What's this about?
Most agent-skill-building approaches distill either text documentation or an agent's own past action traces into reusable skills. RESOURCE2SKILL instead mines multimodal human-created resources — tutorial videos, screen recordings, and similar — and distills them into executable agent skill libraries, the day's runaway HuggingFace paper at 117 upvotes.
What exactly is an "agent skill"?
A skill in this context is a packaged, reusable procedure an agent can invoke — like a subroutine that encodes "how to do X" so the agent doesn't have to figure it out from scratch every time. Anthropic's Claude Skills feature (covered in earlier digests) is a production example of the same idea. The novelty here is the source material: instead of writing skills by hand or extracting them from an agent's own trial-and-error, you point the system at the same how-to videos humans already use to learn things.
Why does it matter?
Human tutorial content is a vastly larger and richer training signal than the hand-crafted documentation or self-generated traces most skill-extraction methods rely on — if this generalizes, it's a scalable way to bootstrap agent capability from the same resources people already use to learn.