RLM Core Instructions v4

Recursive Learning Model (RLM) + claude-flow integration for Claude Code

A structured workflow system that enforces complexity-based task decomposition, quality gates, and learning patterns for Claude Code sessions.

Features

• Complexity-based routing: LOW/MEDIUM/HIGH task classification
• Phase pipeline: Structured execution (Phase 0-4)
• Quality Gates: Automated verification checkpoints
• claude-flow MCP integration: Memory, trajectory tracking, pattern learning
• Metacognition: Self-reflection and adaptive learning
• Stuck recovery: Automatic detection and escalation

Quick Start

Option 1: CLAUDE.md (Recommended)

Copy CLAUDE.md to your home directory:

# Global (all projects)
cp CLAUDE.md ~/.claude/CLAUDE.md

# Or project-specific
cp CLAUDE.md ./CLAUDE.md

Option 2: UserPromptSubmit Hook (Auto-inject every prompt)

Windows:

.\install.ps1

Linux/Mac:

chmod +x install.sh && ./install.sh

What It Does

Complexity Classification

Level	Criteria	Required Phases
LOW	Info lookup, 1-2 line changes	None (direct execution)
MEDIUM	1-3 files, bug fixes	Phase 1, 2, 3, 4
HIGH	5+ files, new features	Phase 0, 1, 2, 3, 4 + Plan Mode

Phase Pipeline

Phase 0: Initialize (HIGH only)
    └── Agent routing, trajectory start

Phase 1: PDE Q&A
    └── Expert consultation, requirement gathering

Phase 2: Block Decomposition
    └── 5-10 blocks, max 30min each, workflow_create

Phase 3: Recursive Execution
    └── Per-block implementation + reflection (max depth: 3)

Phase 4: Verification
    └── Quality Gates, Playwright E2E, pattern storage

Quality Gates

Gate	Threshold
Test Coverage	>= 80%
TypeScript Errors	0
Lint Errors	0
Build	Pass
Playwright E2E	Pass
Console Errors	0

File Structure

~/.claude/
├── CLAUDE.md                    # Core Instructions (auto-loaded)
└── hooks/
    ├── inject-system-prompt.ps1 # Windows hook
    ├── inject-system-prompt.sh  # Linux/Mac hook
    └── system-prompt.md         # Instructions content

Requirements

• Claude Code 2.0+
• (Optional) claude-flow MCP for advanced features

Configuration

Add to ~/.claude/settings.json:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "powershell -NoProfile -ExecutionPolicy Bypass -File ~/.claude/hooks/inject-system-prompt.ps1"
          }
        ]
      }
    ]
  }
}

License

MIT

Contributing

PRs welcome! Please test your changes before submitting.

Credits

Inspired by the Claude Code community and Anthropic's best practices.