What problem does it solve? Improving an AI agent skill's quality is usually guesswork: you edit the prompt, eyeball a few outputs, and hope it got better. This Skill treats SKILL.md as trainable parameters of a frozen agent and runs a benchmark-driven optimization loop with mandatory validation gating, so only edits that measurably improve scores are committed. ## Core Features & Use Cases - Validation-gated optimization loop: Runs forward rollouts, reflection-based edits, and a median-of-3 plus epsilon=0.05 validation gate before any SKILL.md rewrite is accepted, with atomic versioned snapshots and an audit trail. - Benchmark bootstrapping: Generates a starter benchmark of ~15 tasks with rule judges directly from a SKILL.md via --bootstrap-from-skill, requiring human review and strengthening before optimization. - Safety guardrails for bundled skills: Refuses to mutate shipped skills in place unless both --allow-mutate-bundled and an independent held-out set are provided; otherwise writes a proposed.md for review. - Use Case: You authored a meeting-notes skill and want it to produce tighter agendas. Run the bootstrap, strengthen the generated judges, then launch the optimizer with a 1:1:1 split to iteratively improve the skill body against your benchmark. ## Quick Start Ask the agent to run skillopt on your skill, for example: optimize my meeting-notes skill against its benchmark using gbrain skillopt with a dry run first to preview cost.
