Home › Digital Tracks

3 November 2026 · 5 min read · Silvia Volpe

How to build a website with AI from scratch (an honest guide)

I use AI every day, and I do not hide it. This site was built with it. In my experience, a website that used to need a team and a long timeline now takes about half the time and half the resources. In my view, that is also why many AI-built sites end up looking alike. Using AI and using it well are two different things, and the gap between them is where a website is won or lost.

AI is not the artist. It is another tool to draw and create with, like Photoshop or Canva. What makes the work good is still the person using it: critical thinking, creativity and experience.

1. I learned AI the way I learned the web: on my own

I started using AI in 2023, in the early days, with no course and no guide. I tried, broke things, compared results and kept notes on what worked. It is the same path I took at 13 with HTML and CSS.

There was no school behind either of them. There was curiosity, a lot of trial and error, and the habit of asking one question after every result: can this be better?

2. The prompt is the last step, not the first

Before I type a single prompt, I write the brief: who the site is for, which pages it needs, which languages, which tone, and which limits cannot be broken. For this site the limits were fast loading on every device, a layout that works on phones first and the same quality in English and Italian.

A clear brief makes every prompt better. If you do not know what you want, the AI will fill the gaps with generic choices.

3. Prompting that actually works

There is no magic formula. But a few practices are consistently recommended by the people who build these models, and they are the ones worth learning.

• Set the rules first. Put your standing instructions (role, tone, limits, things to avoid) in a system prompt or in your tool’s custom instructions, so every request starts from them.
• Brief it like a smart new colleague. State the goal, the audience, the format and the constraints. A simple test: if a colleague with no context would be confused by your prompt, the AI probably will be too.
• Explain the why. “No heavy images” works better as “no heavy images, because the page must load fast on a phone”. The reason lets the AI apply the rule to cases you did not list.
• Show an example. A few relevant examples of the result you want are one of the most reliable ways to steer format, tone and structure.
• Give it your material and allow “I don’t know”. Paste the real information and tell the AI to say when it is not sure. This reduces made-up facts. It does not remove them.
• Split big jobs, then ask for a review. Have it draft, then check the draft against your criteria, then refine.

These come from Anthropic’s prompt engineering guide. Even with all of them, a model can still get things wrong, so the checking stays with you.

4. This is where the fan club school shows

When I built my first fansite, there were no tools doing the optimisation for me. I did it by hand, because a slow page meant visitors left. That habit never went away, and AI does not have it unless you ask.

So I check what no prompt checks by itself:

• the weight of every image, and whether a smaller one would look the same;
• one title and one H1 per page, and the two saying the same thing;
• alt text, links between pages and a sitemap that matches the pages;
• a clean structure for two languages, with hreflang that points both ways;
• how the page loads on a slow connection, in more than one browser, on a real phone.

5. AI sounds sure. That does not make it right

An AI assistant can sound confident even when it is wrong. It can invent figures or state things it cannot source, and sometimes it changes things you never asked it to touch. I treat every output as a draft.

My rule is simple: I never publish a number I cannot verify. If it is uncertain, I remove it. I also check what changed after each edit, because a quiet “improvement” can move a layout that was already right.

Two small examples. I rejected a pointing-hand cursor because it looked like a middle finger. And I caught a header that glitched only on the first visit in a clean Chrome and never again after a refresh. The AI could not see either, because I opened the site like a visitor would.

6. What AI cannot give you

It cannot give you taste, a story or a point of view. The glitter, the Y2K look and the pink are mine, and they come from the fansite years. The AI helped me build them faster. It did not decide that they belonged.

It also cannot decide what to leave out, and that is often the most important choice. And once the site exists, there is the next question: can machines read it? That is the idea behind SEO for AI search.

Before you publish: a checklist

• Is the brief written down? Audience, pages, languages, tone, limits.
• Did you set your rules and give examples before asking?
• Is every fact and figure verified?
• Do titles, H1s, alt text and links check out?
• Did you test on a real phone and in two browsers?
• Does it still sound like you? Remove generic phrases and anything that reads like a template.

The takeaway: AI makes it faster, judgement makes it good

Anyone can generate a website now. Fewer people can look at it and know what is wrong with it. That takes critical thinking, creativity and experience, like every tool before it. I built that second skill on fansites, and it is the one that stays yours.

♪♪ What do I listen to for focus and fresh ideas?My songs to work and daydream to.▶︎ Play

No Spotify? Listen on: YouTube