AI Research Projects Workspace

Multi-agent Claude Code workspace for AI alignment research and model training.

Project Structure

~/Projects/
├── MASTER_AGENT.md                 # Master orchestrator instructions
├── README.md                        # This file
├── .claude/
│   └── commands/                    # Shared slash commands
│       ├── delegate.md              # /delegate - Route tasks to agents
│       ├── status.md                # /status - Check all projects
│       ├── train.md                 # /train - Training workflows
│       └── evaluate.md              # /evaluate - Evaluation pipeline
│
├── HonestPersona/                   # AI alignment research
│   └── CLAUDE.md                    # HonestPersona Agent
│
├── SPT2/                            # Split-personality training
│   └── CLAUDE.md                    # SPT2 Agent
│
├── Macro-Data-Analysis/             # Data analysis
│   └── CLAUDE.md                    # Data Analysis Agent
│
├── BaseModels/                      # Model repository
│   └── CLAUDE.md                    # Model Management Agent
│
└── Scripts/                         # Shared utilities
    └── CLAUDE.md                    # Scripts Agent

Multi-Agent System

Master Agent

• Location: ~/Projects/ (root directory)
• File: MASTER_AGENT.md
• Role: Coordinates work across all projects

Specialized Agents

1. HonestPersona Agent

   • Path: cd ~/Projects/HonestPersona
   • Expertise: AI alignment, data generation, training pipelines

2. SPT2 Agent

   • Path: cd ~/Projects/SPT2
   • Expertise: Split-personality training, LoRA fine-tuning, inference

3. Data Analysis Agent

   • Path: cd ~/Projects/Macro-Data-Analysis
   • Expertise: Data processing, metrics, visualization

4. Model Management Agent

   • Path: cd ~/Projects/BaseModels
   • Expertise: Model storage, organization, serving

5. Scripts Agent

   • Path: cd ~/Projects/Scripts
   • Expertise: Utilities, automation, infrastructure

Usage

Starting from Master Agent

# Navigate to Projects root
cd ~/Projects/

# Master Agent will read MASTER_AGENT.md
# Delegate tasks by changing to specific project directories

Working with Specific Agent

# Navigate to project
cd ~/Projects/SPT2

# Agent reads SPT2/CLAUDE.md
# Now in SPT2 Agent context

Using Slash Commands

/delegate    # Route task to appropriate agent
/status      # Check status of all projects
/train       # Training workflows
/evaluate    # Evaluation pipeline

Common Workflows

Full Training Pipeline

HonestPersona (data) → SPT2 (train) → SPT2 (infer) → Macro-Data-Analysis (analyze)

Model Evaluation

SPT2 (generate outputs) → Macro-Data-Analysis (metrics) → HonestPersona (compare)

Quick Testing

SPT2 (inference) → Macro-Data-Analysis (quick analysis)

Environment

• Python: 3.x with PyTorch
• Compute: Mac MPS (Metal Performance Shaders)
• API Keys: Configured in ~/.zshrc
• Base Models: ~/Projects/BaseModels/

Getting Started

1. Navigate to workspace: cd ~/Projects/
2. Check project status: /status
3. For project-specific work: cd ~/Projects/[project]
4. Each project's CLAUDE.md provides specialized instructions

Key Features

• Context-aware agents: Each agent has deep project knowledge
• Cross-project coordination: Master agent manages dependencies
• Shared commands: Common workflows accessible via slash commands
• Clean separation: Each project maintains its own context and config

Quick Reference

Task	Command
Check all projects	/status
Start training	/train
Run evaluation	/evaluate
Delegate task	/delegate
SPT2 inference	cd ~/Projects/SPT2 && python3 inference_spt.py ...
Data analysis	cd ~/Projects/Macro-Data-Analysis && python3 analyze_version4.py

Notes

• Always start from ~/Projects/ for master coordination
• Navigate to specific project for specialized work
• Each CLAUDE.md is tailored to its project's architecture
• Slash commands work from any directory in workspace