continuous-learning-v2

Capture Claude Code session activity by enabling PreToolUse and PostToolUse hooks pointing to an observe.sh hook, which records prompts, tool calls, and results for deterministic observation of coding preferences.

Atomic confidence-weighted instincts are minimal trigger-action behaviors with assigned confidence scores, designed to replace large brittle skills with small, reliable automated actions applied during interactions.

Automatically detect recurring debugging patterns by using a background observer to analyze logged PreToolUse and PostToolUse events, converting repeated workflows into atomic instincts with provenance for later application.

Yes, continuous learning requires configuring PreToolUse and PostToolUse hooks directed at an observe.sh script to capture session events and generate instincts for review.

Manage and share evolved micro-skills by clustering related instincts into full skills, commands, or agents, and using export and import workflows for sharing and review.

Atomic instincts solve the problem of noisy, unreliable end-of-session skill extraction by creating precise, minimal learning units that can be automatically applied with confidence scores for reliable behavior learning.