Empty AI login pages are a layering problem: five CSS-only layouts lift coverage from 5.8% to 24.8%
Teams burning time on AI mockups get a reusable checklist that converts "make it look good" into numbers a machine can follow and a human can verify. A light-dependent layout that reads as filled at just 8% coverage also means the fix is cheap even with zero brand material on hand.
AI defaults to a 340px login form centered on a white 1440×900 canvas, which covers 5.8% of the viewport and leaves zero structure behind it. The article's claim is that emptiness is a layering problem, not a field-count problem: one layer of focus with nothing behind it reads as an unfinished page. It measures any layout with three metrics — content block count, coverage (the union of content bounding boxes over the viewport, sampled on a 6px grid), and the content bounding box — and argues block count predicts emptiness better than area does.
Five pure-CSS layouts each anchor a second layer differently. A 54/46 split puts brand narrative, three numbered value points, and a client logo wall on the left. A centered glass card over a radial-lit dark field gets the most counterintuitive result: 8.1% coverage and 91.9% blank space, statistically close to the default, yet it reads as filled because background light replaces dead white. A typographic layout makes the 46px headline itself the visual weight. A product-stage layout fakes the product UI in CSS so users see the interface before logging in, reaching the top coverage of 24.8%. A narrative split answers "why register" on the left while the right side runs a three-step flow.
The article also ships a prompt for AI tools: seven rules that are all countable, like form width locked to 360–420px, a non-form layer taking at least 25% of the viewport, second-layer content requiring three readable blocks, and a minimum of eight content blocks on the first screen. Every rule is something a human can verify after generation, which is what makes the AI's output fixable instead of a lucky draw.
Perceived emptiness and measured emptiness diverge sharply: V2 and V0 are statistically close (91.9% vs 94.2% blank) but read completely differently, so a single coverage number won't tell you whether a design feels finished — you have to know what the background is doing.
Block count predicts emptiness better than pixel area, which matches how users actually scan: a sparse layout built on a few large regions still feels hollow, while many structured elements read as populated regardless of proportion.
Countable prompt rules generalize beyond login pages to any AI UI work; the same principle that turns "add a spring" into "spring to 1.05 then settle to 1.0" turns "keep it airy" into a floor of eight content blocks and a 25%-viewport second layer.
The cheapest fix in the set is the one that adds nothing but light — a radial gradient over a dark field — which suggests background treatment should be treated as a layout decision, not a decoration pass.