What problem does it solve?
Manually capturing coding patterns and workflow preferences across Claude Code sessions is tedious and error-prone, while project-specific conventions often get lost when switching contexts or contaminated when applied globally without isolation.
Core Features & Use Cases
- Hook-Based Observation: Captures 100% of tool use events via PreToolUse and PostToolUse hooks, ensuring no patterns are missed due to probabilistic skill firing.
- Atomic Instincts with Confidence Scoring: Learns small, specific behaviors with confidence weights (0.3-0.9) that evolve based on observation frequency and user corrections.
- Project-Scoped Learning: Automatically isolates instincts per project using git remote hashes, preventing React patterns from leaking into Python projects while promoting universal patterns to global scope.
- Evolution Pipeline: Clusters related instincts into full skills, commands, and agents through the
/evolve command.
Quick Start
Use the continuous-learning-v2 skill to automatically observe your Claude Code sessions, learn your coding patterns with confidence scoring, and evolve them into project-specific skills and global best practices.