当AI开始“看教程”：Resource2Skill 让智能体从多模态资源中学会技能

TL;DR - Resource2Skill distills videos, code repositories, articles, and reference works into a hierarchical, multimodal Skill Wiki that agents can retrieve, combine, and execute. Across seven software-creation domains, it improved average performance by 11.9%.

• Skills combine text, visuals, code, and metadata, with deterministic checks for completeness, traceability, consistency, deduplication, and executability.
• Agents access skills and software backends through MCP, enabling verified code to execute directly in tools.
• The evaluation covered 4,893 skills across PPT, CAD, Web, Excel, Blender, UE5, and Reaper; human evaluators preferred skill-assisted outputs in 85.5% of non-tied comparisons.
• Video was especially valuable: removing it reduced the aggregate score from 68.9 to 59.4, while using video alone achieved 66.8.