Computer-use agents can handle increasingly complex GUI tasks, yet unfamiliar software and workflows that require human domain knowledge remain challenging. Human demonstrations convey procedural knowledge within the same GUI interaction space; reuse demands interpreting actions in context and separating workflow rules from task‑specific details. We therefore introduce the Work2Skill framework to capture human GUI experience and induce reusable agent skills. WorkRecorder supplies the infrastructure to align screen recordings, mouse and keyboard events, and accessibility‑tree data. Using WorkRecorder we build WorkDemo-350, the first synchronized multimodal dataset of human desktop GUI demonstrations, comprising 350 recordings screened for completeness and task success. We then propose an agentic skill‑induction method that inspects evidence on demand to produce reusable skills that capture workflow rules, step dependencies, and completion checks. When four execution models are equipped with skills induced by this method under shared settings, Work2Skill achieves the highest scores and success rates. Compared with the No‑Skill baseline, mean score improvements are 4.67‑10.54 points on MyPCBench Bounded Action and 3.11‑7.50 points on a 286‑task subset of OSWorld‑Verified. Code and data are available at https://github.com/china-qijizhifeng/Work2Skill.
Review