upskill

Why we included this project

Code agents like Claude Code, Open Code, and OpenAI Codex often repeat the same mistakes on tricky tasks. UPskill turns those failures into reusable skills: you describe the task or point it at an agent trace, and a slower, more capable teacher model writes a SKILL.md plus synthetic test cases. The tool then measures how well cheaper student models do with and without that skill, and you can benchmark several models side by side and review past runs. For heavier evaluation, it can offload the work to Hugging Face Jobs. If you want small, fast models to handle harder work without hand-writing prompt libraries, the generate-and-measure loop is worth a look.