Skill Optimizer

If your skill has accumulated a dozen or more learn-rule corrections and is starting to feel bloated with overlapping guidance, this runs an offline training loop that treats those corrections like labeled data. It mirrors Microsoft SkillOpt's six-stage pipeline: rollout from SQLite, reflect via an optimizer LLM, aggregate patches, select within a learning-rate budget, update a candidate skill, and gate acceptance against a held-out validation set. Only candidates that beat the current score ship. Default budget is fifty cents, three epochs, and it writes the improved skill back to disk with full provenance in the database. It's the closest thing to gradient descent you can do on a markdown file without retraining the model itself.

skills.sh installs: 140
GitHub stars: 2.9K
last updated: Jun 8, 2026

View on GitHub

Install

$ npx -y skills add rohitg00/pro-workflow --skill skill-optimizer --agent claude-code

Installs into .claude/skills of the current project.

Files

Skill Optimizer

Train an existing SKILL.md the way a deep-learning optimizer trains weights: via rollouts, gradient-like reflections, validation-gated acceptance. No model retraining; only the skill markdown changes.