Microsoft Turns YouTube Tutorials Into Agent Skills

Microsoft Research has a paper called RESOURCE2SKILL that answers a question the whole Skills wave has been dancing around: where do the skills actually come from? Everyone agrees agents get better when you hand them reusable procedural knowledge. Nobody wants to write all of it by hand. The answer here is to mine it, from tutorial videos, code repos, articles, and reference artifacts, the same messy pile of human-made material a new hire would learn from.

The trick is what they build with it: a hierarchical multimodal Skill Wiki. Each entry fuses structured text, executable code, visual examples, metadata, and a pointer back to the source. The insight is that different media teach different things. Videos show you the temporal workflow and what success looks like, code captures the actual tool calls, articles give you the why. Cram them into one entry and the agent gets a fuller picture than any single format could give.

The numbers hold up. Across seven authoring domains the framework lifted average performance by 11.9 points over agents with no skills, and beat comparable baselines in 26 of 28 model-domain combinations. The bigger point is a supply-chain one. If a company can point a pipeline at its existing documentation and out come executable agent skills, then every tutorial video ever recorded is suddenly training data for your workforce of agents. Paper at arXiv 2606.29538.