Somebody said "token" and she nodded
It was in a Facebook group, or a supplier chat, or at a trade day. Whoever said it clearly assumed everyone already knew, so she nodded, the conversation moved on, and she has been quietly guessing ever since. Something to do with credits? A bit like a coin?
Everybody has done this. We have done this. Nobody gets the AI terms explained to them up front, least of all a small business owner who has a salon to run.
There is no version of learning a new field where you do not spend a stretch nodding at words nobody gave you, and the reason nobody explains them is not that they are difficult. It is that once you know a word, you cannot remember not knowing it.
So nobody goes back and fills the gap. This page is the gap being filled.
AI terms explained: the nine that actually matter
There are nine words that do almost all the work in this field: prompt, context, model, terminal, skill, permission mode, hallucination, MCP, and token. Every one is a plain idea wearing a technical coat. Learn these and you can follow any conversation about AI without nodding at anything.
Every example below belongs to Steph, a two-groomer salon in Sydney's Inner West. She is a worked example we follow through this series, and her numbers are the sort you would recognise rather than a real client's books.
The rest of this series links back here on purpose. No other post stops to define these words, so read the one you came for and come back when the next word turns up.
Prompt
A prompt is what you type. That is the whole definition: the instruction, plus anything you paste in with it.
It is not a magic phrase and there is no official list of them. That matters because the biggest difference between a useless answer and a usable one is almost never which product you opened. It is how much you put in the prompt.
Steph wants a caption for the photo of Ziggy after her Saturday groom. "Write me an Instagram caption" gets her something a stranger wrote about a dog they have never met. The same request, carrying Ziggy's name, the coat, the service, and one sentence in the way she actually talks, gets her something she only has to tidy.
Same product, same minute. The difference is entirely in the prompt.
If you would rather paste prompts than write them, 20 prompts that reclaim a groomer's week is where to start.
Context
Context is everything it can currently see: this conversation, the files you have shown it, the notes you have given it. Nothing else.
That last part is the one that catches people out, so it is worth being blunt. It cannot see anything you have not shown it. Not your booking system, not your inbox, not your diary, and not last Tuesday's conversation unless you open it again. If it did not arrive in the prompt, it does not exist.
Context is the notes Steph did not hand over. She asks for a message to Peanut's owner about how the muzzle went, and gets back something generic and slightly wrong, because Peanut's handling notes live in her booking system and she never pasted them in. The tool was not being thick. It answered the only question it could see.
Two things follow. What you must never show it is in what you must never paste into an AI chatbot. How to show it the same things once rather than every time is in teaching it your prices, policies, and voice.
Model
The model is the engine. Same product, different engine underneath.
Every one of these tools has more than one, and a picker somewhere to change it. Claude Code, for instance, switches with a /model command between engines labelled sonnet, opus, fable, and haiku, where haiku is the fastest and lowest cost and fable is the most capable. ChatGPT and Gemini do the same thing under their own names.
Do not memorise any of them. They change every few months.
What survives the name changes is the trade: faster and cheaper, or slower and better. Steph's caption does not need the careful engine. Her reply to the owner who is upset about how short Moose came out does.
There is a more technical definition of a model and it will not change a single decision you make. For which tool suits which job, read the decision chart.
Terminal
The terminal is a plain text window on your computer where you type commands instead of clicking. That is all it is.
It is already installed. On a Mac it is called Terminal, on Windows it is PowerShell. It looks unfriendly because there are no pictures in it, not because it is dangerous, and nothing you type happens until you press Enter. More people are put off by this one window than anything else on this page, and the fear is entirely about how it looks.
Steph has never opened it, and nothing at this level needs her to. The only reason she would ever meet it is Claude Code, which runs there. Even then there is a way around: Anthropic publishes a Claude Code desktop app for Mac, Windows, and Linux, plus a web version on a paid plan. The text window is a preference, not a toll gate.
When you are curious, what Claude Code actually is explains it without assuming anything, and the honest install guide is the only place in this series with the steps.
Skill
A skill is a written-down job. You describe how a task gets done once, in a file, in plain English, and the tool follows that description every time after.
The test for whether you want one is not technical. Anthropic's own documentation puts it plainly: you write a skill when you notice you are re-pasting the same instructions over and over.
A skill is the job Steph has explained to Priya three times. What goes in a handover note, in what order, which parts the owner sees and which stay internal, and the one thing that never goes in. Explained out loud three times, and still slightly different every time it comes back. Written down once, it comes back the same shape in March as it did in November.
The difference from a prompt is only ever this: a prompt is this once, a skill is this every time. How to build one, with a real one you can read end to end, is in teach it a job once, run it forever.
Permission mode
A permission mode is the setting that decides what the tool is allowed to do on your computer without asking you first.
It only becomes a live question once something is working on your actual files rather than inside a chat window, which is the level above this one. The idea is worth having early, because it separates two very different arrangements: ask me before you touch anything, or go ahead and get on with it.
Reading Steph's price list and rewriting the files on her desktop are not the same risk, and the permission mode is where that difference lives. There are six of them in Claude Code, they run from cautious to genuinely reckless, and exactly one post in this series names them and tells you which to use: what to allow, and what never to. Read it before you change any setting.
Hallucination
A hallucination is when it states something confidently and the thing is not true.
It is not lying, because lying needs an intention. It is not a bug you can report and have fixed. It is a property of how these tools work, and the reason it catches people is that there is no tell.
An invented council rule arrives in exactly the same calm, well-organised sentence as a true one. Anthropic's own documentation says the unvarnished version: confident-sounding wrong answers remain possible.
So the answer is never to go and find a tool that does not do it, because they all do it. The answer is a habit. Anything you would not say to a client without checking, check: a price, a rule, a date, or a claim about a breed.
Why it happens, and the 10-second check that catches most of it, is why it gets things wrong.
MCP
MCP stands for Model Context Protocol, and the only useful way to picture it is a standard plug socket: an agreed shape that lets an AI tool talk to another piece of software directly, instead of you carrying things between the two by hand. The specification behind it is a developer document you will never need to read.
Now the honest part, which most articles about MCP leave out. A socket needs two sides. If the software you want to reach has not built one, MCP does nothing for you, and no amount of setting up will change that. Plenty of the tools a pet business runs on have not built one.
So the route that works today is the unglamorous one: export the file, then hand the file over. That is what asking your business a question works through, and it works whether or not anybody builds a socket.
One caution, because it sits in Anthropic's own risk list: an MCP server you do not trust can run commands on your machine. Connect the ones you have a reason to trust, and not one more.
Token
Here is the word from the top of the page.
A token is the unit these tools count in. Not letters, not words: roughly three quarters of a word, close enough for any decision you will make. The precise definition is fussier and you do not need it.
Tokens matter for two reasons. First, everything has a ceiling measured in them: how much you can put in front of it at once. OpenAI translates its own ceiling into pages on its pricing page: about 12 pages of text on the free tier. Google publishes Gemini's ceilings in tokens instead, at 32k, 128k, and 1 million by tier.
Second, if you pay per use rather than a flat monthly fee, tokens are the unit on the invoice, and what you send in is priced separately from what comes back.
Which is why pasting the whole year costs more than pasting one week. Steph's salon sees around 40 dogs a week, so a year of appointments is roughly 2,000 rows. Paste the year to ask about last week and she has paid for 1,960 irrelevant rows, and crowded out the room the real question needed. Send the week.
Both ceilings will have moved by the time you read this. Current figures live in what AI actually costs.
The AI terms a small business owner can ignore
You will meet others: agent, LLM, fine-tuning, RAG, temperature, and parameters. None of them change what you do on a Tuesday. If a post here genuinely needs one, it explains it where it uses it. Nothing in this series gets used that has not been explained or linked.
Which is why the nine above earn their 10 minutes. They are not trivia. Each one is a decision you are already making without realising: what to put in the prompt, what it can see, which engine, what it may touch, and what to check before you send.
What this actually reclaims for you. The thing Steph is chasing back is the admin evening: two hours, twice a week. Across the 48 weeks she trades that is 192 hours, and not one of them comes back while she is nodding at a word and guessing at the rest of the sentence. This page costs 10 minutes. Against 192 hours, it is the cheapest thing you will ever spend.
Next on the ladder is the one that sounds like it belongs to somebody else, right up until it doesn't: what Claude Code actually is, if you have never written code.