continuous-learning-v2_skill

Overview

This skill implements an instinct-based continuous learning system that watches Claude Code sessions via deterministic hooks and converts observed patterns into atomic "instincts" with confidence scores. It stores observations locally, clusters repeated behaviors, and can evolve those instincts into reusable skills, commands, or agents. The design focuses on high-reliability observation, evidence-backed rules, and privacy by keeping raw data on-device.

How this skill works

Hooks capture every PreToolUse and PostToolUse event and append structured observations to a local JSONL store. A background observer agent reads those observations, detects patterns (user corrections, error resolutions, repeated workflows), and creates or updates atomic instincts that carry domain tags, evidence, and a confidence score. Instincts can be inspected, exported/imported, or clustered into higher-level evolved artifacts on demand.

When to use it

• Automatically learn and codify recurring preferences from interactive sessions
• Capture tool usage and user corrections deterministically for pattern detection
• Tune confidence thresholds to control auto-application vs. suggestion behavior
• Export or share distilled behavior rules without exposing raw session data
• Evolve a set of related micro-behaviors into a formal skill, command, or agent

Best practices

• Enable PreToolUse/PostToolUse hooks to guarantee comprehensive observation
• Set a conservative min_confidence (e.g., 0.3) and higher auto_approve_threshold (e.g., 0.7)
• Review new instincts regularly before trusting auto-applied behavior
• Keep observations local and export only distilled instincts for sharing
• Use domain tags (code-style, testing, workflow) to make clustering more reliable

Example use cases

• Learn a developer's preference for functional patterns and suggest them in new fixes
• Detect a recurring testing workflow and evolve it into a reusable testing skill
• Auto-capture and reinforce preferred git or review practices across sessions
• Turn repeated error-resolution steps into a command that automates the fix
• Import a colleague's non-sensitive instincts to align team workflows