Large language models (LLMs) can now build complete interactive websites from natural‑language prompts, yet non‑expert users often struggle to describe sophisticated visual and interaction designs. The open web already hosts a massive pool of such design knowledge in mature site implementations. To bridge this gap we introduce Website‑derived Visual Skill Extraction (WebViSE), a framework that automatically abstracts existing sites into reusable Visual Skills and organizes them into a structured skill library.\ \ WebViSE arranges its skills in a multi‑parent directed acyclic graph (DAG) and jointly evolves skill descriptions and the DAG topology using Skill‑conditioned Potential Queries. This enables efficient matching between diverse user instructions and relevant skills. On Qwen3.8‑Max, the auto‑evolved DAG achieves better retrieval performance than non‑hierarchical LLM baselines while costing only about 39.09% of the estimated retrieval budget.\ \ The framework can be injected as plug‑and‑play guidance for any model, helping them produce more visually appealing and interaction‑consistent web pages. For open‑source models, WebViSE can drive a data‑synthesis pipeline that internalizes visual design intelligence into model weights. Experiments across multiple web‑development benchmarks and human evaluations consistently show that WebViSE improves visual quality and interaction fidelity of the generated artifacts.\ \ All extracted Visual Skills, the associated skill DAG, and the code are released at https://github.com/YifeiSun-0710/WebViSE.\ \ Review: WebViSE offers a scalable approach to systematically harvest design knowledge from real‑world websites, balancing retrieval efficiency with generation quality, and paving the way for practical LLM‑driven website creation.