RESOURCE2SKILL Distills Agent Skills From Tutorial Videos

A new paper called RESOURCE2SKILL, posted on arXiv, makes a small but interesting bet about where agent capability is actually bottlenecked. Rather than train on more traces or ask humans to hand-write skill libraries, the authors describe a pipeline that ingests tutorial videos, code repositories, articles and reference artifacts, then distills them into a hierarchical wiki of executable skills that combines structured text, code, visual examples, metadata, and provenance.

The headline number is a +11.9 percentage point improvement in average overall scores against agents without the skill wiki, evaluated across seven practical authoring domains. The authors report leading results on 26 of 28 main-aggregate model-domain cells, and their ablations argue that the multimodal format, the hierarchical organization, the diversity of sources, the selection strategy, and online skill acquisition each pull weight in the final score.

Why this is worth watching if you build agents rather than train them: it points at ordinary tutorial content, the kind that already exists in abundance for authoring tools, as a first-class input to agent competence. If the pipeline generalizes, the moat for agent skill quality is less about proprietary trajectories and more about who curates and structures existing multimodal resources well.

The honest caveat is that this is a preprint reporting the authors' own benchmark, and the abstract does not tell you which seven authoring domains were chosen, what the pipeline costs to run, or whether video-heavy skill extraction holds up when tutorials are scarce, contradictory, or outdated. Take the specifics as reported, not settled.

The direction still matters. If the next wave of useful agents is built by wrapping general models around well-organized domain knowledge, tooling that turns tutorial libraries into structured, executable skills is exactly the kind of quiet infrastructure that quietly becomes valuable.