Somewhere on your phone or in a folder called "inspo" is a screenshot of a website you liked enough to save. Most of the time it stays there, because turning a static image into a real, working page used to mean briefing a designer and waiting. That gap is smaller than it used to be.
The old way: hand the screenshot to a designer or developer, describe what you liked about it, wait for a mockup, then wait again for it to be built. Days to weeks, depending on who you ask, plus a round or two of revisions once you see it doesn't quite match what you had in mind.
The AI way: describe the layout in the screenshot, or the type of page it is, in a prompt, and get a working structure back in seconds. You're not waiting on anyone's schedule but your own, and if the first attempt isn't quite right, adjusting the description and regenerating costs you another few seconds, not another few days.
An AI builder working from a description of a screenshot is good at picking up structure: where the hero section sits, how many feature blocks there are, whether there's a pricing table, how the footer is organized, and the general style (minimal, bold, image-heavy, text-heavy). It's reconstructing the pattern, not tracing pixels.
What it won't do is reproduce hidden interactions, backend logic, or anything not visible in the image. A screenshot doesn't show you what happens when someone clicks "sign up," so that part is still on you to define. It also won't know why a design choice was made. If a competitor buried their pricing at the bottom of the page for a strategic reason, the screenshot alone won't tell you that; it'll just replicate the position.
Say the screenshot is a SaaS homepage: a centered headline over a product screenshot, three feature cards in a row below, a row of customer logos, and a pricing table near the bottom. Describing that structure as "headline and hero image, three feature cards, a logo row, then a pricing table" gets you a page with exactly that skeleton. From there you write your own headline, describe your own three features, and either add real logos or remove that section if you don't have any yet.
This approach is useful for competitor teardowns (seeing how a similar business structures its site and adapting the parts that make sense for you), client requests that start with "can you make it look like this," and saving a design idea before the original page changes or disappears, which happens more often than you'd expect.
The same rule applies here as with any cloning approach: you're allowed to reuse a layout, not someone else's actual content. Before publishing, make sure every word, image, and brand element is genuinely yours. A screenshot is a starting point for structure. The finished site should look and read like your business, not a copy of someone else's.
If the reference has a distinctive visual style (a specific accent color, a particular font mood, a dark or light theme), mention it explicitly in your prompt rather than assuming it'll carry over from a vague description. "Dark navy background with a bright orange call-to-action button" gets you much closer to the original feel than "modern and clean." Color and type are exactly the kind of detail a screenshot conveys clearly that words alone tend to lose, so it's worth the extra sentence to spell it out. That's the difference between a draft that merely has the right sections and one that actually captures the feeling you saved the screenshot for in the first place.
Before a redesign, it's easy to lose track of what actually worked about the version you're replacing. Screenshotting your current site and describing its structure before you start over gives you a reference to compare against, and sometimes a section from the old version turns out to be worth keeping even as everything else changes. This is also a useful habit any time you're about to make a big change to a live site. A quick record of what it looked like before costs nothing and can save you from losing something that was quietly working well.
Long pages rarely fit in a single screenshot, and that's fine: describe what you can see, publish that as a first section, and add a second prompt for whatever continued below the fold. Trying to guess at content you can't actually see defeats the purpose of working from a reference in the first place; it's better to be explicit about what you're building from and fill any gaps with your own judgment rather than invented guesses about what the rest of the page might have contained.
No, describing a screenshot gives you the structure (sections, layout, general style), not the original words or images. You still write your own copy and add your own visuals.
Describe each page you want separately: a prompt for the homepage structure, then another for a pricing or about page, rather than expecting one description to cover an entire site at once.
Add more specific detail (the exact number of sections, their order, anything distinctive you noticed) and regenerate. Since it takes seconds, it's a cheap thing to iterate on.
If you've got a screenshot sitting in a folder right now, that's the test case. Describe it, generate a draft, and see how close it gets before you touch anything.