Continuous Learning v2 - Instinct System

Overview

Instincts are micro-patterns learned from coding sessions. Unlike full skills, instincts are:

•Lightweight - Single pattern, minimal context
•Confidence-scored - Track success rate
•Evolvable - Cluster into skills when mature

Instinct Structure

json

{
  "id": "inst_abc123",
  "pattern": "When encountering CORS errors in Next.js API routes...",
  "solution": "Add headers object with Access-Control-Allow-* fields",
  "confidence": 0.85,
  "uses": 12,
  "successes": 10,
  "created": "2026-01-15T10:30:00Z",
  "lastUsed": "2026-01-28T14:20:00Z",
  "tags": ["nextjs", "api", "cors", "debugging"],
  "context": {
    "framework": "next.js",
    "version": "15.x",
    "category": "debugging"
  }
}

Commands

`/instinct-status`

View learned instincts with confidence scores:

code

📊 Instinct Status

Total: 47 instincts | Avg Confidence: 0.78

High Confidence (>0.8):
  ✅ CORS handling in Next.js API [0.92] - 15 uses
  ✅ Prisma transaction patterns [0.88] - 8 uses
  ✅ React useEffect cleanup [0.85] - 23 uses

Medium Confidence (0.5-0.8):
  ⚡ Supabase RLS policies [0.72] - 5 uses
  ⚡ Tailwind responsive patterns [0.68] - 7 uses

Low Confidence (<0.5):
  ❓ Edge function cold starts [0.45] - 2 uses

`/instinct-export`

Export instincts for sharing:

bash

# Export all instincts
/instinct-export

# Export by tag
/instinct-export --tags=nextjs,react

# Export high confidence only
/instinct-export --min-confidence=0.8

Output: .ai_state/instincts/export-2026-01-28.json

`/instinct-import <file>`

Import instincts from team:

bash

/instinct-import shared-instincts.json

Imported instincts start with 0.5 confidence and adjust based on local use.

`/evolve`

Cluster related instincts into a skill:

bash

# Interactive evolution
/evolve

# Target specific tags
/evolve --tags=authentication

Learning Triggers

Instincts are captured when:

•Successful Debugging - Problem → Solution pattern
•Pattern Recognition - Repeated similar solutions
•User Confirmation - Explicit "remember this" signals
•Code Review Feedback - Accepted improvements

Confidence Scoring

code

confidence = successes / uses * decay_factor

where:
  decay_factor = 0.95^(days_since_last_use / 30)

Confidence increases with successful uses, decreases with failures or time.

Storage

code

.ai_state/instincts/
├── instincts.json      # Main instinct database
├── index.md            # Human-readable index
├── exports/            # Export history
│   └── export-*.json
└── evolved/            # Evolved skills
    └── skill-*.md

Evolution Process

When instincts cluster around a topic:

code

1. Identify Related Instincts
   - Same tags (>3 instincts)
   - Similar patterns
   - High combined confidence

2. Generate Skill Draft
   - Merge patterns
   - Synthesize solutions
   - Create SKILL.md

3. User Review
   - Present draft
   - Cunzhi confirmation
   - Install or iterate

4. Deprecate Instincts
   - Mark as "evolved"
   - Link to new skill

Integration

With Experience Skill

code

Instincts → lightweight, auto-captured
Experience → heavyweight, manually curated

Workflow:
1. Instinct captured automatically
2. High-confidence instincts → candidate for experience
3. User confirms → promote to experience

With /learn Command

code

/learn               # Capture to instincts (default)
/learn --experience  # Capture to experience (manual)
/learn --dry-run     # Preview without saving

Quality Filters

Instincts must meet criteria:

•Non-trivial - Not basic language features
•Reusable - Applies to multiple contexts
•Specific - Clear trigger and solution
•Tested - At least one successful use

Team Sharing Best Practices

bash

# Export team-relevant instincts
/instinct-export --min-confidence=0.7 --tags=our-stack

# Import with namespace
/instinct-import team-patterns.json --namespace=team

# Review imported before trusting
/instinct-status --namespace=team