continuous-learning-v2

You can automatically save Claude Code session patterns by enabling deterministic PreToolUse and PostToolUse hooks that capture session activity and convert it into confidence-scored reusable instincts. This prevents wasting time re-teaching Claude Code the same workflows.

Project-scoped instinct learning isolates instinct storage per detected repository, preventing cross-project contamination while still supporting global instincts for universal best practices. It ensures coding style and file-structure conventions remain specific to their intended context.

You can promote project-specific coding conventions to global skills by using cross-project promotion criteria that elevate high-confidence instincts when they recur across multiple repositories. This evolves repeated atomic instincts into higher-level skills and commands.

Yes, you need git, python3, bash, and claude installed to enable deterministic hook-based observation capture. These dependencies support scrubbing local observations and extracting workflow patterns from your session activity.

Confidence scoring works by creating atomic instincts from observed tool-usage evidence and clustering them into higher-level skills. Instincts gain confidence as evidence accumulates, eventually meeting promotion criteria to evolve into global commands or agents.

Workflow conventions fail to persist when deterministic observation hooks are not enabled in your Claude settings. Enabling hook-based capture ensures session activity is reliably recorded and converted into confidence-scored, project-scoped instincts for future use.