SkillOpt-style training loop for agent content skills — bounded add/delete/replace edits, validation gate, rejected-edit buffer. Optimize writing/edit-planning skills without changing model weights.