How to talk to Claude Code so you get what you meant

Claude Code will do exactly what you ask, fast, and it will not stop to check whether you meant it. The skill of this course is not coding. It is asking well and checking honestly. This lesson teaches that on something you can throw away, so the mistakes are free.

You will build a single web page for a fictional business. It has nothing to do with the client tracker. That is the point.

Step 1: Make a scratch project

In a terminal:

cd ~/projects
mkdir scratch
cd scratch
claude

Say yes when it asks whether you trust the folder. Then press Shift + Tab until the status bar says manual mode on. You are in a brand-new empty folder with Claude Code running inside it. Anything that goes wrong here stays here.

Step 2: Ask badly, on purpose

Type this:

Make me a website.

Watch what happens. If it asks permission to create files, say yes; nothing here matters. It will probably ask you questions, or it will make several guesses and build something. Either way, notice: it does not know what business, what pages, what colors, what it is for. Everything it decided, it decided for you.

Type /clear to throw that conversation away. The files it made can stay; we are about to replace them.

Step 3: Ask well

Now the same request, done properly. Press Shift + Tab until the status bar says plan mode on. Then paste:

I want a single-page website for a fictional dog-walking business called Good Boy Walks in Austin, Texas.

Two files: index.html and style.css. No frameworks, no JavaScript.

Sections, in order:
1. A header with the business name and a one-line tagline.
2. Three services with a name, a one-sentence description, and a price: 30-minute walk $20, 60-minute walk $35, weekend sitting $60/day.
3. A short "about" paragraph. Make up a friendly owner named Maria.
4. A contact section with an email address and a phone number. Make them obviously fake.

Style: warm, simple, readable on a phone. One accent color, a soft green.

Before you write anything, tell me the plan in plain English: which files you will create and what will be in each one. Wait for me to say yes.

Read the plan it gives you. It should be two files and a short description of each. When the plan is ready, Claude shows a small menu asking how to proceed. Choose Yes, manually approve edits. Do not pick the option that mentions auto mode; you want to see each change. Then approve each file as it proposes it.

This menu comes up every time you approve a plan, in every lesson from here on. The answer is always the same: manually approve edits. It takes one extra keypress per file and it is how you keep the habit of looking.

When it says it is done, look at the page. Open Finder (Mac) or File Explorer (Windows), go to your home folder, then projects, then scratch, and double-click index.html. It opens in your browser. (Terminal alternative: in a second terminal window, cd ~/projects/scratch then open index.html on Mac or start index.html on Windows.)

Compare it to what you asked for, section by section. This is Habit 1, and you will do it after every single task in the course.

Step 4: Change one thing at a time

Now iterate. Each prompt asks for exactly one change:

The prices are hard to read. Put each service in its own card with the price large and bold.

Refresh the browser and look. Then:

Add a fourth service: "puppy visit, 20 minutes, $15". Match the existing style exactly.

Look. Then:

The accent color is too bright. Make it a muted sage green.

Look. Notice the rhythm: one change, one look. When people get into trouble with AI tools, it is almost always because they asked for five things at once, three came out wrong, and now nobody can tell which change caused which problem.

Step 5: Say no

Ask for something and then reject the plan:

Add a booking form to the contact section.

It will probably propose a form and maybe some JavaScript to handle it, or it will ask where the form should send its data. Whatever it proposes, say:

No. I said no JavaScript. Make it a plain form with name, email, and a message box, and a submit button that does nothing yet. We will wire it up later.

Two things to learn from this. First, it will not argue; it will just do the new thing. You are in charge. Second, “does nothing yet” is a normal thing to ask for. You do not have to know how the whole feature works to ask for the visible half of it.

Step 6: Ask why

The most underrated prompt in the course:

Explain the files you created, one at a time, as if I have never seen HTML. What does each part do?

Read it. You will not remember it all, and you do not need to. What you are building is the reflex to ask, so that when the app is real and something matters, you already know how to get a plain-English explanation of what is going on.

Try one more:

If I wanted to change the business name later, which exact lines would I change?

The six rules

You just practiced all of them.

1. Be specific. Name the thing, the count, the order, the style. Vague in, guesses out.
2. Plan first for anything bigger than a one-liner. Plan mode, read the plan, then approve it with manual edits.
3. One change per prompt. One change, one look.
4. Read what it says it will do, not the code. The description is written for you.
5. Say no. It is a tool, not a colleague with feelings.
6. Ask why. “Explain this like I have never coded” is the prompt to reach for.

A seventh, for later: if you have gone back and forth three times and it is still wrong, stop. Do not keep pushing. Either go back to the last working snapshot (Lesson 5) and ask differently, or ask a person. Three failed rounds is the signal.

Check your work

• The page opens in your browser and has four services, a contact form, and a sage-green accent.
• You used plan mode at least once and chose “manually approve edits” from the menu.
• You rejected one proposal and got what you actually wanted.
• You asked it to explain something and read the answer.

If something went wrong

It built something completely different from what I asked. Read your prompt back. Was there a number, an order, a style, a “do not”? If not, add them. Then /clear and try again with the better prompt. This is not a failure; this is the lesson.

It keeps asking me permission for every file. Good. That is the mode you want right now. It gets faster once you trust it, which is after Lesson 5.

It says it is done but the page looks unchanged. Refresh the browser. If it still looks the same, ask: “I refreshed and don’t see the change. Which file did you edit, and what did you change?”

It made a lot of files I did not ask for. Ask: “Why did you create these files? Delete anything not needed for a single-page site with one HTML and one CSS file.” And note it for next time: “two files” was in the prompt for a reason.

Next

Lesson 5 is Git: the undo button. It is the one thing separating people who recover from a bad AI change in a minute from people who lose a week.