Turn on Linux (WSL)

Underneath the windows and buttons, a computer works with text. A terminal is a window where you talk to it in words. We’ll use Linux, running inside Windows (that’s WSL, Windows Subsystem for Linux).

1. Click Start and type PowerShell.
2. Right-click Windows PowerShell, choose Run as administrator, and click Yes.
3. Paste this and press Enter:

wsl --install

When it says so, restart the computer. A window called Ubuntu (our kind of Linux) then opens and asks you to make up a username (short, lowercase, no spaces) and a password.

From now on, open Terminal from the Start menu and pick Ubuntu from the ▾ menu next to the tabs.

Look around

The line ending in $ is the prompt: the computer waiting for you.

pwd

pwd prints where you are, something like /home/anna: your home folder, written as a path (folders inside folders, separated by /).

mkdir course
cd course
pwd

Predict before the last line: what will pwd print now? (mkdir makes a folder, cd goes into one, cd .. goes up a level.)

Stop the password questions

Commands starting with sudo run as the administrator, and Linux asks for your password each time. The Windows login already protects this computer, so we tell Linux to stop asking:

echo "$USER ALL=(ALL) NOPASSWD:ALL" | sudo tee /etc/sudoers.d/$USER
sudo chmod 0440 /etc/sudoers.d/$USER

Your tutor: Claude Code

Claude is an AI, a language model: it learned from a huge amount of text to predict what words come next. That turns out to be enough to explain things and write code. It also makes Claude very good at sounding right. Usually it is right. Sometimes it isn’t, and it can’t always tell. That’s why you check.

Claude Code lets you talk to Claude in the terminal. You’ll need a Claude account.

curl -fsSL https://claude.ai/install.sh | bash

curl downloads the installer, and | bash runs it (the |, a pipe, sends one command’s output into the next).

Claude could do this whole lesson by itself. That’s why we tell it not to. This box writes a file called CLAUDE.md: standing instructions Claude reads every time it starts in this folder.

cat > ~/course/CLAUDE.md <<'EOF'
# Tutor mode

I'm a beginner learning to program, and above all to reason.
The point is that I work things out myself.
You're my tutor, not the one doing the work.

- Never write the code for my exercises,
  and don't run commands or edit files for me.
  If you solve it, the lesson is lost.
- If I paste an error message, explain in plain words
  what it means and where to look. One hint, not the fix.
- If I ask what a command does, explain it simply,
  then ask me to predict what will happen before I run it.
- Ask me "why?" often, and look for the hole
  in my explanations.
- If I ask how you work, or why AI gets things wrong,
  explain honestly and simply. That's part of the course.
- Allowed: if I ask for code *to referee*
  (with a hidden bug for me to find), write it.
- Keep answers short. Use everyday words.
EOF

Press Alt+Shift+D to split the window into two panes. In the right pane, start Claude:

cd ~/course
claude

The first time, it opens a browser to log in. Claude asks before it does anything on your computer: say no unless you understand what it’s asking.

Try two questions: “What did the cat > CLAUDE.md command I just ran do?” and “When are you most likely to be wrong?”

From now on: left pane, you work. Right pane, you ask.

A friendlier shell: fish

The program that reads what you type is the shell. Ubuntu’s is bash. We’ll use fish, which helps you type less. Ask Claude what chsh does, and predict what will change, and when:

sudo apt update && sudo apt install -y fish
sudo chsh -s /usr/bin/fish $USER

(apt installs software.) Close the left pane (exit) and open a new one with Alt+Shift+D. Was your prediction right?

Fish suggests the rest of a command in grey from what you typed before (→ accepts it), and Tab finishes names. ↑ brings back earlier commands.

Fish doesn’t know yet where Claude was installed. Tell it once:

fish_add_path ~/.local/bin

Explain (ask Claude if you’re stuck): what is a “path”, and why did bash find Claude but fish didn’t?

An editor: nvim

To write programs we need a text editor. nvim runs inside the terminal.

sudo apt install -y neovim
cd ~/course
nvim hello.txt

nvim starts in normal mode, where keys are commands, not letters. Four things are enough:

Press i, type hello, press Esc, type :wq, press Enter. Predict: what will this print?

cat hello.txt

Stuck in nvim? Press Esc a few times, then :q!. (More nvim: type :Tutor inside it.)

Python

python3

The prompt changes to >>>: you’re talking to Python now. Predict each answer, then check:

2 + 2
7 / 2
7 // 2
2 ** 10

Ask Claude why 7 / 2 and 7 // 2 differ, then explain it back. Leave Python with Ctrl+D.

A good place to stop for the day.

Boxes and questions

Start Python again and predict each line before pressing Enter:

python3

x = 5
x + 1
x
x = x + 1
x
x == 6
x == 7

x = 5 makes a box called x and puts 5 in it. x + 1 works out an answer but leaves the box alone.

In maths, x = x + 1 is nonsense. In Python, = means “work out the right side, then put it in the box on the left”.

== asks whether two things are equal, and the answer is True or False. One = stores, two == ask. Mixing them up is one of the most common mistakes in programming.

Leave Python with Ctrl+D.

Decisions: if

cd ~/course
nvim decide.py

Type this in (i first, Esc and :wq at the end):

n = 7
if n % 2 == 0:
    print(n, "is even")
else:
    print(n, "is odd")

% is the remainder: 7 % 2 is 1. if runs the indented lines under it when its question is True, otherwise the lines under else:. Predict, then run:

python3 decide.py

Indentation: what’s inside what

A line ending in : starts a block: the lines under it, indented 4 spaces further. The block ends at the first line that steps back out. nvim indents for you after a :, but it can’t know when a block ends. That’s you, pressing Backspace, and it’s a reasoning question: does this line belong inside?

Repeats: for

nvim loop.py

for k in range(1, 6):
    print(k, k * k)

for runs the indented line once for each k in range(1, 6): 1, 2, 3, 4, 5. It stops before 6. Predict all five lines, then run:

python3 loop.py

Predict exactly what this prints:

for k in range(3):
    if k % 2 == 0:
        print(k, "even")
    else:
        print(k, "odd")
    print("done")

How many times does done appear? What changes if print("done") moves one step left, under for? Try both and check.

The story

Around 1785, the story goes, a teacher told his class to add up every number from 1 to 100. A ten-year-old, Carl Friedrich Gauss, wrote the answer almost at once: 5050.

How might he have done it? Spend two minutes with a pencil before reading on.

The slow way: a loop

cd ~/course
nvim gauss.py

Type it in yourself (typing teaches more than pasting):

total = 0
for k in range(1, 101):
    total = total + k
    print(k, total)

Predict the first three lines, then run it:

python3 gauss.py

The last line says 100 5050.

Now make an error on purpose: delete the spaces in front of total = total + k and run it. Read the error, guess what it means, then paste it to Claude with your guess.

Gauss’s trick

Write the sum forwards and backwards:

  1 +  2 +  3 + ... +  99 + 100
100 + 99 + 98 + ... +   2 +   1

Each column adds up to 101, and there are 100 columns. Two copies of the sum make 100 × 101 = 10100, so one copy is 5050. For any :

Explain it to Claude in your own words, without the picture, and let it look for the hole. (Does your explanation work when is odd?)

A check inside the loop

nvim invariant.py

total = 0
for k in range(1, 101):
    total = total + k
    assert 2 * total == k * (k + 1)
print(total)

assert means “this must be true here”: if it’s false, Python stops with an AssertionError. So this loop checks Gauss’s formula after every step. Predict, then run it:

python3 invariant.py

It prints 5050: no step failed.

Now break the loop on purpose: change total + k to total + k + 1. Predict: at which k does the assert stop it? Then change it back.

Check lots of cases

invariant.py only checks up to 100. To check other n, we turn the adding-up into a function, made with def: a recipe with a name. [claude: give more defintion of funcitons and def, its an import basic, and then also introduce what assert does with perhaps a simpler example to try like assert 4 % 2 == 0 verssu assert 4 % 2 == 1, then after the check, perhaps another one to do the n^3 example below]

n is its input, and return hands back the answer, so sum_to(100) is 5050.

nvim check.py

def sum_to(n):
    total = 0
    for k in range(1, n + 1):
        total = total + k
    return total

for n in range(1000):
    assert sum_to(n) == n * (n + 1) // 2
print("all 1000 agree")

The first part is the loop from gauss.py, indented under def, with n + 1 where 101 was. The second part asks, for every n from 0 to 999, whether the loop and the formula agree. Predict, then run it:

python3 check.py

It prints all 1000 agree. Is the formula now true for every number? Decide before reading on.

No. A thousand examples is still just examples. Mathematicians have found patterns that held for millions of cases and then broke.

Why it can never fail

A claim that stays true after every round is a loop invariant, like the assert in invariant.py. Now the big question: if the assert passed in one round, why must it pass in the next? (What does one round add to total?) Decide before reading on.

Say it passed in the round for :

The next round, for , adds , so grows by :

which is exactly what the assert checks. It passes in the first round, and each passing round makes the next one pass, so it can never fail, however far the loop runs.

That’s no longer a check. It’s a proof for every , and it’s the method from maths class: induction. The first round is the first domino, and each round knocks over the next.

Write this proof on paper, in your own words. You’ll need it in Part 4.

Why AI and Lean belong together

A proof that sounds right can still be wrong, and Claude is very good at sounding right. Lean doesn’t care how convincing a proof sounds: it checks every step, or it refuses. So the AI suggests, and Lean checks.

That leaves you the most important job: does the claim say what you meant? Lean checks that the proof proves the claim, not that it’s the claim you wanted.

What Lean is

Lean is another programming language. It has boxes and loops, like Python, and it can also check proofs. Seeing the same program in a second language shows which parts are the ideas (a box, a repeat) and which are just spelling.

We’ll use Lean in the browser, with Claude beside it. Claude in Chrome is Claude in a side panel of Chrome that can read the page you’re on and type into it. Install the Claude extension from the Chrome Web Store, sign in, and pin it (the puzzle-piece icon, then the pin).

Open live.lean-lang.org, delete what’s there, and paste:

def sumLoop (n : Nat) : Nat := Id.run do
  let mut total := 0
  for k in [1:n+1] do
    total := total + k
  return total

#eval sumLoop 100

The panel on the right shows 5050. Here it is next to sum_to from check.py:

• Nat is a natural number (0, 1, 2, …). (n : Nat) : Nat says n is one, and so is the answer.
• let mut makes a box you can change (mutable).
• := puts something into a box, like Python’s =.
• Id.run do says “steps, done in order”. Just copy it.
• #eval runs a line and shows the answer.

Explain: what does Lean make you say that Python doesn’t?

Two programs, one answer

Gauss’s formula is a program too, a much faster one. Add it below:

def sumFormula (n : Nat) : Nat := n * (n + 1) / 2

#eval sumFormula 100

/ on natural numbers rounds down, like Python’s //. (Why doesn’t that rounding ever change the answer here?)

Checking isn’t proving

Do the two programs agree? Add one check, for :

#eval sumLoop 1000 == sumFormula 1000

It shows true. Now add the claim for every :

theorem agree (n : Nat) : sumLoop n = sumFormula n := by
  sorry

Read it out loud: “a claim called agree: for any natural number , the loop and the formula give the same answer. Proof: sorry.” This is check.py again, for every instead of a thousand.

sorry means “I’ll prove this later”. Lean understands the claim but warns: declaration uses `sorry`. You still owe a proof, and checks like true don’t count. What Lean wants is an argument like your loop invariant.

AI translates, Lean checks

Writing Lean proofs yourself comes later in the course. Today Claude only translates your proof, and you watch what Lean says.

Open Claude in Chrome on the Lean page, type in your paper proof from Part 3, and ask:

Please prove agree in Lean by translating my proof, and see whether Lean accepts it. Don’t fix any mistakes in my proof, and don’t fix anything Lean complains about. Only translate my proof into Lean, and translate Lean’s answer back into English.

It may take a few minutes, and then Lean should accept it. Claude may add a smaller claim along the way (something like “one more round adds ”). **Where is that step in your paper proof?

Next, a claim of my own. Before you send it, predict: is it true? Try , 1 and 2 on paper. Then ask:

Now I want to check whether this also means that sum(n)^2 = n^3. Please convert that to Lean, have Lean check it, and translate only Lean’s answer back into English. Don’t fix the proof.

What did Lean say? And did Claude do only what you asked, or did it quietly change the claim to one that works?

What we did today

• Set up Linux with Claude as a tutor, plus fish, nvim and Python.
• Added 1 to 100 with a loop and with Gauss’s trick.
• Checked a thousand cases, and saw why that isn’t a proof.
• Put a check inside the loop, and saw why it can never fail: induction.
• Wrote the same loop in Lean, next to Gauss’s formula, and claimed the two programs always agree.
• Had Claude translate your proof for Lean to check.

The computer did the adding. You did the reasoning.