AI & MCP

Build it in Claude. Finish it in the builder.

AI is brilliant at the first 80% of a website and wasteful at "change this headline". Split the job so each half happens where it's cheap.

Build it in Claude. Finish it in the builder.
Published
November 6, 2026
Read time
11 min
Category
AI & MCP
Written by
BYQ Supply
2,900+ sections. Zero filler.

Copy-paste sections for Webflow, Figma, Framer, Shopify and your AI tools.

Explore library

Picture the moment. You've just watched Claude build a whole website. Pages, copy, a hero you'd have happily drawn yourself. It took a few minutes, and you didn't open a single file.

Then you scroll down and see it. The button says "Get started". It should say "Book a call".

Three words. And you're about to do the thing almost everyone does next: you type it into the chat.

The small fix that isn't small

Here's what happens when you ask an AI to change one button. It reads the page. It finds the button. It changes the button. And, because it's helpful, it notices the headline above it, decides that could be tighter, and nudges the spacing around it while it's in there.

You wanted one word swapped. You got three files touched, and now you're reading a diff to work out whether it broke anything.

Two panels. By prompt: Claude changes three files to rename one button. By hand: click the button, type, one change, no AI tokens.
The same three-word change, done two ways.

None of that is the model misbehaving. It's the model doing what it's good at, which is working across a whole site at once. That's a superpower for building and a liability for polishing.

Polishing is the opposite job. It's one thing, exactly, and nothing else. You can see the thing you want to change. You shouldn't have to describe it.

Two jobs, two places

So we stopped thinking of building a website with AI as one job.

The first job is making something out of nothing. Pick a direction, choose a template, lay out the pages, write a first draft of every word. It's wide, it's tedious, and a person would need days. This is where Claude is worth every minute.

The second job is making it yours. The headline that almost says what you mean. The photo that isn't quite right. The colour that's one shade off. It's narrow and fiddly, and it rewards someone who can see the page.

Claude is the right tool for the first. The BYQ Website builder is the right tool for the second. The trick is that they're looking at the same site.

The heavy work happens in a chat, where AI is good at it. The finishing happens on the page, where you are.

What it looks like in practice

Say you're building a site for Atelier Mar, a made-up architecture studio in Porto. You connect BYQ to Claude (here's the two-minute setup) and ask for the whole thing in one go.

Prompt
Build a site for Atelier Mar, an architecture studio in Porto, from a BYQ template. Rewrite it for us, keep it to five pages, and tell me when it's ready to open in the builder.

Then you watch. Claude searches the templates, picks one, creates the site, rewrites the copy, and checks every page in a real browser before it moves on. You can open the live editor and see it appear while it works.

Four steps: Claude builds, open it in the builder, fix by hand, publish.
Claude does the first step. The builder does the next two. Publishing works from either.

When it says it's done, you don't carry on in the chat. You open the site in the builder.

Now you're looking at the actual page, not a description of it. Click the headline and type. Click the photo and swap it. Open the Style panel and change a colour or a font, and the whole site follows.

This is the part that feels different. You're not negotiating with anything. You change a word and the word changes.

Three evenings where this matters

It's easy to nod at a workflow in the abstract. It's more useful to see where it lands in a real week. Here are three moments we think it earns its keep.

Three scenarios: a client on the phone at 4:40pm, one more tweak at 11pm, and a new teammate on Monday.
Three moments where the second half of the flow is the whole point.

The client on the phone. It's late afternoon and you've shared the link. They like it, mostly. Then: "Could that headline be a bit softer?" Without the builder, that's a promise to get back to them, a prompt, a wait, a check, and an email. With it, you click the headline, type a softer one, and read it out while they're still on the line. They hear the change happen. It's a surprisingly good moment.

The late-night tweak. You're nearly done, and you're also tired, which is when a one-word change quietly becomes a rewrite. This is the time to stop prompting. Open the page, move the photo two inches, swap the second sentence, close the laptop. Nothing gets regenerated because nothing was asked to.

The person who joins on Monday. A colleague, a client, someone who has never typed a prompt in their life and has no wish to. In a chat-only world, that person is locked out of their own website, because the only door is a text box. In this one, they can click a heading and change it. That alone is worth a lot.

A rule of thumb

If you can point at it, click it. If you can't point at it, ask for it.

"Change this button", "swap this photo", "make this heading shorter": all things you can point at. Do those by hand. "Add a testimonials section", "build a pricing page", "rework the whole structure of the home page": nothing to point at yet. That's a request, and it belongs in the chat.

Most of the time the line is obvious. When it isn't, we find that asking "could I do this in ten seconds with a mouse?" settles it.

Why undo is the quiet hero

There's one more thing that makes the second half feel safe, and it's the least glamorous. Every change in the builder is a version.

A version list: v1 Claude built the site, v2 headline rewritten, v3 hero photo swapped, v4 new accent colour in the Style panel.
Claude builds v1. Everything after it is a step you can walk back.

That changes how you edit. You stop being careful. You try the other photo, you try the other colour, and if you don't like it you go back. The site Claude built is always there as the first step in the list, so there's no version of this where you lose the good draft by fiddling with it.

The bit about money

This is the part we like most, because the workflow splits the cost the same way it splits the work.

Building from Claude runs on your own Claude plan. It uses none of your BYQ AI tokens. And everything you do by hand in the builder, the clicking, typing, photo swaps and styling, uses none either. The only time BYQ tokens come into it is if you ask the builder's own chat to do something for you.

Three rows: the big build runs on your Claude plan, hand edits use no BYQ tokens, the builder chat uses a little each.
What each part of the flow draws on.

The two parts that feel the most expensive, the enormous build and the endless little tweaks, don't touch your BYQ tokens at all.

What it's good at, and what it isn't

Worth being honest about the edges.

A whole site is a big job for Claude, so a long build counts against your usage there. On a smaller plan you'll notice it.

Publishing from Claude needs a Pro or Ultra plan, or hosting. Building and editing don't.

And it's a first draft, not a finished brand. Claude gets you a good site fast. The last stretch, the bit that makes it sound like you, is yours. That's why the builder is the second half of the flow and not an afterthought.

Try it tonight

If BYQ is already connected in Claude, you're one prompt away. Start with the build prompt above, or this one if you'd rather leave the words to yourself:

Prompt
Build a site for my studio from a BYQ template. Keep the copy simple and generic, because I'll rewrite the words myself in the builder. Tell me when it's ready to open.

When you're happy with it, you can go live without leaving the chat.

Prompt
Publish it to a byqsite.com link.

Then open it, click the headline, and fix it yourself.

See the Website builder →

Explore more articles

All articles