continuous-learning-v2

Hook-based observation captures every tool call in Claude Code sessions to build a dataset of user behaviors, which background Haiku analysis then processes to extract reusable coding conventions automatically.

Project-scoped isolation stores coding instincts within specific git repositories to prevent cross-project contamination, keeping React patterns in React projects and Python conventions in Python projects while sharing universal practices globally.

Yes, continuous learning hooks work across multiple git repositories, capturing project-specific conventions for solo developers and teams while maintaining isolated storage to prevent cross-project pattern contamination.

Confidence scoring weights extracted behavioral patterns by frequency, allowing high-confidence instincts to be promoted from project scope to global scope or evolved into full skills, commands, and specialized agents.

Automating extraction with atomic instinct patterns and confidence scoring converts repeated coding corrections into reusable instincts by observing tool use and detecting behavioral patterns without requiring manual documentation.

No, manual documentation is unnecessary because the system automatically observes tool use during coding sessions, detects behavioral patterns, and converts them into confidence-scored instincts stored with project-scoped isolation.