chemistrywow31/Continuous Learning V2

Continuous Learning V2

Description

Capture team design patterns from A-Team sessions using hook-based observation. Detected patterns become "instincts" — atomic learned behaviors with confidence scoring that improve future team designs.

Users

Infrastructure-level skill. Operates via hooks in settings.json and benefits all agents automatically. No explicit agent invocation required.

Core Concept

An instinct is a small learned behavior:

---
id: always-propose-qa-role
trigger: "when designing a team that produces deliverables"
confidence: 0.9
domain: "role-design"
---

# Always Propose QA Role

## Action
Proactively propose a quality checker/reviewer role during requirements exploration.

## Evidence
- User added QA role in 3/4 recent designs after initial omission

Properties:

• Atomic — one trigger, one action
• Confidence-weighted — 0.3 tentative, 0.7 auto-applied, 0.9 core behavior
• Domain-tagged — role-design, skill-planning, model-selection, workflow
• Evidence-backed — tracks observations that created it

How It Works

Team Design Session
      │
      │ Hooks capture tool use + user feedback (100% reliable)
      ▼
  observations.jsonl
      │
      │ Observer agent (background, Haiku)
      ▼
  PATTERN DETECTION
  • User corrections → instinct
  • Repeated structures → instinct
  • Successful designs → instinct
      │
      ▼
  instincts/personal/
  • always-propose-qa-role.md (0.9)
  • writers-use-sonnet.md (0.7)
  • check-tech-feasibility.md (0.8)

Team Design Patterns to Detect

Role Structure Patterns

• Roles users consistently add or request (QA, tech feasibility)
• Role granularity preferences per team complexity
• Coordinator scope patterns across teams

User Correction Patterns

• Roles the user added after initial design
• Model selection corrections (e.g., writer agents downgraded from opus)
• Skill/rule additions the user requested

Workflow Patterns

• Common phase structures across designed teams
• Typical agent/skill/rule counts per team size
• Deployment mode preferences (subagent vs Agent Teams)

Quality Patterns

• Cross-validation issues that recur across teams
• Common reference consistency problems
• Naming convention preferences

Hook Configuration

Add to ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "*",
      "hooks": [{
        "type": "command",
        "command": "~/.claude/skills/continuous-learning-v2/hooks/observe.sh pre"
      }]
    }],
    "PostToolUse": [{
      "matcher": "*",
      "hooks": [{
        "type": "command",
        "command": "~/.claude/skills/continuous-learning-v2/hooks/observe.sh post"
      }]
    }]
  }
}

Commands

| Command | Description | |---------|-------------| | /instinct-status | Show learned team design instincts with confidence | | /evolve | Cluster related instincts into reusable skills | | /instinct-export | Export instincts for sharing | | /instinct-import | Import instincts from other A-Team instances |

Integration with A-Team Memory

Instincts complement the auto memory system (MEMORY.md):

• MEMORY.md: Manually curated, high-confidence project-level patterns
• Instincts: Automatically detected, confidence-scored session-level patterns
• Promote instincts at 0.9+ confidence to MEMORY.md for persistence

Example

Input

After 5 team design sessions, observer detects: in 4/5 sessions, the user asked to add a "technical feasibility assessor" role that was not in the initial design.

Output

---
id: propose-tech-feasibility-role
trigger: "when designing a technology-oriented team"
confidence: 0.8
domain: "role-design"
---

# Propose Technical Feasibility Role

## Action
When the team's output involves technology, proactively ask the user if a technical feasibility assessment role is needed.

## Evidence
- User added tech feasibility role in 4/5 recent tech team designs
- Pattern first observed: 2025-01-20