The Hypothesis

Can an AI sub-agent crawl an existing website and extract a complete, usable brand stylesheet automatically?

The Concept

Every AI content generation tool needs brand constraints — or it produces generic output. But most businesses don't have a neatly packaged brand stylesheet. Their brand lives in their existing website. This experiment tests whether an AI sub-agent can crawl a URL, extract the colour palette, typography, button styles, spacing, border radius, and link treatments, then auto-generate a locked brand CSS that constrains all future AI-generated content.

Brand extraction via crawling

The hypothesis

Can an AI sub-agent crawl an existing website and extract a complete, usable brand stylesheet automatically?

The concept

Every AI content generation tool needs brand constraints — or it produces generic output. But most businesses don’t have a neatly packaged brand stylesheet. Their brand lives in their existing website. This experiment tests whether an AI sub-agent can crawl a URL, extract the colour palette, typography, button styles, spacing, border radius, and link treatments, then auto-generate a locked brand CSS that constrains all future AI-generated content.

How it works

1. User pastes website URL
2. AI crawls key pages — homepage, product pages, about — sampling visual patterns
3. Extract visual DNA — colours, fonts, spacing, button styles, border radius, link treatments
4. Generate brand CSS — structured stylesheet with variables and component styles
5. User reviews and approves — preview applied to sample components, tweak if needed
6. Lock as immutable constraint — all AI-generated content inherits this CSS

From URL to locked brand stylesheet in under 2 minutes.

What it explores

• Can automated crawling reliably extract a coherent brand system from a live website?
• How many pages need to be crawled before the extraction is representative?
• Do AI-extracted brand styles match what a designer would define manually?
• Can the locked CSS prevent AI-generated content from drifting off-brand over time?
• What happens when a site has inconsistent styles — can the AI pick the “intended” brand?

What we found

• Extraction accuracy was 85% compared to manually defined brand guides
• Colour palettes and typography were near-perfect
• Spacing and layout patterns were the weakest — inconsistent responsive styles produced noisy results
• Sites with clear design systems extracted well; older sites with accumulated style debt did not
• User review step caught most errors in under 30 seconds

Learnings

• Crawling 3–5 pages is sufficient — more pages add noise without improving accuracy.

• The quality of extraction depends on the quality of the source — sites with a design system or CSS custom properties produce dramatically better results than those with style debt.

• The “immutable constraint” model works — locking the CSS prevents AI-generated content from drifting off-brand over time.

• A lightweight human review step is more practical than chasing full automation — 30 seconds of approval catches what the AI misses.

Where it goes next

This is now a core onboarding step for Flywheel. When a new client connects, the first thing the system does is crawl their site and lock in the brand CSS. The question we’re exploring now: can the same approach extract tone and voice — not just visual style?