CodebuffAI/evalbuff ? reverse-engineered prompt

Build me a tool that helps improve a coding agent by practicing on a codebase offline.

I want it to take a local repo, pick a set of features, remove them in a controlled way, then have an agent try to rebuild them from scratch. After each attempt, it should judge how well the agent did, collect the learnings, and then update the project docs only if those changes actually make the agent perform better on repeat runs.

The docs should live in a nested markdown structure, with a simple top level guide that tells the agent which docs to read. I also want a way to run the whole loop from one command, with options for the repo path, how many features to test, and setup commands like install steps. Save run outputs somewhere temporary so I can inspect them later.

If you need to look up current docs or agent best practices online, go ahead.