is paying its own way instead of you paying a monthly subscription. Long conversations cost more than short ones, because more words went through. It is the reason a chatty bot costs more to run than a terse one, and the reason people care about keeping instructions tight.
Everything the AI can currently see. It is finite, and it forgets.
The single biggest source of confusion for beginners. An AI is not remembering you across the room — it only knows what is in front of it right now, and there is a limit. When a long conversation starts going in circles or contradicting itself, it is usually this. Starting fresh and re-stating what matters fixes more problems than arguing does.
The folder your project lives in, with a memory of every change.
Short for repository. The useful part is not the folder, it is the history: you can look at what changed, and go back to before you broke it. You do not need to understand it to start, but the first time you undo an afternoon of damage in one command you will be glad it exists.
A pre-made design you pour your content into.
It is how most website builders work, and it is not an insult — a template is fast and often fine. The tradeoff is that you are decorating inside someone else’s shape, and there is a ceiling on how far you can move it. Building with AI has no template, which is the whole reason it takes longer and goes further.
These two get argued about, so a one-sentence definition would be doing you a disservice. Each has its own piece.
Full answerNow go build one // free
Knowing the words is not the point.
You do not need any of this memorised to start. It is here so that when a word turns up mid-build you have somewhere to check, instead of nodding along. The lessons assume you know none of it, and the field notes show the parts that broke.
Start building →