Hypothesis
If AI design agents learn by imitating the pages they read, then the sites they scrape are their de facto design curriculum. Design-inspiration roundups, agency blogs, and framework documentation sit near the top of that reading list. Our hypothesis: these sources reliably teach system-level conventions — reusable tokens, dark-mode switching — but rarely demonstrate component-level responsiveness through container queries, leaving agents fluent in fixed-viewport visual conventions and underexposed to the one CSS feature built for adaptable components. To test it, we ran an automated CSS-technique scan dated 2026-08-30 across ten popular sources and counted which of four techniques each page actually exposes.
Method
We scanned 10 publicly listed websites: Reddit’s web-design inspiration discussion, Hook Agency, Figma’s web-design trends resource, Chili Piper, Marketer Milk, Bootstrap, Tailwind CSS, Fairfield Programming, Ionic Framework, and React Bits. The detector checked for four patterns:
- CSS custom properties: reusable, author-defined values (MDN)
- clamp(): a function that constrains a value between a minimum and maximum (MDN)
- Container queries: queries based on an element’s containing context (MDN)
- Dark-mode detection
This review is based on automated CSS scans of public pages and cited documentation — we did not run the sites hands-on. Two scope notes matter. First, when the analyzer requested DesignRush’s minimal-website-design roundup it received HTTP 403, so that source contributed no retrieved evidence and is excluded from every count below. Second, the scan reports presence or absence of CSS patterns only; it does not measure visual quality, accessibility, performance, or author intent, and it says nothing about industry practice beyond these ten pages.
Results
Six of 10 scanned sites used CSS custom properties. Five of 10 included dark-mode detection. Two of 10 used clamp() for fluid typography.
Container queries: 0 of 10. Read that carefully — it means no scanned page exposed the detector’s target pattern. It does not mean these sites cannot respond to viewport changes; media queries and other responsive techniques may well be present. It means the specific CSS feature that teaches component-level adaptation never appeared in what the scanner could see.
For ecosystem-visibility context (not a proxy for design quality), the GitHub snapshot dated 2026-08-30 recorded 174,678 stars for Bootstrap, 97,398 for Tailwind CSS, 52,635 for Ionic Framework, and 46,467 for React Bits; related resources in the same snapshot — Docz at 23,581, Awesome Design at 17,467, and Dev Landing Page at 1,346 — show how widely these references circulate. Visibility, again, is not quality.
Conclusion
The pattern is consistent with our hypothesis. The corpus agents most easily reach teaches tokens (6/10) and color-scheme awareness (5/10) far more often than it teaches container queries (0/10). An agent trained on these pages would learn to define reusable values and respect user color preferences while rarely — if ever — seeing a component adapt to its own container rather than the viewport.
Three caveats keep this honest. The zero could partly reflect what one automated detector can see on one day; a different scanner or a deeper crawl might find container queries elsewhere on these domains. The scan cannot say whether pages use the techniques well. And ten pages is a sample, not a census.
Still, the practical implication holds: if you are building or fine-tuning a design agent, audit your inspiration corpus the way we did. If container queries are absent from it, the agent will not absorb them by osmosis — seed the training mix with explicit container-query exemplars, or measure whether the absence shows up in what the agent ships.