Claude Code for client websites: what to decide before the first prompt
Can Claude Code build a website for a client? Yes, if you decide the brief, rules, review and ownership first. AI agents for client work start with this list.
A dentist's practice manager emails on a Monday: the new hygienist starts next month, the practice now offers teeth whitening, and could the website be "a bit more modern" while you're at it. You open a terminal, start Claude Code, and type "build a modern website for a family dentist". Forty minutes later there's a clean site with smiling stock families, a whitening page, and opening hours that belong to no dentist on earth.
The build was quick. What was missing happened before the first prompt, and it's the part AI agents for client work still leave to you.
Claude Code can build a website for a client. It writes the pages, the styles and the forms, and it changes them when you ask. What it can't do is decide what this particular practice needs, which facts are true, and what the client agreed to pay for. Those decisions are yours, and they're easiest to make once, in writing, before the agent starts.
Can Claude Code build a website for a client?
Yes. For a small business site (a few pages, a contact form, a booking link) an agent like Claude Code can do most of the building. The question is not whether it can write the code. It's whether the result says true things about the business, matches what was sold, and can be looked after later.
That depends on four decisions:
| Decision | The question | Where it lives |
|---|---|---|
| The brief | What pages, for whom, and what's not included? | Your proposal, copied into the project |
| The facts | What's true about this business? | A facts file the agent reads |
| The review | Who checks what, and when? | Your review checklist |
| Ownership | Whose accounts does the site live in? | The client's name, from day one |
Each one is ten minutes of work before the first prompt and hours saved afterwards.
What should the agent know before it writes anything?
Start with the brief. If you sent a proposal, you already have it: what the client told you, what you'll build, and what's not included. Copy it into the project folder as a plain text file.
Then write the facts, one per line, with where each came from:
- Business name, exactly as on the sign.
- Address, phone number and email.
- Opening hours, including any closed days.
- Services and prices, if the site shows prices.
- Staff names and roles, if there's a team page.
- Anything the client said must never appear (an old address, a former partner's name).
This is the part people skip, and it's why agent-built sites end up with plausible hours and invented testimonials. An agent fills gaps with likely text. Give it fewer gaps.
Claude Code reads a file called CLAUDE.md in the project at the start of a session. Anthropic's documentation describes it as the place for instructions and context you want Claude to have every time. For client work, that's where the brief, the facts and the rules go, or where you point to the files that hold them. Your rules might read:
- Use only the facts in facts.md. If something is missing, ask; don't invent it.
- No reviews, testimonials, awards or numbers unless they're in facts.md.
- Plain words a patient would use. No jargon.
- Keep the pages agreed in brief.md. Ask before adding a page.
Skills and plugins can add more (a house style, a set of checks you run on every site); Claude Code skills covers what those are. But a short CLAUDE.md with the brief, the facts and four rules does more for a client site than any add-on.
How should you work with the agent during the build?
Ask for a plan before it builds. Claude Code has a plan mode, where it reads the files and proposes a plan but makes no edits until you approve it. For a client site, the plan is where you catch a page that wasn't in the brief, or a booking system nobody paid for, before it exists.
Then build in small steps. One page or one change per request is easier to check than "build the whole site". After each step, look at the result in a browser. Not the code: the page, the way the client will see it.
Keep the earlier versions. Claude Code saves checkpoints before each change, and Anthropic's September 2025 announcement describes rewinding to an earlier state by pressing Escape twice or using the rewind command. That's useful inside a session. For client work you also want versions that outlast the session, with a note of what changed and why, because the client will ask "what did the home page look like last week?" and "why did the booking button move?"
If you connect the agent to other tools, such as a client's files or your email, what is MCP and what it changes for agency work explains those connections and what to check before you turn them on.
What should you check before the client sees it?
Everything the agent wrote about the business. It is the step where your name is on the line.
- Every fact. Name, address, phone, email, hours, prices, staff. Against facts.md, one by one.
- Every claim. Anything that sounds like a promise ("same-day appointments", "pain-free") must be something the client said and can keep.
- Every link and form. Click them. Send a test message through the contact form and check it arrives.
- On a phone. Most people will see the site on a phone first. Open it on one.
- Against the brief. Every agreed page is there; nothing unagreed has crept in.
Write those five into a checklist you tick every time, so the review doesn't depend on how tired you are.
Who owns the site the agent built?
The client, and that's easier to arrange at the start than at the end. The domain, the hosting account and any booking or payment accounts should be in the client's name, with you added as a helper. The code and content the agent wrote should be handed over as ordinary files the client could give to another designer.
Decide this before the build, because it changes where you deploy and which accounts the agent's work goes into. A site built into your personal hosting account is a site the client has to ask you for later.
What does the agent do, and what do you do?
A fair split for a small business site:
| The agent | You |
|---|---|
| Writes the pages from the brief and facts | Agrees the brief and facts with the client |
| Proposes a plan | Approves the plan |
| Builds and changes pages | Checks each change in a browser |
| Suggests wording | Checks every fact and claim |
| Fixes what you point out | Chooses the version that goes live |
| Hands over accounts and files |
The right-hand column is the client work. It's also the part the client is paying you for, whoever writes the code.
Where does Volant fit?
Most of what you decide before the first prompt (the brief, the facts, the versions, what the client approved) needs a home that outlasts one chat with your agent. Building, reviewing and delivering client websites is how Volant handles that: your agent builds the website, and Volant has version history for every website, so you can compare versions and go back to any earlier one. You choose which one goes live. If you work as a developer and already use Claude Code, building client sites with an AI agent shows how it fits your workflow, and approving what your agent sends covers the approval step.
Whatever you use, make the four decisions first: the brief, the facts, the review and who owns what. The build goes better when the agent starts from them.