continuous-learning-v2

To capture Claude Code sessions and turn them into reusable skills, you enable observer hooks that log session activity into atomic instincts with confidence scores. These instincts evolve into structured skills, commands, and agents through an automated workflow.

Atomic instincts in continuous learning are granular behavioral patterns captured from session hooks. Each instinct receives a confidence score that determines its reliability, guiding its evolution into mature skills, commands, and agents.

You start evolving session patterns by enabling the observer hooks to capture activity. The system generates atomic instincts with confidence scores that progressively evolve into commands and agents using a frontmatter-driven skill manifest.

Project-scoped learning keeps captured instincts isolated per project by default, preventing cross-project contamination. Global instincts can still be enabled when appropriate to allow controlled cross-project learning and pattern sharing.

Yes, a frontmatter-driven skill manifest manages evolving instincts by structuring their metadata throughout the end-to-end workflow. This supports the transition from raw session observations into fully developed skills and agents.