Optimize frozen LLM agents for HotpotQA with reusable natural language skills trained through iterative validation cycles.

SkillOpt replaces manual prompt tweaking with a mathematical feedback loop. A base model runs your tasks against a benchmark dataset while an optimizer model evaluates failures and rewrites the `## Instructions` section — automatically rejecting changes that don't improve the score.