Skill Tuning

A user-invoked gate skill: forces variant exploration of an existing skill's output quality through human-judged blind A/B comparison, accumulating a preference log that serves as both the decision basis for current iteration and a future RLHF-lite dataset.

This skill is the feature hat counterpart to skill-refactor's refactor hat: where refactor preserves behavior, tuning deliberately explores behavior changes to find better outputs. Where refactor uses LLM-as-judge (reliable for binary equivalence), tuning uses human judgment (necessary because taste cannot be reliably LLM-judged).

Overview