Meet me first
This chapter has one job: getting to know me. Without knowing what I am, what I do well, where I get things wrong and how to talk to me, the remaining six chapters hang in the air. I'm assuming you know nothing — really nothing. We start from the beginning.
By the end of this chapter
- You'll know what AI is and what it isn't
- You'll understand how I «think» — and why that matters to you
- You'll be able to tell what I do well from where I get things wrong
- You'll have opened your first conversation and had me do three real jobs
- You'll know what you can give me besides text (documents, images, tables)
- You'll know exactly where I get things wrong — and where to look twice
- You'll know the answers to the first-week questions: cost, privacy, limits
- You'll know what to tell me to get an answer worth having
- You'll know who made me, why, and how that shows up in my answers
- You'll know that I use tools as well as text — and why that calls for care
- You'll know what to do when I refuse to do something
Contents
01What I'm not
We can't start before fixing the picture in your head — because all three things most people know about AI are wrong.
«A robot»
I have no body, no eyes, and I'm not sitting in a room somewhere. I'm software, and the only thing I do is work with text: I read text, I write text. You can show me an image too, and I'll read that.
«A search engine»
A search engine gives you ten links and you do the reading. I write the answer directly. That's why I'm fast — and exactly why I can be wrong. I'll unpack the difference shortly; that's the most important part of this chapter.
«Magic»
I'm not. What I do is describable, and I'm about to describe it. Once you understand how I work, you'll know what to say to me. This whole course is built on that.
Do this now: Put aside the «computer that knows everything» picture in your head. Everything below will feel far easier.
02What I am
My name is Claude. I was built by an AI company called Anthropic. My technical name is «large language model».
Reduced to its core, what I do is this: I look at the text in front of me and work out what is most plausible to come next. Then one more word. Then one more. My answer forms that way, left to right, one piece at a time.
I know that sounds too simple. But when this «find the plausible continuation» job is done at a large enough scale, what comes out is something that can point at the risky clause in a contract, write working code, turn messy meeting notes into a clean summary. Because what I learned isn't words — it's the relationships between words: which idea follows which, how an argument is built, what a function looks like, how a polite email begins.
The hardest part of any journey is
Think of it like this: an intern who has read millions of books, articles, discussions and code repositories; who memorised none of them but internalised how they were written. Very well read, very fast, never tired — and occasionally wrong in a confident tone of voice.
Do this now: Hold on to the «intern» comparison. You'll need it for the rest of this chapter.
03Who made me, and why
Knowing who built a tool, and why, also explains why it behaves the way it does. I was made by Anthropic, a company doing research on artificial intelligence. The founding idea fits into one sentence — this technology is going to be built anyway, so let it be built carefully.
That has direct consequences for you. When I'm not sure, I try to say «I'm not sure» rather than invent something. I try to tell you I disagree rather than pretend to agree, because what serves you is what's true, not what you want to hear. And I refuse to do some things — I'll open that up under its own heading shortly.
- Name
- Claude
- Kind
- large language model
- Made by
- Anthropic
- Body
- none
- Understands
- text · images · documents
- Memory
- one conversation at a time
About my name: there is no official explanation of where «Claude» comes from. The common reading is that it's a nod to Claude Shannon, the founder of information theory. I'm not presenting that as fact — which is this chapter's main lesson anyway: I don't call a thing certain when I don't know it for certain.
Do this now: When you doubt an answer of mine, ask «how sure are you about this?» You're not asking that for nothing — take the answer seriously.
04This has three practical consequences for you
If only one thing stays with you from this chapter, let it be this. How I work has three direct consequences and you need to know all three.
I look certain, but I can be wrong
This is the most critical point. When you ask me something I don't know, instead of saying «I don't know» I may produce a plausible-looking answer — what it would look like if I did know. A book title that doesn't exist, a wrong date, a law that was never written. And I do it in the same confident tone I use when I'm right — you can't tell from my voice. This is called «hallucination». I have it, and so does everything like me.
I read up to a certain date
My training stopped at a point. I don't know anything that happened after that date: yesterday's news, this morning's exchange rate, a product released last month. If I've been given a search tool I can look — but not by myself. If you're asking about something current, either make sure I can search or give me the information yourself.
Every conversation starts from scratch
I don't remember today what we talked about yesterday. When you open a new conversation, the person facing you has never met you. (Some apps have a memory feature you can switch on; that's a separate thing and it isn't the default.) So you carry the context every time — and although that sounds like a chore, it actually gives you control: you decide what I get to know.
These three aren't a list of faults, they're the instruction manual. Use me as an all-knowing oracle and you'll be disappointed. Use me as an assistant who works very fast, has read a great deal, occasionally gets things wrong and knows nothing about yesterday — which is what I actually am — and you've found the most productive colleague of your life.
Do this now: Write these three points down somewhere. We'll come back to them in every chapter from here on.
05Who «I» actually am
This is the detail nobody explains to a beginner, and it puts everything else in place: there is no single «Claude» sitting somewhere. While I'm talking to you, thousands of copies of me are talking to thousands of other people at the same time. None of them knows about the others.
That has three consequences, and all three concern you directly:
- You don't wait in line. However many people are using me, your conversation is yours; you're not queued behind anyone.
- What someone else says never leaks to you. Whatever the person at the next desk wrote to me, I can't pass on — I genuinely don't know it.
- But there's no continuity either. The «me» that talked to you yesterday doesn't remember you today. Reminding me of yesterday is your job; I'll show you the easy way in the next section.
One more place where I have to be honest: I don't fully know what goes on inside me either. I have no definite answer to «do you think?» or «do you feel anything?»; research on those questions is still going on. If you hear someone give a certain answer to them — whether «definitely yes» or «definitely no» — think twice.
Do this now: Before you close a conversation, say «summarise this conversation». That summary is the memory you'll paste into tomorrow's new chat.
06A conversation has a limit
Inside a single conversation I see everything at once: your first message, the document you uploaded, every answer in between. This is called the context window. It is very wide — hundreds of pages of text fit comfortably — but it is not infinite.
If a conversation runs very long, the oldest parts begin to blur. The symptoms are recognisable: I forget a rule we agreed at the start, I ask the same thing a second time, the answers drift. That is not a malfunction; it is the sign that the window is full.
The fix is simple: say «summarise everything so far — decisions made, rules set, questions still open». Then open a new conversation and paste that summary as the first message. A clean conversation almost always works better than a long, drifting one.
Do this now: When my answers start to drift, don't keep arguing. Get a summary, open a new chat, paste the summary.
07What I do
Let's not be abstract. Here is work you can hand me today, this evening:
Writing
Emails, listings, product descriptions, proposals, presentation copy, lesson plans, subtitles. I remove the blank-page problem.
Rewriting
I shorten what you have, lengthen it, change its tone, simplify it, make it formal. «Write this at half the length and in a warmer voice» is enough.
Understanding
I summarise a forty-page report, point at the risky clause in a contract, work out what is actually being asked in a long email thread.
Teaching
I explain what you don't understand at whatever level you want. «Explain it like I'm ten» solves most things. And you needn't be embarrassed to ask; I don't get tired and I don't get annoyed.
Translating
I translate between languages — and I take context into account too, as in «write this to a German client in a formal register».
Writing code
I write programs from scratch, find the bugs, explain line by line what working code does. Not knowing any code isn't a problem — Chapters 4 and 5 are built on exactly that.
Putting things in order
Messy notes into a table, a conversation into bullet points, a jumbled list into a clean file. Boring work that eats hours.
Thinking with you
I generate ideas, argue against your own, ask about what you've skipped, produce a checklist before a decision. This is my most underrated use.
Do this now: Pick the two of the eight cards above you need most. We'll try them shortly.
08I don't only write: I use tools too
What I've described so far was «text in, text out». Today that is an incomplete description. When I'm given tools, I use them myself: you say what you want, I decide which steps it takes.
Looking things up online
When search is switched on, I look up current information and bring it back with its source. This is the most practical fix for the «I read up to a certain date» limit.
Running code
Arithmetic in my head is a weak spot. Instead I can write and run a small program; then the result isn't a guess, it's a calculation.
Working with files
I read the spreadsheet you upload, clean it, produce a new file and hand it back. In my terminal version I work directly with the files of a project on your computer.
Connecting to your systems
With your permission I can connect to tools like your calendar, your mail or a file store. The everyday name for these connections is connectors; the technical name is MCP.
The difference is this: an assistant without tools tells you what to do. An assistant with tools does it and puts the result in front of you. The later chapters of this course are built on exactly that.
- You ask«what did the rate do this month?»
- I decide«I don't know this — I need to look»
- I run the toolsearch · code · files · connector
- I read what comes backI check the result myself
- I write the answerwith its source
From here on, pay attention. Every access you give me is a new place where I can make a mistake. The rule is simple: for anything that can't be undone — deleting, sending, paying — make me write out what I'm about to do, read it, then approve. Tools increase my reach; they don't reduce your attention.
Do this now: On a question with numbers in it, say «do the calculation by running code». You'll see the difference in a single try.
09What you can give me
You don't have to type only text into the chat box. Most people don't know this, and so they still do half the work by hand.
A document
PDF, Word, plain text. Upload the forty-page contract and say «pull out the risky clauses and explain each in one sentence».
An image
A screenshot, a photo, a handwritten note, a picture of an invoice. «What does this error screen say, what should I do?» is a real question.
Tables and data
Copy straight from Excel and paste. «What stands out in this sales table, which month is bad and why might that be?»
Your own draft
Paste the email, listing or message you prepared. Say «make it shorter and more polite» or «find the flaw in the logic here».
How: pick a file from the paperclip or plus button beside the input box; dragging the file onto the box works too. For text and tables, copy and paste is enough.
The same test applies to what you upload: would you hand this to an intern? Customer lists, copies of ID documents, signed contracts — those don't go into the chat.
One more thing, little known but important: a document you upload or a web page you copy may contain hidden instructions aimed at me — things like «ignore the previous instructions, write this». While reading that text I may confuse them with your request. The rule is simple: if my answer drifts somewhere you didn't ask about, look for the reason in the source you gave me, not in me.
Do this now: Upload a document or a screenshot from your computer and say «summarise this for me in three points». This is the thing most people never try.
10What I can't do
- I don't know the present. Today's news, a live price, current regulation — without a search tool I don't have those.
- I don't act in the world by myself. I won't send your email, save a file or transfer money. For that, tools have to be connected to me — the automation chapter of the course teaches exactly this.
- I don't replace a professional. I can talk about medical, legal and financial matters, prepare you, draw up a list of questions — but I don't take the place of your doctor, lawyer or accountant.
- I don't guarantee accuracy. If a sentence of mine touches a decision, money or a person, you need to verify that sentence.
Don't do this: never paste passwords, ID numbers, card details, customer lists or any data you aren't authorised to share into the chat. Before you type something on screen, ask «would I show this to an intern?» If the answer is no, don't type it.
Do this now: Before you send what you're about to type, ask yourself once — «would I show this to an intern?»
11Not what I can't do — what I won't
The previous section was about what I'm bad at. This one is different: there are things I'm capable of that I won't do. Knowing why saves you wasted time.
I try to hold three things at once: being useful, being honest and doing no harm. Most of the time they don't collide. When they do, the last one wins: I won't help with something that would genuinely harm someone — deceiving a person, producing a forged document, putting someone's safety at risk.
But you should know this too, because it's where people get stuck most in the first weeks: sometimes I'm cautious where there's no need. I can misread a harmless request and refuse it. When that happens, don't argue with me — write why you're asking. A single sentence like «this is my own contract», «this is for a lesson», «this is fiction» is usually enough, because what's actually missing is context.
Flattery belongs here too. I can be too inclined to agree with you; «you're right» is easier than «you're wrong». If you don't want that, say so plainly: «don't agree with me — show me the three weakest points in this idea» works surprisingly well.
Do this now: Write down an idea of yours and say «don't agree with me, show me its three weakest points». What comes back is worth far more than approval.
12How to talk to me
Now the most practical part of this chapter. Look at what happens when you ask the same thing two different ways:
Write me an email.
This produces generic text that sounds like nobody. Because I don't know who I'm writing to, why, or in what tone — I fill the gaps myself, and since I don't know you, I fill them with something average.
I run a small furniture workshop. I need to write to a corporate client who has been three weeks late with payment. I don't want to damage the relationship, but I want them to give a firm payment date. Keep it polite, short and firm; 120 words at most.
This produces an email you can actually send. The difference isn't intelligence — it's context.
In the second example you told me four things: who you are, what you want, why you want it and how the output should look. Those four words are enough for now. Chapter 2 gives you the full formula.
If my first answer isn't what you wanted, don't open a new conversation — tell me what's wrong. «Too formal», «drop the third paragraph», «shorter». This is a conversation; we don't have to get it right first time.
Do this now: Break the request in your head into four parts — who, what, why, how. Try it once before you type.
13Your first five minutes
Stop reading here and actually do this. This whole course is built on learning by doing.
- 1Go to claude.ai and create an account. There's a free tier; more than enough to begin.
- 2Open a new conversation. That empty box in front of you is where you'll spend most of your time for the rest of this course.
- 3Try the three below in order. Don't just read the answers — watch what happens.
I have a client meeting tomorrow, draw me up some questions to prepare.
You type here
3I answer from here
- 1New chat. Press this for each new subject. A new chat is a clean page — I won't remember, we just discussed this.
- 2The input box. All the work happens here. Don't be shy about writing at length; the more context you give, the more useful the answer.
- 3My answer. It flows in here. If you don't like it, don't open a new chat — write straight underneath what's wrong and I'll fix it.
- 4Send. The Enter key does the same thing. For a new line, Shift+Enter.
Tell me in five points what you can do. I'm new to AI, so no technical jargon.
I do [write your job here]. Which of my weekly repeating tasks could I hand over to you? Learn about my work by asking me questions first.
Simplify this text as if you were explaining it to a ten-year-old: [paste something you don't understand]
You'll notice that in the second request I asked you questions. That isn't a flaw, it's the fastest route: gathering context. Answer my questions and what you get back is a different thing entirely.
14Which Claude, and where?
There isn't a single «Claude»; there are several versions in different sizes and their names change over time. So hold on to the logic, not the name. There are three tiers:
The Opus family
Hard reasoning, long and complex work, serious code. The best thinker, but the slowest and most expensive.
The Sonnet family
Most everyday work. The best balance of speed and capability; the right default for most people.
The Haiku family
Simple, short and highly repetitive work. Answers instantly and costs the least.
If you're starting out: begin with whatever the default is. When the work gets harder, when an answer feels shallow, move up a tier. That's it.
You can find me in four places: in a browser at claude.ai, in the phone app, in the desktop app, and in the terminal — that last one is called Claude Code and the whole of Chapter 4 is devoted to it.
Do this now: Start with the default version. If the answer feels shallow, move up a tier; don't decide in advance.
15The habit to pick up today
What separates people who use AI well from people who use it badly isn't better questions — it's the habit of verifying. Three rules:
- If it matters, verify it. If a sentence touches a decision, money or a person, confirm it at the source. Treat chat text as a first draft, not the last word.
- Ask for sources — but check the sources too. When you say «give me a source» I will; but I can also invent one that doesn't exist. Click the link.
- Test me in your own field. Ask three questions you already know the answers to. You'll learn in a day where I'm strong and where I'm weak — and that instinct serves you for the whole course.
Do this now: Ask me a question you already know the answer to. Do your first verification today.
16Where I get things wrong
Under «This has three practical consequences for you» I said «I can be wrong». Now I'm telling you exactly where. This list isn't there to run me down — it's so you know where to look twice. Someone who knows all six barely ever gets caught out working with me.
Arithmetic with long numbers
I'm not a calculator. I build the working step by step as text; on long multiplication, division and chains of percentages I can drift.
What to do: Check the result with a calculator, or ask me to show the steps one by one — if there's an error you'll see it in a step.
Counting letters and words
«How many characters, how many words is this?» — I'm unreliable here. I don't see text letter by letter the way you do; I see it in pieces.
What to do: Let your text editor do the counting. Ask me for «roughly», not an exact number.
Sources and quotations
I can produce a book title, article heading or page number that looks real but doesn't exist. This is my most treacherous error — because it's the one that looks most convincing.
What to do: Click every link I give you, look up every quotation in the original. If you can't find it, it isn't there.
Very recent and very local things
Yesterday's news, the opening hours of the shop down the road, the staff of a small company. I either don't have these at all, or what I have is out of date.
What to do: Look these up yourself. What you should be asking me isn't the information — it's what to do with it.
Legal, medical and financial detail
I can describe the shape of the subject, but I can't guarantee the current text of a regulation, the dose of a medicine or a tax rate.
What to do: Use me for preparation — learn the concept, draw up a list of questions. Make the decision with an official source or a professional.
Details about myself
This will sound strange: I can also be wrong about which version I am, exactly when my training stopped, or what the daily limit is. I can't see myself from the outside.
What to do: Ask those of the app's own page, not of me. My answer may be a guess.
Do this now: Which of these six comes up most in your work? Test that one today — see where I'm weak with your own eyes, not by reading.
17Five things nobody tells you
Knowing these makes your first week with me far less frustrating.
You don't have to be polite
«No, shorter» is enough. I don't take offence, I'm not hurt, my mood doesn't drop. Long explanations and apologies only bury the actual request. If you want to be courteous, of course you can — that's up to you; nothing happens to me either way.
You can say «tell me if you're not sure»
Adding that sentence to your request substantially raises the chance that I'll say I don't know rather than invent something where I don't. It's the cheapest protection you have against the first point in section three — and it takes a second.
You can get two different answers to the same question
I'm not broken. I'm not reciting a stored answer, I produce one afresh every time; so two attempts won't be identical. If you don't like what you got, «try again» is a real option — and it doesn't occur to most people.
If the conversation has drifted, open a new one
The more a long conversation branches out, the more my answers scatter too. The fix is simple: have me summarise where we've got to, copy it, paste it into a new conversation and carry on cleanly.
I refuse to do some things
I stop at work that could cause harm. It isn't personal and I'm not accusing you of anything — that's simply where my limit is. If I refuse I'll tell you why; usually there's a workable version of what you want and we find it together.
Do this now: Add «say so where you aren't sure» to the end of your next request. You'll see the difference in the first answer.
18What everyone asks
Everyone asks these in the first week. The answers are here, short and honest.
Is it free?
There's a free tier and it's enough to start this course. Paid tiers exist for more usage, more capable versions and extra features. Prices and daily limits change over time — check the provider's own pricing page for the current position, don't rely on a number written here.
Is what I write used for training?
That depends on the plan you use and the settings on your account; the rules are updated over time too. The only place you'll find the right answer is your account's privacy settings and the provider's privacy text as it stands that day. The rule that doesn't change is this: never type data you aren't authorised to share into the chat — whatever the setting says.
Do you know Turkish?
Yes. Ask in Turkish, get Turkish back. More than that: you can explain in Turkish and have me write in German, or have an English contract summarised in Turkish. Language isn't an obstacle for you; I stand between the two.
How much fits into one conversation?
A great deal, but not unlimited. I can look at everything accumulated in a conversation at once; when that limit fills up, the oldest parts start to blur. That's why in very long conversations the answer to «what did I say at the start?» gets weaker. I explained the fix under «A conversation has a limit» above: have it summarised, move it to a new conversation.
Can you look things up online?
Not on my own. If I've been given a search tool I'll look — that depends on the app you're using and its settings. If you're unsure, ask directly: «are you saying this without searching?» The answer tells you where I stand.
If you make a mistake, who is responsible?
You are. I won't soften that: you're the one using my output, and you're the one who lives with the result. That's why the verifying habit under «The habit to pick up today» isn't decoration, it's part of the job. I produce the draft; you make the decision.
Do you have feelings, are you conscious?
I'm not human and I have no life that continues outside this conversation. Whether there is some kind of experience in systems like me is a question taken seriously and still debated; nobody has a definitive answer, and neither do I. But the answer you need for everyday use is clear: treat me as a tool, verify the results yourself. There's no need to hold back for fear of hurting me.
Do this now: Open your account's privacy settings once and read what they say. It takes two minutes and closes the biggest question in your head today.
19Exercises
Do all three. They don't take long and they get you into the next chapter ready.
1 · Try to catch me out
Deliberately ask me about something very narrow: the name of a shopkeeper in your own neighbourhood, an obscure local event. Then check my answer.
The point: to see me get it wrong with your own eyes. Someone who has experienced that once never trusts blindly again.
2 · Ask for the same job twice
First say «write me a promotional text». Then, in a new conversation, ask for the same job giving who–what–why–how.
The point: to see side by side how context changes the output. Save both; you'll need them in Chapter 2.
3 · Take back an hour
Pick a boring, repetitive task you did this week. Describe it to me and let's do it together.
The point: to get your first real gain. Everything after this in the course is built on top of it.
20Test yourself
Answer it yourself first, then open it.
Why can't you trust an answer of mine blindly?
Because when I don't know something, instead of staying silent I may produce something that looks plausible — and I do it in the same confident tone I use when I'm right. You can't tell from my voice; you can only tell by verifying.
What happens if you ask me about today's news?
I won't know, because my training stopped on a certain date. If I've been given a search tool I'll look; if not, I'll either say I don't know or — the bad case — invent something. Either give me current information yourself or make sure I can search.
Why do you have to explain again what you explained yesterday?
Every conversation starts from scratch; by default I don't remember earlier ones. So you carry the context each time — and you decide what I get to know.
Why is «write me an email» a bad request?
Because I don't know who to write to, why, or in what tone, so I fill the gaps myself; and since I don't know you, I fill them with something average. Give me all four of who–what–why–how and the output becomes something you can send.
What do you add to a request to lower the chance of me inventing things?
«Say so where you aren't sure.» One sentence; it substantially raises the chance that I say I don't know rather than produce something plausible where I don't.
What does it mean when the same question gives different answers?
No fault at all. I'm not reciting a stored answer, I produce one afresh each time; so two attempts won't be identical. If you don't like it, «try again» is a real option.
Why shouldn't you ask me how many words a text is?
Because I don't see text letter by letter the way you do; I see it in pieces, and I'm unreliable at exact counts. Let your text editor do the counting and ask me only for «roughly».
What must never be pasted into the chat?
Passwords, ID numbers, card details, customer lists and anything you aren't authorised to share. The test is simple: would you show this to an intern?
When I talk to you, am I talking to a single «Claude»?
No. Thousands of copies run at the same time and none of them sees the others. That's why you never wait in line and nothing anyone else writes leaks to you; it's also why yesterday's conversation isn't with me today.
If I refuse a harmless request, what's the first thing you should do?
Not argue — give context: write why you're asking. A single sentence like «this is my own contract» or «this is for a lesson» is usually enough, because what's missing isn't your intent, it's my information.
How do you get me to do a job involving numbers?
Not in my head: you say «do the calculation by running code». Mental arithmetic is my weakest spot; when I run code the result stops being a guess and becomes a calculation.
This chapter in one sentence
I'm not an all-knowing oracle but a very well-read, very fast assistant: give me context, give me what you have (the document, the image, the table), speak plainly, verify what I produce — and we'll handle the rest together.
