continuous-learning

Continuous Learning Skill

At the end of a session, automatically evaluate the Claude Code session and extract reusable patterns that can be saved as Learned Skills.

When to Activate

• Set up automatic pattern extraction from Claude Code sessions
• Configure Stop Hook for session evaluation
• View or organize ~/.claude/skills/learned/ Learned Skills
• Adjust extraction thresholds or pattern categories
• Compare the v1 (current) and v2 (Instinct-based) approaches

How It Works

This skill runs as a Stop Hook at the end of each session:

1. Session Evaluation: Check whether the session has enough messages (default: more than 10)
2. Pattern Detection: Identify extractable patterns from the session
3. Skill Extraction: Save useful patterns to ~/.claude/skills/learned/

Configuration

Edit config.json to customize:

{
  "min_session_length": 10,
  "extraction_threshold": "medium",
  "auto_approve": false,
  "learned_skills_path": "~/.claude/skills/learned/",
  "patterns_to_detect": [
    "error_resolution",
    "user_corrections",
    "workarounds",
    "debugging_techniques",
    "project_specific"
  ],
  "ignore_patterns": [
    "simple_typos",
    "one_time_fixes",
    "external_api_issues"
  ]
}

Pattern Types

Pattern	Description
error_resolution	Solutions for specific errors
user_corrections	Patterns from user corrections
workarounds	Solutions for framework/library defects
debugging_techniques	Effective debugging methods
project_specific	Project-specific conventions

Hook Setup

Add to your ~/.claude/settings.json：

{
  "hooks": {
    "Stop": [{
      "matcher": "*",
      "hooks": [{
        "type": "command",
        "command": "~/.claude/skills/continuous-learning/evaluate-session.sh"
      }]
    }]
  }
}

Why Use a Stop Hook?

• Lightweight: Runs only once at the end of a session
• Non-blocking: Does not add latency to every message
• Full context: Can access the complete session record

Related Content

• The Longform Guide - Chapters on continuous learning
• /learn command - Manually extract patterns mid-session

---

Comparison Notes (Research: January 2025)

Comparison with Homunculus

Homunculus v2 uses a more complex approach:

Feature	Our Approach	Homunculus v2
Observation Mechanism	Stop Hook (end of session)	PreToolUse/PostToolUse Hook (100% reliable)
Analysis Method	Main context	Background agent (Haiku agent)
Granularity	Full skills	Atomic "instincts"
Confidence	None	0.3-0.9 weight
Evolution Path	Directly generate skills	Instinct → Clustering → Skill/Command/Agent
Sharing Method	None	Export/import instincts

Key insight from Homunculus:

"v1 relies on Skills for observation. Skills are probabilistic—they trigger about 50-80% of the time. v2 uses Hooks for observation (100% reliable) and treats 'Instincts' as the atomic unit of learned behavior."

Potential v2 enhancements

1. Intuition-based learning - Smaller, atomic behaviors with confidence scoring
2. Background observer - Haiku agent analyzing in parallel
3. Confidence decay - Intuition confidence decreases if conflicts occur
4. Domain tags - Code style, testing, Git, debugging, etc.
5. Evolution path - Cluster related intuitions into skills/commands

See:/Users/affoon/Documents/tasks/12-continuous-learning-v2.md for the full specification.