Self-learning from past tasks

[NVolcz]

As a user, I find it hard to define agent skills and keep AGENTS.md up to date upfront. I also often forget to go back and review past conversations, even when they contain useful patterns or corrections.

A command like /learn or /reflect could help by reviewing recent task conversations and extracting small, reusable learnings over time. This idea is directly inspired by https://github.com/scottfalconer/self-learning-skills.

This would fit naturally alongside the existing /init flow. Instead of capturing context only at the start, this command could be run during development to reflect on recent work and update AGENTS.md or create or refine agent skills based on what actually worked.

If this proves useful, it could later be extended to run automatically (on a schedule or after successful tasks).