designseeddocs

Generator

Supported sites & limitations

2 min read

The current extractor is deterministic — it samples computed styles and runs heuristics, no AI synthesis (yet). That sets a clear ceiling on what it can do.

Works well on

  • Public marketing sites with consistent brand expression (Stripe, Vercel, Airbnb, Linear).
  • Single-page apps that hydrate within ~500 ms of DOM ready.
  • Sites under 5 MB of DOM — we cap the in-page walk at 1500 visible elements.
  • Sites that don't block bots — we use a custom DesignDropBot UA. Hostile rate-limiters may serve us a 429 or a captcha page, in which case the extraction captures whatever the bot saw, not the real brand.

Doesn't work yet

  • Auth-walled pages. The extractor runs unauthenticated. Anything behind a login produces a degraded result (often a blank or near-blank palette). Workaround: extract from a public marketing page that uses the same design system.
  • Pages requiring user interaction to render real content (e.g. accordions that hide everything until clicked). Workaround: pick a content-rich landing page.
  • Iframes. We only sample the top document. Embedded UIs in iframes are invisible to the extractor.
  • Canvas-heavy interfaces. Pixel data inside <canvas> is opaque to computed-styles sampling.

Heuristic ceiling

Even on a perfect input, the deterministic mapper has known weaknesses:

  • Brand primary detection picks the most-saturated visible color. On minimalist sites where the primary is a near-black, the heuristic surfaces an accent instead.
  • Mono font detection is missing. The extractor returns a system fallback for --font-mono until we add a fixed-width digit detector.
  • Dark theme synthesis is opt-in: only triggered when the source page declares color-scheme: dark or sets a dark theme-color meta. We don't invent dark themes from scratch.

What's coming

The next iteration ships an AI synthesis layer that takes the raw harvest plus a screenshot and names tokens semantically. That closes most of the heuristic gaps above and bumps the average confidence score noticeably. Track progress in the TASK-4 thread.