Automating SKILL.md Generation for Computer-Using Agents via Interaction Trajectory Mining

arXiv·low signal

A three-stage pipeline (segment → cluster → train skill-aware policy) mines skill libraries from GUI interaction trajectories to auto-generate SKILL.md files for computer-using agents. Skills were interpretable (5 of 8 clusters >=0.95 purity) but policy gains were marginal — +1.95pp on one benchmark and negligible on another — suggesting trajectory mining isn't yet reliable for cross-domain transfer. A useful negative signal for teams betting on auto-generated agent skills.