VISUALSKILL: Multimodal Skills for Computer-Use Agents

In three linesVISUALSKILL introduces hierarchical multimodal skills for computer-use agents. Combining authored documentation with live UI exploration, the system improves Claude Opus 4.6 performance by +15.3 points on CUA-World and OSExpert-Eval (0.456 vs 0.303 baseline). Visual figures outperform text-only descriptions (+8.3 points).

Summary generated by Claude — human-verified