Text to website: how AI turns a prompt into a site

What is text to website?

Picture a pottery teacher with a Saturday beginners' class, a shared studio and no website. She can describe the class in two minutes: six weeks, what people will make, what to wear, how to sign up. Turning those two minutes into a web page used to be the hard part.

Supacode, an AI app builder by Transcend Technologies, does this with real code you can download. Type a description into the composer and it writes HTML, React or Next.js files, runs them in a live preview and saves the result as a version. You then refine it in the same chat: “make the dates bigger”, “add a FAQ”. People also call this prompt to website or, more broadly, vibe coding.

That's the promise. The useful question is how it works, because once you see the mechanism you can predict where it will do well, where it will stumble, and what to write to get a better first draft.

What happens after you press Build?

Between your text and the preview there are several steps. Knowing them explains most of what you'll see.

1. A vagueness check

If the first message of a new project says what kind of thing to build but nothing about what it's for (“make me a website”), Supacode asks one short question with three or four suggested answers you can tap, instead of guessing. The question uses no credits. Anything with a real subject, even a short one like “portfolio for a photographer”, builds straight away.

2. A stack choice

Rules read the prompt for signals. Sites with pages and content get plain HTML; interactive apps such as dashboards, games and trackers get React; anything that needs a server for sign-in, saved data, payments or sending email gets Next.js. Naming a stack in the prompt overrides the rules. You can see the pick before you send: the composer shows it next to Auto, the stack menu gives a one-line reason, and you can choose another stack there.

3. The model writes files

By default, Supacode's builds run on Anthropic's Claude models. The model doesn't paint a picture of a website; it writes the files: index.html, stylesheets, scripts, more pages. On later requests it writes targeted edits to existing files rather than rewriting everything, which is why a small request should only change what you asked about. An edit that's cut off halfway is thrown away, never half-applied.

4. Checks and automatic fixes

Before you see anything, Supacode checks the result: does the code build, does every image or file it references exist, do the navigation links point to real pages? For HTML and React sites it also runs the page for a moment to catch crashes and blank screens. If problems remain, it makes up to two automatic fix attempts. If those fail too, the credits for the fix attempts are refunded.

5. A preview and a version

The result appears in the live preview and is saved as a version. Every later change is another version, and you can undo, preview or restore any of them.

How does the AI remember your site between messages?

A language model doesn't remember your project the way a person does. Each request is assembled from scratch, so what it “knows” is what Supacode sends with your new message:

• The Memory file. SUPACODE.md, shown in the studio's Memory tab: goals, design choices and rules for this project. It's read with every request.
• Your files. The project's file list, plus the contents of the files most relevant to the request.
• The recent conversation. The last twelve messages word for word, and a short running summary of older ones.

Two practical rules follow. First, anything that must hold for the life of the site, such as “never use photos of people” or “prices always include VAT”, belongs in the Memory tab, not in message thirty. Second, when a change touches something that already works, say so in the request: “keep the header and colours as they are”. The model sees your files, but it doesn't know which parts you're attached to.

What can't text to website do well?

Generated sites are good at structure, layout and first-draft copy. They're weak in predictable places, and knowing those places saves most of the frustration:

None of these is a reason not to use it. They're the reason to review. Google's guidance on AI-generated content makes the same point from the search side: it rewards helpful, original content however it's produced, so a generated page that says something specific and true about your business is fine, while one full of filler isn't. Our accessibility checklist covers the twenty-minute test worth running before you publish.

How do you write a prompt that gets a better site?

Anthropic's own prompting guide suggests treating the model like a brilliant but new employee who lacks context on your norms. That framing works well for websites. Compare two prompts for the same pottery class:

Make a website for my pottery class.

A one-page website for Saturday Clay, a beginner pottery class in a shared studio. Sections: what you'll make in six weeks, the teacher, dates and price, what to wear, and a sign-up button that opens an email to hello@saturdayclay.example. Earthy, relaxed, lots of white space.

The first will build, but every decision is a guess. The second answers the questions a designer would ask in a first meeting. Use it as a pattern:

1. What it is and for whom. “A beginner pottery class”, not “a pottery website”.
2. One main action. Sign up, book, call, buy. One, so it can be obvious.
3. The pages or sections, named. The navigation comes out right first time.
4. Your real content. Paste it. Long, pasted briefs are fine.
5. The look in a few words. Two or three adjectives and your colours beat a paragraph of styling.
6. What must not change (from the second request on). “Keep the header and colours.”

For anything you're unsure about, switch the composer from Build to Plan first: Supacode answers with a plan in the chat and changes no files, so you can fix the plan before anything is built. Ask mode answers questions about your project without editing it. For prompting habits that apply to apps as well as sites, see prompt patterns for AI app builders.

Should you start from a prompt or a template?

A prompt is best when your site is specific to you: a clinic, a class, a portfolio with your own work. A template is best when you want a proven layout fast. Supacode's gallery has official templates across HTML, React and Next.js; copying one uses 0 credits, and from there it's text to website again: you describe what to change.

Either way, the second request is where the real work happens. Our walkthrough AI website generator: build a site, then fix what's off takes one prompt from first draft to published, including what to do when the result isn't right. For the basics of a small business site, see how to build a website with AI.

What does it cost to turn text into a website?

Supacode meters AI work in credits. In the studio, the chat shows an estimate before you send a request; Supacode holds that amount, charges what the work actually cost and gives back the rest. (A first build started from the home page begins straight away and shows its estimate in the progress line.) Fix attempts that fail are refunded, and the vagueness question is free. The Free plan includes free credits every day; current plans and limits are on the pricing page.

When the site is right, press Publish and it goes live on your own supacode.app address, visible only to people with the link until you add it to the gallery. The code is yours: any version can be downloaded as a zip. The credits guide and publishing guide cover the details.