What Claude actually is

Claude is two things sharing one name, a model that turns text into text and an app around it that decides what the model gets to see, and telling them apart explains most of what is confusing about using it.

There is the model, which is the part that produces text. It has no memory, no clock, no access to anything, and no existence between the moment you press send and the moment the answer finishes. It is a function: text in, text out.

And there is the app around it, which is everything else. Your conversation history, the files you upload, the settings you set, the ability to search the web. The app decides what the model sees, hands it over, and shows you what comes back.

Nearly every “why did it do that” question has the same answer: because of what the app did or did not put in front of it.

Everything it knows about your question is in the conversation

The model reads your entire conversation, from the top, every single time you send a message. That whole thing is called the context.

Which means:

It cannot see anything you have not shown it. Not the file on your desktop, not the email you are describing, not the thing you told it last Tuesday in a different chat.

It cannot remember your last message except by re-reading it. Nothing is retained between messages. The conversation is re-read from the beginning each time, which is why a long chat gets slower and more expensive as it goes.

”It forgot” usually means “it was never told”. That is lesson 4.

There is one deliberate exception. Claude has a memory feature that saves facts about you between conversations. It is not a recording of your chats; it is a set of individual entries, and Settings > Memory lists all of them grouped by category, where you can delete any one of them or turn the whole thing off.

Whether it is on right now depends on your account, so lesson 2 has you go and look rather than telling you. Everything else starts empty.

It does not give the same answer twice

Ask the same question three times, in three new chats, and you get three different answers. Here are three real ones, same question, same model, seconds apart:

A stock take is a physical count and verification of all inventory a business currently holds, to check it matches recorded stock levels.

A stock take is the process of physically counting and verifying a business’s inventory on hand against its recorded stock levels.

A stock take is the process of physically counting and verifying the quantity of inventory a business currently holds, typically to reconcile it against recorded stock levels.

The same answer, worded three ways. That is the normal case and it is harmless.

The case that is not harmless: when the question is hard, the three answers can differ in substance rather than in wording, and there is nothing on screen to tell you which kind of question you asked. If it matters, ask twice in two fresh chats and see whether the substance holds. That takes twenty seconds and it is the cheapest quality check there is.

This is also why “it worked yesterday” is not evidence of anything, and why sending the same message again after a bad answer sometimes just works. You are not fixing anything. You are rolling again.

What is Claude good at?

It is good at anything where the input is language and the output is language, and where you can tell whether the answer is right. Rewriting, summarising, explaining, translating between formats, drafting a first version of something you will edit, extracting structure out of a mess.

It is bad at anything where being subtly wrong looks the same as being right and you have no way to check. That sentence is the whole of lesson 9, and it is worth carrying from here.

What this course assumes

Part 1, which is these ten lessons, needs nothing but a browser, and works on a free account. Where a feature needs a paid plan, the lesson says so at the point it comes up.

Part 2 is for developers and needs a paid plan and a terminal. If that is not you, part 1 is a complete course on its own and you can stop at lesson 10 without missing anything that depends on part 2.

Your turn

Open Claude and start a new chat. Ask it something with a factual answer you can check.

In one sentence, what is a stock take?

Read it. Then start a second new chat, not a follow-up in the same one, and paste the identical question.

Check: the two answers say the same thing in different words. Now, still in the second chat, ask “what did I ask you a minute ago?”

Both outcomes are worth having. If it does not know, you have seen that a conversation is sealed: nothing carried across. If it does know, memory is on and something was saved, and the useful next step is Settings > Memory to read the entry it made. What it saved will be a fact about you, not a transcript of the message, and seeing the difference is worth more than the original answer was.

Then a third thing, in a fresh chat: ask it something about a document on your computer without attaching the document.

What does the first paragraph of my employment contract say?

Check: it tells you it cannot see the file. That sentence is the whole lesson: it knows nothing you did not put in the conversation.

Recap

Two things share the name. The model turns text into text and has no memory, no clock and no access to anything. The app decides what the model gets to see.

Everything relevant to your question has to be in the conversation, because the conversation is all there is. “It forgot” nearly always means “it was never told”.

The same question gives different answers each time. Usually that is just wording. When it matters, ask twice in two fresh chats and compare the substance.

Next: getting set up, and which of the three ways to use it is right for which job.