oem-design-capture - SKILL.md Agent Skill

name: oem-design-capture description: Brand and layout analysis using vision models. Captures OEM brand tokens, page layouts, and component styles via a three-pass pipeline using Kimi K2.5 for visual analysis.

OEM Design Capture

Analyses OEM websites for brand identity and layout patterns.

Three-Pass Pipeline

Pass 1: Brand Token Extraction (per OEM, infrequent)

Input: Desktop screenshots of homepage + 2 vehicle pages
Model: Kimi K2.5 Thinking Mode (temp=1.0)
Output: brand_tokens.v1 JSON

Pass 2: Page Layout Decomposition (per page type)

Input: Desktop + mobile screenshots + cleaned DOM HTML
Model: Kimi K2.5 Instant Mode (temp=0.6)
Output: page_layout.v1 JSON

Pass 3: Component Detail (per unique component)

Input: Cropped screenshot + DOM fragment
Model: Kimi K2.5 Instant Mode (temp=0.6)
Output: CSS-equivalent JSON per component

Prerequisites

SUPABASE_URL and SUPABASE_SERVICE_ROLE_KEY for storing design data
GROQ_API_KEY for pre-screening
TOGETHER_API_KEY for Kimi K2.5 vision analysis
LIGHTPANDA_URL for browser rendering (primary) or WORKER_URL + CDP_SECRET (fallback)

Note: specs_json is now always populated on every product upsert via orchestrator.buildSpecsJson(), so page generation can rely on specs data being present for all products.

Triggers

Initial onboarding: full capture
Screenshot pHash change >30%: re-capture affected page
Manual trigger via admin API
Quarterly: full audit all OEMs

Input

{
  "oem_id": "ford-au",
  "trigger": "onboarding",
  "page_type": "homepage"
}

Output

{
  "brand_tokens_updated": true,
  "layouts_captured": 5
}