Module 4 · Build your site
Watch your website get built
This is Stage 3, and it is the one people remember. Your two planning documents go in, and a real website comes out — every page you planned, in your colors, running on your own computer where nobody else can see it yet. You will spend most of this lesson looking at it and asking for changes in plain English.
Two minutes of pre-flight
Before you paste anything, check three things. This takes longer to read than to do, and it saves the only kind of afternoon worth saving.
site-architecture.mdexists in your project folder and you have actually read it. It is the blueprint; if a page is missing from it, that page will be missing from your site.brand-strategy.mdexists, and you were happy with the wordmark logo it produced.- You have a snapshot saved. If you are not sure, say “save a snapshot before we start” and Claude will take one.
One thing your architecture document should already contain, from the planning module: a primary and a secondary call to action on every page, and empty placeholder pages for the intake form, the thank-you page, and the disclaimer. The build fills those shells in. Module 5 wires them to your inbox and your calendar. If they are not in the document, add them now with one sentence to Claude — retrofitting a contact path onto a finished layout is the single most annoying job in this course.
Start in Plan mode — the “show me first” button
Plan mode is a setting that tells Claude to think out loud and change nothing. It reads your documents, writes out exactly how it intends to build the site, and waits. You read the plan, ask for changes, and only then say go.
You are a person who reads a document before signing it. This is the same instinct, and it is the most useful setting in the whole application for a cautious beginner. Switch to it with the mode selector next to the send button, or press Shift+Tab until the mode reads Plan. When the plan is ready you will be offered three choices: approve it and let Claude work on its own, approve it but review each edit, or keep planning.
The build prompt
Replace nothing here except what is in square brackets — the kit’s skills carry the detail, so this prompt stays short on purpose.
Build my firm's website
Let's do Stage 3 — build my website. Follow your frontend-design and copywriting skills, and build from my site-architecture.md and brand-strategy.md. 1. Set up a new Next.js + Tailwind website in this project. 2. Build every page described in site-architecture.md — the sections, the copy and the design notes — using my brand colors, fonts and logo. 3. Put the primary and secondary calls to action on every page exactly where the architecture document says they go, and make the phone number a real tel: link everywhere it appears. 4. Create the empty shells for /contact, /thank-you and /disclaimer, /privacy. Leave them plain for now — we wire them up in the next module. 5. Build it phone-first. Most people reading a law firm's website are holding a phone, often at a bad moment. It has to be excellent there. 6. Start the local preview and give me the link. 7. Save a snapshot when the first version runs. Explain what you're doing in plain terms as you go. Use no photographs of people anywhere — leave a clearly-marked placeholder where a photo belongs, and I'll add real ones in the next lesson.
Left to itself, an AI will happily fill your bio page with a confident-looking stranger in a suit. For a firm that is not a taste problem, it is a misrepresentation problem — the next lesson is entirely about that line. Blocking it in the build prompt means you never have to find it later.
Now go and look at it
Claude will hand you a localhost link. That is a private address that only works on your own computer — your website is running there, and nothing is on the internet yet. In the Claude app the site also opens in a preview pane right beside the chat, which matters more than it sounds: Claude can look at the running page itself. You can say “the spacing under the headline is wrong” and it goes and looks rather than guessing.
That loop is the rest of the module. Real things to say:
- “The headline is too long. Cut it to eight words and make it bigger.”
- “The phone number should be the first thing I see on a phone, not the third.”
- “This practice-area page reads like a brochure. Make the first paragraph speak to someone who is frightened.”
- “Add more space between the sections — it feels cramped.”
- “Here’s a screenshot from my phone. The button is under the keyboard.” (You can drag a screenshot straight into the chat.)
Check every page at phone width before you move on. In the preview pane you can drag the edge in until it is phone-shaped; better still, open the localhost link on your actual phone if it is on the same wi-fi. Ask Claude how, in those words, and it will tell you.
The undo button you did not know you had
Claude quietly saves a copy of your files before every message you send. So when a change makes things worse, you do not have to describe your way back out of it.
Type this in the chat box
/rewind
You get a list of everything you have asked in this session, and the option to put your files back exactly as they were before any one of them. You can roll back the files, the conversation, or both. Pressing Esc twice on an empty message box does the same thing.
It cannot undo something a command did on its own — a file that was deleted or moved by a command stays deleted or moved. And it is not a replacement for snapshots, which are your real backup and the thing that gets copied off this laptop in Module 7. Use /rewind for “that last change made it worse,” and keep saying “save a snapshot” at the good moments.
Next to the send button is a model name — that is which version of Claude is doing the work. If an identical problem returns a second time, click it, choose a stronger model, and say “try that again.” Stronger models use more of your monthly allowance, so keep them for the genuinely hard parts. The default is right for almost everything here.
A long build session on the Pro plan can hit a limit message. Nothing is lost — your work lives in your files and snapshots, not in the conversation. The main limit resets every five hours, so make a coffee, come back, and open with “where did we leave off?”
There is also a separate weekly limit that a short break will not clear. If you reach that one, spread the rest of the build across the following days. It is the reason this module is written as two sittings rather than one marathon.
Every page named in site-architecture.md loads on your localhost link, in your colors, with your wordmark at the top. The phone number is tappable. /contact, /thank-you and /disclaimer exist, even if they are nearly empty. Every page still looks right when you squeeze the preview down to phone width, and there is not one photograph of a person anywhere yet.
An error on the screen is normal. Copy it, paste it back, and say “this didn’t work — please fix it and tell me what happened in one sentence.” Errors during a build are routine and Claude clears them quickly.
Your change did not appear? Say “I don’t see the change — is the preview still running? Please restart it and give me the link again.”
Claude went down a rabbit hole? Say “stop — let’s try a simpler approach,” then use /rewind to get back to before the detour started.
A file called AGENTS.md appeared. That is a note the website framework writes for AI tools. Leave it alone; your CLAUDE.md from the Build Kit is still the one in charge.
Watch this step
Here is the build stage on screen — the same describe-it-then-refine-it rhythm you're about to run, with a builder deliberately steering away from the look that reads as machine-made. Your kit already enforces most of what he does by hand. It’s from an independent creator — credited below, so go give them a follow.
How to Build Beautiful Websites with Claude (That Don't Look AI-Generated)
More learning resources
Still want another angle? These are hand-picked and vetted to line up with this lesson — from creators and sites worth your time.
