Your first conversation with a coding agent.

Step 1 — give it a workspace

Claude Code works inside whatever folder you start it from — that folder is its whole world, which is a feature: it can't wander off and touch the rest of your computer. So we make it a sandbox. In a fresh terminal:

Mac:

cd ~/Desktop
mkdir first-build
cd first-build

Windows (PowerShell):

cd ~\Desktop
mkdir first-build
cd first-build

Translation: go to the Desktop, make a folder called first-build, step inside it. Three commands, and they're 80% of the terminal you'll ever need: cd means "go to," mkdir means "make a folder."

Step 2 — wake it up and ask for something

claude

You're at the prompt. Now type your first real instruction. Use mine or freestyle:

Now watch what happens, because this is the difference between a chatbot and an agent: it doesn't print code at you and wish you luck. It says something like "I'll create index.html" and asks for permission to write the file. The whole exchange looks like this:

Simulated run — the wording on your screen will differ, the shape of the exchange won't.

Step 3 — the permission prompt

Claude Code asks before it creates, edits, or runs anything. Early on, read what it's proposing — not because danger lurks, but because those prompts are a free education in what's happening. Here's the anatomy of one:

● I'll create index.html with your headline
  and the date.                    ← what it wants to do

  Create file index.html?          ← the file it touches
  ❯ Yes                            ← your approval
    No, tell Claude what to do differently

Approve it. It writes the file and tells you it's done. (You can tell it to stop asking for routine stuff later. For now, the asking is useful.)

Step 4 — look at what you made

The file is real. It's sitting in the first-build folder on your Desktop right now. Open the folder, double-click index.html, and it opens in your browser like any web page — because it is one.

"I built this with AI." Big letters. Dark background. There's the moment.

Step 5 — change something (this is the actual skill)

Building isn't one perfect request — it's a loop. Back in the terminal, without doing anything special, just keep talking:

Approve, then refresh the browser tab (the page doesn't update itself — refresh re-reads the file). New colors:

Do it a couple more times with your own ideas: a different font, an emoji, a second line of text. Each round trip takes seconds.

That loop — describe → approve → look → refine — is the whole workflow. Burn this picture in; every lesson from here is this loop pointed at bigger targets:

What to notice before you move on

• You never wrote code. You read English, approved English, and judged results with your eyes.
• The files are just files. You can see them in your folder, copy them, back them up. Nothing lives in a cloud you can't reach.
• Vagueness showed up already. "Subtle gradient" — did it match what was in your head? If not, you just met the problem Lesson 4 solves.