Your first Claude Code session on a website, step by step

Your first Claude Code session on a website should be small, slow and reversible. Get a copy of the site in git (the system that records every change to your files), install Claude Code or the Claude desktop app, open the copy, switch to a mode that asks before acting, and start by asking questions. Then run /init to create a project instructions file, make one small change on a new branch, read the diff (the list of lines removed and added), check the result in a browser and commit it. Don't push, don't deploy, and stop there.

Claude Code is Anthropic's AI coding tool: it reads your project, edits files and runs commands on your computer. A first session shows you what it does with those permissions, on a copy where a mistake costs nothing. Each step below, as of October 2026, has an example prompt and the check that proves it worked.

What you need before you start

Get these in place first, or ask your developer or host to. If the site has no version control and no staging copy, set those up before anything else.

Step 1: Install Claude Code or the desktop app

The desktop app (no terminal). Download the Claude desktop app for macOS or Windows (Linux is in beta), sign in, and open the Code tab. Nothing else needs installing. If clicking Code asks you to upgrade, your account isn't on a paid plan yet.

The terminal. The terminal is the text window where you type commands (Terminal on a Mac, PowerShell on Windows). Anthropic's recommended native installer is one line. On macOS, Linux or WSL (Windows Subsystem for Linux):

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

In Windows PowerShell:

irm https://claude.ai/install.ps1 | iex

The installer shows no progress while it downloads, so leave the window open. Install commands change, so copy them from Anthropic's quickstart, not from any article. Native installs update themselves.

Check that it worked: open a new terminal window and run claude --version. A working install prints a version number followed by "(Claude Code)". If claude isn't found, your computer doesn't yet know where the program was installed (its PATH setting), and the quickstart explains the fix.

Step 2: Open your site's folder and sign in

1. In the terminal, move into the project folder and start Claude Code: cd followed by the folder path, then claude. In the desktop app, choose Local, click Select folder and pick the project folder.
2. Sign in when asked. In the terminal, a browser window opens for you to log in with your Claude account. Your credentials are stored, so you only do this once.
3. The first time you open a folder, Claude Code asks whether you trust it before using the project's own settings. Accept only for a folder you know: a project's settings can include commands that run automatically.

Check that it worked: in the terminal, the prompt appears with the version, current model and working directory shown above it. Confirm the directory is your copy of the site, not some other folder.

One trap: if an Anthropic API key (a password for pay-per-use API access) is set on your computer and you approve it when Claude Code asks, Claude Code uses it instead of your subscription and bills per use. Type /status to see which account the session uses.

Step 3: Check the permission mode before you type anything

A permission mode decides what Claude can do without asking you. As of October 2026, Anthropic's docs say: "With Claude Code v2.1.283 or later, auto mode is the built-in starting permission mode for interactive terminal and VS Code sessions." In auto mode, a separate AI model checks actions in the background instead of asking you. Anthropic warns that it "reduces permission prompts but does not guarantee safety."

For a first session you want to see every action, so switch to Manual:

• Terminal: press Shift+Tab. From auto mode, the first press switches to Manual (the docs' name for the mode whose setting value is default). The status bar shows the current mode, and Manual reads "manual mode on".
• Desktop app: use the mode selector next to the send button and pick Manual. Anthropic's desktop docs don't say which mode a new desktop session starts in, so look before you type. In Manual, Claude proposes each change and your files aren't modified until you accept.

Don't turn on bypass permissions mode. Anthropic reserves it for isolated containers and virtual machines and says it "offers no protection against prompt injection or unintended actions." Prompt injection means hostile instructions hidden in content Claude reads. Every mode is explained in our guide to Claude Code permission modes.

Check that it worked: the mode indicator says Manual before you send your first message. Plan mode, which proposes without editing your files, is also fine.

Step 4: Ask questions before asking for changes

Questions don't change your files, and they show you how the site is built. Try these:

• What does this project do, and what is it built with?
• Where do page titles and meta descriptions come from on this site? Show me the files.
• What command builds or previews this site locally? Don't run anything yet.
• Which files would I need to change to edit the footer text?

In Manual mode, Claude asks before running any command beyond simple read-only ones. When a permission prompt appears, read what it wants to run. If you don't understand it, choose No and ask Claude to explain the command first.

What a person must review: check one or two answers against what you know. If Claude says the site runs on WordPress and you know it's Shopify, you're in the wrong folder.

Step 5: Run /init and edit the CLAUDE.md it writes

CLAUDE.md is a plain text file of project instructions that Claude Code reads at the start of every session. Type /init and Claude analyzes the codebase and drafts one with build commands, test instructions and conventions it discovers. If a CLAUDE.md already exists, /init suggests improvements instead of overwriting it.

The draft covers what Claude can find in the code. Add what it can't, such as:

## How we work
- Work on a branch. Never commit to main, never push, never deploy.
- Run the build before saying you're done, and show me the diff.
- Prices, phone numbers and addresses come only from docs/company-facts.md.

Two cautions. Never put passwords or API keys in CLAUDE.md: it usually gets committed to git. And Claude treats these files "as context, not enforced configuration," in Anthropic's words: a rule in CLAUDE.md steers Claude, while a permission rule in settings actually blocks an action. Our guide to writing a CLAUDE.md for a website goes deeper, and our copy-ready templates include a permissions file that blocks pushes and keeps environment files out of reach.

Check that it worked: open CLAUDE.md and read every line. Delete anything wrong.

Step 6: Make one small change on a new branch

Pick a change you can judge by eye, such as a typo or the copyright line in the footer. Avoid forms, checkout, tracking, redirects and payments for now.

1. Ask for a branch first: Create a new branch called first-session-footer and switch to it. A branch is a separate line of work you can throw away.
2. Describe the change precisely: In the site footer, change "Copyright 2025" to "Copyright 2026". Change nothing else. Tell me which file you changed.
3. In Manual mode, read each proposed edit before you accept it, then accept it in the permission prompt.

If Claude heads somewhere you didn't expect, press Esc to stop it. To undo its file edits, press Esc twice while the input box is empty, or run /rewind. These checkpoints are for quick undo inside a session; git is your permanent history.

Step 7: Review the change and check it in a browser

1. Run /diff in the terminal, or click the change counter (it looks like +1 -1) in the desktop app. The diff shows removed lines and added lines for each changed file.
2. Confirm only the file you expected changed, and only the line you expected.
3. Ask Claude to run the build or preview command it found in Step 4. Ask for evidence, not a summary: Run the build and paste the last lines of the output.
4. Open the page on your local preview or staging copy and look at it yourself.

What a person must review: the diff and the page. Claude can say "done" about work that isn't right. Our guide to reviewing a diff when you aren't a developer lists the red flags.

Step 8: Commit, then stop

1. Ask: Commit this change with a short, descriptive message. Don't push. A commit is a saved snapshot in git; a push would send it to the shared online copy.
2. Ask: Show me the last 3 commits. Your commit should be at the top.
3. Type /usage to see how much of your plan's limits the session used. On a subscription, read the usage bars, not the dollar figure, which is meant for API users.
4. Leave the session with /exit, or type /clear before starting an unrelated task.

The branch now holds one reviewed, committed change. Getting it live is a separate step: push the branch, open a pull request (a request to merge it, with the diff attached for review), have it reviewed, then deploy. Leave that to whoever owns deploys until you've done a few sessions.

What to leave out of a first session

One of the house rules we kept from rebuilding eseospace.com with Claude Code applies from day one: one task per session, with /clear between unrelated jobs. Anthropic notes that when a long conversation is compacted (summarized to free space), detailed instructions from early on may be lost.

Conclusion

A good first session ends with one small, reviewed change committed on a branch, a CLAUDE.md you've read line by line, and a clear picture of what Claude Code does with your permission. Repeat that loop on low-risk tasks before anything larger, and keep a person in charge of what goes live. If you'd like your team's first sessions set up with safe permissions, project instructions and a review process, our Claude Code consulting covers exactly that. More guides are in our Claude Code Guide.

Frequently asked questions