A skill library for software agents

Resource2Skill

Resource2Skill gives software agents a library of real human skills, so they ship polished sites, decks, spreadsheets, 3D scenes, and music instead of rough drafts.

Yijia Fan1,*, Zonglin Di2,*, Zimo Wen3,*, Yifan Yang1,†, Mingxi Cheng1, Qi Dai1, Bei Liu1, Kai Qiu1, Yue Dong1, Ji Li1, Chong Luo1

1Microsoft   ·   2UC Santa Cruz   ·   3Shanghai Jiao Tong University

*Equal contribution   ·   †Corresponding author: yifanyang@microsoft.com

Microsoft UC Santa Cruz Shanghai Jiao Tong University
See the showcase Paper Code Dataset Watch Video

Project Video

Watch the system in motion.

A short walkthrough of skill construction, selection, and execution across authoring domains.

YouTube ↗

How it works

One library, built once — browsed, selected, and run by the agent.

Resource2Skill distills tutorials, repositories, and reference artifacts into a hierarchical Skill Wiki. At run time an agent browses candidates, selects from text, visual, and code views, and executes the chosen skills through MCP into each domain's real software backend — the same construction operator is reused online when the offline pool falls short.

Resource2Skill pipeline: distill resources into a Skill Wiki, then browse, select, and execute through MCP into domain backends.

Results

Skills lift every domain.

Task scores on the 7-domain benchmark with a GPT-5.4 backbone — the same agent, with vs without the Resource2Skill library. Skills help everywhere, from +4.1 on Reaper to +38.2 on UE5, for a +15.0 average lift.

Without skillsWith skills
68.7
82.4
Web
+13.7
58.6
76.4
Excel
+17.8
73.2
77.3
Reaper
+4.1
55.4
64.8
PPT
+9.4
29.5
44.1
Blender
+14.6
48.7
55.7
CAD
+7.0
29.1
67.3
UE5
+38.2
51.9
66.9
Avg
+15.0

The Showcase

Same task. With skills vs without.

Large, synchronized comparisons across the released domains. Each row uses the same task and model; the only visible difference is whether the agent can browse and compose Resource2Skill skills.

Resource2Skill distills tutorials, code, documents, and reference artifacts into executable skill entries that agents can browse, select, compose, and run through real software tools.

Technical details & paper →