Strip the tutorial video from a skill wiki, lose 9.5 points

Notes on RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources (arXiv:2606.29538) — Yijia Fan, Zonglin Di, Zimo Wen, Yifan Yang, Mingxi Cheng, Qi Dai, Bei Liu, Kai Qiu, Yue Dong, Ji Li, Chong Luo · June 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill evolution

Most agent skill libraries come from one medium — hand-written instructions, or traces an agent left behind after solving a task once. Resource2Skill goes wider: it mines the tutorials, code repositories, articles, and reference files humans already produce for seven software-authoring domains — slide decks, spreadsheets, web pages, Blender scenes, Reaper audio projects, CAD drawings, Unreal Engine 5 builds — and runs each through multimodal skill distillation. A resource becomes a skill entry that keeps its structured instructions, a code snippet, a visual example, and provenance metadata bundled together. Entries land in a hierarchical Skill Wiki that an agent searches and composes from at task time. When nothing in the wiki fits, the same pipeline runs live as online skill acquisition, minting a new skill on the spot instead of leaving the agent to improvise.

Across four backend models and the seven domains, wiki access lifted the average task score from 45.0% to 56.8% — an 11.9-point gain — and beat two competing agentic harness baselines in 26 of 28 model-domain pairings, every one significant at p<10⁻³. The gain wasn't even: Unreal Engine 5 jumped 38.2 points while Reaper audio editing moved only 4.1, because agents were already decent at Reaper with no skill library at all. Ablations point to one ingredient driving most of that gain — strip tutorial video out of the wiki and the average falls from 68.9% to 59.4%, a bigger hit than pulling code or articles. A video-only wiki even beat a three-source wiki with everything but video, by 7.4 points.

Key numbers

Overall score lift with skill access	+11.9 pp (56.8% vs 45.0%)
Model-domain cells beating harness baselines	26 of 28
Score drop when tutorial video is excluded	-9.5 pp (68.9% → 59.4%)
Human A/B preference for Resource2Skill	85.5% win rate (200 ratings)
Largest vs. smallest per-domain gain (GPT-5.4)	+38.2 pp (UE5) vs +4.1 pp (Reaper)

What this means for skill authors

For skill authors, this changes what counts as a legitimate source. A hand-written markdown file discards exactly the signal the ablations say is hardest to replace: a timed visual demonstration sitting next to its on-screen effect. Bundle a short walkthrough into a skill entry alongside the instructions, and you're carrying information a purely textual skill can't fake.

For catalogs and marketplaces, the construction pipeline matters as much as the extraction model. Five deterministic gates — completeness, provenance, deduplication, modality consistency, and code executability — sit between a vision-capable model producing something and that entry shipping in the wiki. SkillFed's dual-tool scanning takes the same qualitative stance: a skill earns trust by clearing independent checks before publication, not by virtue of an LM having written it.

The uneven per-domain lift matters more than the headline average. An 11.9-point improvement hides a 34-point spread between UE5 and Reaper, and what decided it was how strong the no-skill baseline already was in that domain. Treat a skill library's payoff as domain-conditional, not a flat multiplier, before assuming a catalog's average score will hold for your specific task.

Skills related to this research

canvas Canvas organizes visual references in Obsidian by letting you add images, PDFs, and wiki pages to canvas files with automatic positioning inside zones. It integrates with the /banana image capture skill and supports multiple operations like creating new canvases, managing zones, and listing existing canvases.★ 10,007 obsidian-bases obsidian-bases enables you to transform vault notes into queryable, interactive views using Obsidian's native database layer. Define filters, computed properties, and multiple view types (table, cards, list) in .base files stored as YAML, then embed them directly in your notes. Perfect for dashboards, entity trackers, and content organization.★ 10,007 save save captures valuable chat content and files it as a permanent wiki page with the right structure and metadata. It analyzes conversations to pick the best note type—synthesis, concept, source, decision, or session—then creates frontmatter, places the note in the correct folder, and updates your wiki's index and log. The skill respects your vault's organization mode and handles concurrent writes safely.★ 10,007 wiki wiki transforms Obsidian into a compounding knowledge system by scaffolding vault structure from your description, organizing sources and entities, and maintaining a hot cache of recent context. Unlike RAG systems, the wiki persists as a structured artifact where cross-references and contradictions surface automatically. Route ingest, query, and maintenance tasks to specialized sub-skills while the skill handles vault setup, indexing, and knowledge layer management.★ 10,007

Related notes

References

Fan, Di, Wen, Yang, Cheng, Dai, Liu, Qiu, Dong, Li, Luo. "RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources." arXiv:2606.29538.