Why Coding Agents Build the Same Landing Page Every Time

Ask four different coding agents for a landing page and you get four versions of the same page. Centered hero, gradient headline, three feature cards, a logo row, a three-tier pricing table. The output is competent and completely interchangeable. The usual explanation is that models are averaging their training data, which is true but not useful, because it does not tell you which word in your prompt to change.

The swaggin registry gives a more specific answer. Every site in it is tagged with one or more aesthetic styles by a human who looked at it. Counting those tags across all 189 sites shows exactly where real designs agree and where they diverge.

What does the style distribution actually look like?

Across 189 curated sites, marketing-landing appears on 124 of them, roughly two thirds. No other style comes close. The next largest are gradient-vibrant and elegant-consumer at 39 sites each, then playful at 34 and minimal-swiss at 24. Sites average 1.83 style tags each.

Why does one tag dominate?

Because marketing-landing describes a structure, not a look. It means the page sells something above the fold, stacks feature sections below it, and closes with a footer. 189 of 189 sites have a navigation bar. 189 have a hero. 165 have a feature section. That skeleton is genuinely near-universal, and it is the part your agent has already memorised.

So when you ask for a landing page, you are naming the layer where real designs already agree. The model returns the average of that layer, correctly. The prompt contained no information about the layer that varies.

Where do real sites actually differ?

In the second tag. 107 of the 124 marketing-landing sites carry at least one more style, and those pairings are what separate them visually. The five most common combinations account for most of the registry's range.

A gradient-vibrant marketing page and a minimal-swiss marketing page have the same skeleton and share almost no visual decisions. One resolves its hero with a mesh gradient and saturated CTA fills. The other resolves it with hairline rules, a strict grid, and no shadow at all. Same structure, opposite result.

How do you get a different page out of an agent?

Name the axis that varies, not the one that does not. Three changes do most of the work:

1. 01Drop the word that means structure. "Landing page" is scaffolding. The agent has it already. Spend your prompt elsewhere.
2. 02Name a style with visual cues attached. terminal-dev, editorial, brutalist, luxury. Each of these implies specific decisions about type, density and colour that landing page does not.
3. 03Give it something to look at. A style name is still text, and the model resolves text with its priors. A screenshot in the vision context replaces the prior with an observation.

The third one matters most, and it is the reason swaggin returns image bytes rather than links. An agent cannot read typography off a URL it is unable to open. It can read it off a screenshot sitting in its context window while it writes the component.

Does naming a style actually change the output?

It changes it more than adjectives do, because a style id in the swaggin taxonomy is backed by written visual cues and by real sites carrying that tag. brutalist resolves to 8 specific sites. retro-computing resolves to 6. The agent is not guessing what you meant, it is retrieving examples.

That is also the honest limit of this. 14 styles is a small vocabulary. It covers the range a coding agent is usually asked for, and it will not describe every design that exists.

The short version

• Two thirds of curated sites share one style tag, and it is the structural one.
• Sites average 1.83 tags. The second tag carries the visual identity.
• Prompting for the structural layer returns the average of that layer.
• Style ids beat adjectives because they resolve to real examples.
• Images beat style ids because they remove the guess entirely.