@patrickc wild seeing this today as i had the exact same thought (probably a lot of people are having this thought at the same time), and i built something this week to do this for my own workflow which i call upskill.
it's a file-based system for recursive, self-improving agent skills (inspired by microsoft’s SkillOpt paper, karpathy’s LLM-wiki, and garry tan’s gstack line of work).
core idea: a skill whose job is to recursively improve skills (including itself) automatically every session. it reads the run, runs the retro, and folds improvements back into markdown files in a git-native, repeatable, low-cognitive-load way.
it auto-loads input files context, stores & evolves prompts/workflows, and handles VCS/snapshots shareable compiled outputs (and works natively as a skill in any agent harness).
gonna write it up properly in a post, but in meantime, demo repo here:
x.com/yungbose/status/206342…
keen for your take - it's been a key lego brick for me, and feels like it aligns with your Autotools × Notion vision 🫰🏽
spent the week developing and testing a system for self-improving agent skills, inspired by microsoft's SkillOpt paper, karpathy's LLM-wiki, and garry tan's gstack line of work.
core idea: a skill whose job is to recursively improve skills (including itself), automatically, every session they are run. it reads the session, runs the retro, and folds in the improvements in a repeatable, consistent, and low-cognitive-load way.
writing it up properly tomorrow; in the meantime, the repo's in the comments; keen for feedback from anyone building with agent skills 🫰🏽