Can tutorial videos teach software agents reusable skills?

Resource2Skill argues that the raw material for agent skills does not have to be an agent's own history. The abstract says existing skill libraries are "mostly hand-written, text-centric, or derived from agent traces, leaving tutorial videos and other multimodal human resources largely underused." The framework distills tutorial videos, repositories, articles, and reference artifacts into executable skills, and the paper reports that skill access improves artifact quality by +11.9 percentage points on average over no-skill agents across seven authoring domains and four backends. It also beats two agentic-harness baselines in 26 of 28 main-aggregate model–domain cells.

The paper frames the need in terms of what authoring tasks demand: producing slide decks, spreadsheets, web pages, 3D scenes, CAD designs, and audio projects depends "less on isolated factual knowledge than on reusable procedural know-how", meaning how to decompose a goal, which tool or API pattern to use, what intermediate state to inspect, and how to recover when an operation fails. The design keeps each resource type for what it carries. Skills live in a hierarchical multimodal Skill Wiki whose entries combine structured text, code, visual examples, metadata, and provenance. Videos capture temporal operations and visual effects, code captures executable tool patterns, and articles or artifacts supply conceptual and stylistic grounding. At inference, agents retrieve and compose skills from the wiki. When coverage is insufficient, "the same construction operator can acquire new skills online", so offline construction and online gap filling share one mechanism.

Against the neighboring notes, the difference is where the skills come from. How can agent systems share learned skills across users? and Should successful and failed episodes be processed differently? both treat trajectories as the source, and Can a separate trained curator improve skill libraries better than frozen agents? curates a repository over task streams. Resource2Skill starts from artifacts people already made, which suggests a second supply line for skills that does not wait on an agent to fail and succeed its way to them. It also parallels Can codified expertise let non-experts match specialist output?, where domain knowledge is captured for the agent by hand. And Can agents learn new skills without forgetting old ones? shares the retrieve-and-compose loop.

The excerpt does not establish several things. It gives no definition of the overall score, no sample sizes, and no statement of what the two harness baselines are. It does not compare against hand-written or trace-derived skill libraries directly, so the claim that human resources are a better source is not tested in what we can see. It does not report how well online acquisition works, what construction costs, or how much resource quality matters. The safe reading is narrower: for these authoring domains, skills distilled from human-created resources beat both no skills and the harness baselines. Whether they beat other skill sources is left open.