July 1, 2026

I had questions about self-improving Skills. How should I judge whether a Skill is better, and when should I roll it back? When a log shows a Skill script being read, does that code go to the model? I was not sure.

I wondered whether a CLI plus a small command map would be easier to maintain than scripts inside Skills. The CLI could change behind a stable command interface.

I noted two possible loops: a post-turn Skill review and a curator lifecycle. I still needed to decide when a finding deserved a lasting Skill, and whether foreground guidance or a post-turn rubric should trigger it.