microsoft/skillopt ? reverse-engineered prompt

Build me a Python tool that helps me improve the instructions my AI coding agent follows, without changing the model itself. I want to be able to run a training loop over past tasks, try small edits to a skill document, test those edits on held out examples, and only keep changes that clearly make things better. The end result should be a reusable best_skill.md file that I can drop into an agent setup and get better results right away.

Please include a simple web dashboard so I can watch runs, scores, and comparisons, and make it easy to run from the command line too. It should work with a few different agent backends and let me plug in new tasks or benchmarks later. If you need to look up current docs online, do that. Keep the defaults safe and simple, but make it feel like a real self improving workflow for local coding agents.