continuous-learning-v2

Instinct-based pattern learning with confidence scoring and auto-evolution

SKILL.md

GitHub

Continuous Learning v2

Overview

Captures atomic behavioral patterns (instincts) from session activity via hooks. Each instinct has a confidence score (0.3-0.9) that increases with repetition and decays when contradicted.

Commands

    • View learned patterns with confidence scores
    • Export instincts as JSON for team sharing
    • Import instincts from teammate
    • Cluster related instincts into reusable skills/commands

How It Works

Capture Pipeline

Instinct Format

Confidence Scoring

Score Meaning Action
0.3-0.4 Weak signal Observe only
0.5-0.6 Emerging pattern Suggest in context
0.7-0.8 Strong pattern Auto-apply
0.9+ Established behavior Promote to skill

Evolution Pipeline

When 3+ related instincts reach 0.7+ confidence:

  1. Cluster by category/domain
  2. Generate skill draft from instinct descriptions
  3. Present for human review
  4. If approved, create skill .md file in skills/ directory

Stack Context

  • .NET 8: Watch for CosmosDB patterns, GDPR anonymization, French HR terms
  • React 18: Watch for Redux Toolkit patterns, Ant Design usage, French labels
  • Azure: Watch for AKS deployment patterns, Key Vault usage