22-self-learning

Self-Learning Pattern — Inspired by OpenClaw/Larry

Autonomous self-improvement through conversation analysis, pattern detection, and knowledge accumulation.

Overview

This skill implements the OpenClaw self-learning pattern where the agent continuously learns from interactions, records successful patterns, identifies mistakes, and evolves its own capabilities over time. The key insight: the agent doesn't just follow rules — it creates and refines its own rules based on experience.

The Learning Loop

User Interaction → Pattern Detection → Knowledge Update → Better Response
       ↑                                                          │
       └──────────────────────────────────────────────────────────┘

Core Learning Mechanisms

1. Mistake Detection & Rule Creation

When something goes wrong:

Error/Correction → Root Cause Analysis → New Rule → Store in CLAUDE.md
                                                          │
                                               Hook Creation (if automatable)

Process:

• Detect when user corrects the agent
• Analyze WHY the mistake happened (not just symptoms)
• Create a specific, actionable rule
• Add to Self-Correction Rules in CLAUDE.md with date
• If pattern is code-related, consider creating a hook to auto-enforce

2. Success Pattern Recording

When something works well:

Successful Outcome → Pattern Extraction → Store as Success Pattern
                                                     │
                                          Promote to Default Behavior

Record when:

• A complex task completes without corrections
• User explicitly says "good" or "perfect"
• A pattern is reused successfully 3+ times
• An approach significantly outperforms alternatives

3. Conversation Pattern Analysis

After each significant interaction:

Conversation → Extract Patterns → Update Knowledge
  - What questions were asked?
  - What tools were most useful?
  - What approach worked?
  - What was the user's communication style?
  - What domain knowledge was needed?