SkillOpt

Why we included this project

SkillOpt is for teams whose LLM agents keep stumbling on the same kinds of tasks. Instead of retraining or swapping the model, it treats the natural-language skill document as the trainable state of a frozen agent, proposing edits and only keeping the ones that pass a held-out validation gate. The output is a best_skill.md file you can load into the agent's prompt or tooling, so the improvement costs nothing extra at inference time. The workflow will look familiar if you have trained models before: rollouts, reflection, aggregation, and gated updates, all driven from a CLI or a WebUI dashboard. That makes it a good fit for coding agents like Claude Code or Codex, and for custom agentic loops where you want to capture recurring task knowledge once and reuse it across model scales or environments instead of hand-tuning prompts each time.