Brand Extract

brand-extract

Turn a live website into a complete, machine-consumable Brand Kit — identity, semantic color palette, typography, voice — by measuring the real page, not guessing from memory. This is the methodology behind a brand-extraction project: the target site is open in a secondary in-app Browser tab, and you drive it with the agent-browser tool.

The live kit page (brand.html)

The extraction project opens with brand.html as the active tab — a self-contained brand-kit page (template: brand-extract/templates/brand-kit.html) that the daemon renders from brand.json. The daemon pre-seeds it with a deterministic first paint — a harvested logo, an approximate palette, font families, and a few cover images — so it is NOT all-skeleton when it opens. Your job is to replace that seed with measured truth and fill in the rest, progressively, so the user watches it complete module by module. You never hand-edit it: you write brand.json, then run od brand preview <brandId> and the daemon re-renders the page (the page soft-reloads itself while extracting). Optimize for fast first paint and progressive fill-in — write a partial brand.json and preview it the moment you have a name, a couple of colors, and a logo, then preview again after each field group rather than batching the whole kit to the end.

The trap to avoid: an LLM left alone regresses to the mean — Inter, an indigo accent, a purple gradient. That is off-brand for everyone. Every value you emit must trace to something you measured on the page.

The three-step chain

Work in order. Skipping straight to writing brand.json is how off-brand, hallucinated kits happen.

1. Measure (drive the open Browser tab)

Use agent-browser against the selected browser tab (its URL/title are in your run context — treat "this page" / "the site" as that tab):

1. agent-browser get url / get title to confirm the target.
2. agent-browser snapshot before extracting anything.
3. Harvest the real design language from the DOM/CSS, not the screenshot alone:
   • Colors — frequency-rank color literals and resolve the seven semantic roles: background, surface, foreground, muted, border, accent, accent-secondary. The most frequent near-white/cream is usually the background; the most frequent chromatic mid-saturation color is usually the accent.
   • Typography — the @font-face names and font-family declarations for display, body, and (if present) mono. Note weights actually used.
   • Logo (save MULTIPLE candidates) — extract every logo asset you find and save each as a file under logos/: the inline header/nav <svg> (write the literal <svg>…</svg> markup verbatim to logos/header.svg — do not just reference it), any <img> logo, apple-touch-icon, favicon, and og:image. Fetch the asset URLs directly — never leave logo.primary empty when the site has any mark. Set logo.primary to the best vector / transparent lockup (SVG wordmark > apple-touch-icon > favicon > og:image) and list the rest in logo.alternates; the kit page renders them as switchable thumbnails. (The daemon auto-fetches a favicon/og:image fallback into logos/ so the page is never logo-less, but that safety net is no substitute for saving the real wordmark.)
   • Imagery (save 6–8 of the site's LARGE / COVER / HERO images) — this is the Images module. Harvest the site's actual big representative pictures and save them into imagery/: the og:image/twitter:image social card, the hero/banner art, the largest <img> (resolve the highest-res srcset / <picture> source), CSS background-image hero blocks, product or app screenshots, and illustration/photography samples. Filter by rendered size — keep only big images (roughly ≥320px on the long edge) and drop icons, sprites, logos, avatars, and tracking pixels. List them in brand.json as imagery.samples (see shape below); the kit page renders them as a clean labeled Images gallery (a thumbnail grid). Pick 6–8 varied, on-brand images — never UI chrome or icons. (The daemon runs a deterministic cover/hero-image fallback at finalize so the gallery is rarely empty, but that safety net is no substitute for picking the real hero images.)
   • Voice — representative headings, taglines, and body copy to ground the voice; quote-level fidelity, not generic marketing speak.
4. Save any self-hosted webfont files you can fetch into fonts/.
5. Capture one page screenshot as visual evidence when it helps.

Anti-bot wall → ask the user (do NOT bypass)

If the page is an anti-bot interstitial instead of the real site — Cloudflare "Just a moment…", "Verify you are human", "Attention Required", DataDome, PerimeterX, Incapsula — stop measuring and emit a <question-form> asking the user to clear it by hand in the Browser tab:

<question-form id="cf-verify" title="Verify in the browser">
[
  {
    "id": "ready",
    "type": "radio",
    "label": "The site is behind a verification wall. Please complete the check in the Browser tab on the right, then choose Continue.",
    "options": ["Continue — I cleared the wall", "Skip — extract from public knowledge instead"]
  }
]
</question-form>

Then end the turn. When the user submits the form, re-run agent-browser snapshot on the now-unblocked tab and resume measuring. Never attempt to solve CAPTCHAs or bypass the wall yourself. If the user picks "Skip", fall back to your knowledge of the brand's public identity and clearly mark each such value (from brand knowledge) in its usage/notes.