skill-tuning

Empirically optimize any existing Claude skill against a measurable reward signal using a closed SkillOpt loop (rollout → reflect → aggregate → select → evaluate) plus a noise-robust held-out A/B comparator. Use when you want a skill's prose tuned by evidence rather than eyeballed — e.g. raising an agent's success rate on a bounded, scoreable task — or to c…

Independent GitHub snapshot · Automated classification · Original documentation governs