Instead of retraining weights, SkillOpt iterates on the instructions and documents an agent is given for a skill — refining the skill definition itself based on observed performance, so improvement happens at the prompt/skill layer rather than the model layer.