New Shanghai AI Laboratory paper finds that training on verified runs of human-written agent skills makes a model a better agent, even without the skill files.

Prompt-time skills depend on retrieval and instruction following, and fine-tuning on verified skill runs reduces that dependence.

Turning each skill file into a sandboxed task with a pass-or-fail checker produced training data that lifted Terminal-Bench 2.1 success by 19.10 points in Claude Code.

Skill files usually sit in the prompt, so they only help if the agent finds and follows them.

SkillGym turns each skill into a sandboxed task with a code checker, then trains on the runs that pass.

In Claude Code, Qwen3.5-35B-A3B jumped from 39.33% to 58.43% on Terminal-Bench 2.1. With no skill files, it scored 26.81% on SkillsBench, beating the base model with skills at 23.34%.

Loading the skills on top still helps, lifting it to 51.47%.