Building a Structured, Transparent, and Well-Documented AI Team that Delegates Its Own Tasks
If this project helps you build better AI systems and you'd like to show your appreciation:
- Buy Me a Coffee: https://buymeacoffee.com/mnehmos
- Check out Vario Research: For advanced Deep Research alternatives (Talking about you Gemini, ChatGPT), visit https://mnehmos.github.io/VarioResearch/ for custom reports in any format (I Prefer customized websites so i can add in API functionalities into your dashboards as needed. (Alpha Vantage, Yahoo Finance, Arxiv, google maps, etc))
- Multi-Agent Framework: Specialized modes for different types of tasks
- SPARC Framework: Structured approach to complex problem-solving
- Agentic Boomerang: Reliable task delegation and tracking system
- Structured Documentation: Consistent, traceable documentation
- Token Optimization: Efficient resource usage through the "Scalpel, not Hammer" approach
The system includes the following specialized modes:
- 🪃 Orchestrator: Task decomposition, assignment, and verification
- 💻 Code: Software implementation and optimization
- 🏛️ Architect: System design and pattern application
- ❓ Ask: Information retrieval, evaluation, and communication
- 🪲 Debug: Problem diagnosis and solution validation
- 💾 Memory: Knowledge storage, organization, and retrieval
- 🔍 Deep Research: In-depth investigation and analysis
Below is an architectural overview of how the Roo framework operates:
┌─────────────────────────────────┐
│ VS Code │
│ (Primary Development │
│ Environment) │
└───────────────┬─────────────────┘
│
▼
┌─────────────────────────────────┐
│ Roo Code │
│ ↓ │
│ System Prompt │
│ (Contains SPARC Framework: │
│ • Specification, Pseudocode, │
│ Architecture, Refinement, │
│ Completion methodology │
│ • Advanced reasoning models │
│ • Best practices enforcement │
│ • Memory Bank integration │
│ • Boomerang pattern support) │
└───────────────┬─────────────────┘
│
▼
┌─────────────────────────────────┐ ┌─────────────────────────┐
│ Orchestrator │ │ User │
│ (System Prompt contains: │ │ (Customer with │
│ roles, definitions, │◄─────┤ minimal context) │
│ systems, processes, │ │ │
│ nomenclature, etc.) │ └─────────────────────────┘
└───────────────┬─────────────────┘
│
▼
┌─────────────────────────────────┐
│ Query Processing │
└───────────────┬─────────────────┘
│
▼
┌─────────────────────────────────┐
│ MCP → Reprompt │
│ (Only called on direct │
│ user input) │
└───────────────┬─────────────────┘
│
▼
┌─────────────────────────────────┐
│ Structured Prompt Creation │
│ │
│ Project Prompt Eng. │
│ Project Context │
│ System Prompt │
│ Role Prompt │
└───────────────┬─────────────────┘
│
▼
┌─────────────────────────────────┐
│ Orchestrator │
│ (System Prompt contains: │
│ roles, definitions, │
│ systems, processes, │
│ nomenclature, etc.) │
└───────────────┬─────────────────┘
│
▼
┌─────────────────────────────────┐
│ Subtask Prompt │
│ (Generated by Orchestrator │
│ with structure) │
│ │
│ ┌─────────┐ ┌─────────┐ │
│ │ Topic │ │ Context │ │
│ └─────────┘ └─────────┘ │
│ │
│ ┌─────────┐ ┌─────────┐ │
│ │ Scope │ │ Output │ │
│ └─────────┘ └─────────┘ │
│ │
│ ┌─────────────────────┐ │
│ │ Extras │ │
│ └─────────────────────┘ │
└───────────────┬─────────────────┘
│
▼
┌─────────────────────────────────┐ ┌────────────────────────────────────┐
│ Specialized Modes │ │ MCP Tools │
│ │ │ │
│ ┌────────┐ ┌────────┐ ┌─────┐ │ │ ┌─────────┐ ┌─────────────────┐ │
│ │ Code │ │ Debug │ │ ... │ │──►│ │ Basic │ │ CLI/Shell │ │
│ └────┬───┘ └────┬───┘ └──┬──┘ │ │ │ CRUD │ │ (cmd/PowerShell) │ │
│ │ │ │ │ │ └─────────┘ └─────────────────┘ │
└───────┼──────────┼────────┼────┘ │ │
│ │ │ │ ┌─────────┐ ┌─────────────────┐ │
│ │ │ │ │ API │ │ Browser │ │
│ │ └───────►│ │ Calls │ │ Automation │ │
│ │ │ │ (Alpha │ │ (Playwright) │ │
│ │ │ │ Vantage)│ │ │ │
│ │ │ └─────────┘ └─────────────────┘ │
│ │ │ │
│ └────────────────►│ ┌──────────────────────────────┐ │
│ │ │ LLM Calls │ │
│ │ │ │ │
│ │ │ • Basic Queries │ │
└───────────────────────────►│ │ • Reporter Format │ │
│ │ • Logic MCP Primitives │ │
│ │ • Sequential Thinking │ │
│ └──────────────────────────────┘ │
└────────────────┬─────────────────┬─┘
│ │
▼ │
┌─────────────────────────────────────────────────────────────────┐ │
│ Recursive Loop │ │
│ │ │
│ ┌────────────────────────┐ ┌───────────────────────┐ │ │
│ │ Task Execution │ │ Reporting │ │ │
│ │ │ │ │ │ │
│ │ • Execute assigned task│───►│ • Report work done │ │◄───┘
│ │ • Solve specific issue │ │ • Share issues found │ │
│ │ • Maintain focus │ │ • Provide learnings │ │
│ └────────────────────────┘ └─────────┬─────────────┘ │
│ │ │
│ ▼ │
│ ┌────────────────────────┐ ┌───────────────────────┐ │
│ │ Task Delegation │ │ Deliberation │ │
│ │ │◄───┤ │ │
│ │ • Identify next steps │ │ • Assess progress │ │
│ │ • Assign to best mode │ │ • Integrate learnings │ │
│ │ • Set clear objectives │ │ • Plan next phase │ │
│ └────────────────────────┘ └───────────────────────┘ │
│ │
└────────────────────────────────┬────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ Memory Mode │
│ │
│ ┌────────────────────────┐ ┌───────────────────────┐ │
│ │ Project Archival │ │ SQL Database │ │
│ │ │ │ │ │
│ │ • Create memory folder │───►│ • Store project data │ │
│ │ • Extract key learnings│ │ • Index for retrieval │ │
│ │ • Organize artifacts │ │ • Version tracking │ │
│ └────────────────────────┘ └─────────┬─────────────┘ │
│ │ |
│ ▼ │
│ ┌────────────────────────┐ ┌───────────────────────┐ │
│ │ Memory MCP │ │ RAG System │ │
│ │ │◄───┤ │ │
│ │ • Database writes │ │ • Vector embeddings │ │
│ │ • Data validation │ │ • Semantic indexing │ │
│ │ • Structured storage │ │ • Retrieval functions │ │
│ └─────────────┬──────────┘ └───────────────────────┘ │
│ │ │
└────────────────┼───────────────────────────────────────────────┘
│
└───────────────────────────────────┐
▼
┌─────────────────────────────────┐ ┌─────────────────────────┐
│ Orchestrator │ │ User │
│ (System Prompt contains: │ │ (Customer with │
│ roles, definitions, │◄─────┤ minimal context) │
│ systems, processes, │ │ │
│ nomenclature, etc.) │ └─────────────────────────┘
└───────────────┬─────────────────┘
|
Restart Recursive Loop
- A compatible AI assistant that supports custom modes
- Basic understanding of the SPARC framework concepts
=======
-
Clone this repository:
git clone https://github.com/Mnehmos/The-Ultimate-Roo-Code-Hack-Building-a-Structured-Transparent-and-Well-Documented-AI-Team.git -
Copy the
.roomodesfile to your project root -
Configure your AI assistant to use the custom modes:
- Click the "Modes" button in the Roo sidebar
- Select "Edit Project Modes (.roomodes)"
- Verify that the content matches the
.roomodesfile from this repository - Click "Save"
-
Set up the custom instructions:
- Click the "Modes" button in the Roo sidebar
- Scroll down to "Custom Instructions for All Modes"
- Copy the contents of
templates/custom-instructions-for-all-modes.md - Paste into the Custom Instructions field
- Click "Save"
-
Create the
.roodirectory structure:- Create a
.roodirectory in your project root - Add subdirectories for each mode:
rules-orchestrator,rules-code, etc. - Create a
logsdirectory for activity tracking - Add a
memorydirectory for knowledge storage
- Create a
-
Initialize the boomerang state:
- Create a
.roo/boomerang-state.jsonfile with an empty JSON object:{}
- Create a
If you prefer to set up the framework directly using an AI assistant like Roo, Claude, or ChatGPT, you can copy and paste the following prompt:
I want to set up the Ultimate Roo Code Hack multi-agent framework. Please help me:
1. Create a .roomodes file with configurations for these specialized modes:
- Orchestrator (task management and delegation)
- Code (software implementation)
- Architect (system design)
- Ask (information discovery)
- Debug (problem diagnosis)
- Memory (knowledge management)
- Deep Research (in-depth investigation)
2. Create a meet-the-team.md file that describes each mode as a team member in plain English
3. Set up custom instructions that implement:
- The SPARC framework for structured reasoning
- Boomerang logic for task delegation
- Token optimization strategies
- Standardized documentation formats
Please provide the complete content for each file and instructions on how to use them.
- Click the "Support Prompts" button in the Roo sidebar
- Select "Enhance Prompt"
- Copy the contents of
templates/enhance-prompt-template.md - Paste into the Prompt field
- Click "Save"
This feature helps transform basic prompts into comprehensive, structured project prompts.
- Start with Orchestrator Mode - This is your project manager who will coordinate everything
- Describe your project - Be as detailed as possible in your initial prompt
- Let Orchestrator break it down - It will create subtasks and delegate to specialist modes
- Review the results - Orchestrator will integrate all the pieces and present the final result
- Click on the current mode name in the bottom left corner of the Roo interface
- Select the desired mode from the dropdown menu
- Type your basic prompt in the chat
- Click the ✨ button next to the send button
- Roo will transform your basic prompt into a comprehensive, structured project prompt
When creating tasks for specialist modes, use the standardized task prompt format:
# [Task Title]
## Context
[Background information and relationship to the larger project]
## Scope
[Specific requirements and boundaries for the task]
## Expected Output
[Detailed description of deliverables]
## Additional Resources
[Relevant tips, examples, or reference materials]This structured format ensures that specialist modes have all the information they need to complete tasks effectively and consistently.
The Boomerang Pattern ensures reliable task delegation and tracking:
- Add new modes by updating
config.jsonand.roomodes - Create corresponding rule files in
.roo/rules-{new-mode}/rules.md - Implement mode-specific logging in
.roo/logs/{new-mode}-activity.md - Update memory indexes to accommodate new artifact types
Note: While
.roomodesis used for mode assignments, the.roodirectory structure is still used for keeping notes, logs, and documenting activity and changes.
- Keep context window utilization below 40%
- Start with the least token-intensive cognitive primitives
- Break complex tasks into atomic components
- Use the most specialized mode for each subtask
Contributions are welcome! Please feel free to submit a Pull Request.
This project is licensed under the MIT License - see the LICENSE file for details.
- The SPARC framework developers
- Contributors to the multi-agent AI research community (Roo Code, huge shoutout)
- All users who provide feedback and suggestions