This project addresses the common issues of undisciplined and token-inefficient AI coding assistants by implementing a structured, multi-agent development process. It targets developers seeking more robust and cost-effective AI-assisted coding, offering a framework that mimics real-world software development lifecycles for improved reliability and reduced token consumption.

How It Works

The system employs a three-agent team—Architect, Builder, and Reviewer—each with distinct responsibilities and a defined workflow. This structure, inspired by multi-agent research, ensures a minimum level of review without excessive coordination overhead. The Architect plans and deploys, the Builder executes tasks precisely as briefed, and the Reviewer validates the work before deployment, preventing common AI failure modes like scope drift or unnecessary token usage. Token optimization is further enforced through five core rules embedded in the AI's behavior, prioritizing existing context, eliminating speculative calls, enabling parallelization, routing large outputs, and preventing redundant statements.

Quick Start & Requirements

Installation can be done per-project by cloning the repository directly into the project folder (.claude/skills/three-man-team) or globally (~/.claude/skills/three-man-team). Both methods require running a ./setup script to configure the agents and provide initial commands. The system integrates with AI environments supporting context files, such as Claude Code, VS Code, and Cursor. For enhanced efficiency, the optional RTK tool can be used for bash output compression.

Highlighted Details

• Token Optimization: Implements specific rules to minimize token usage during AI sessions, such as trusting existing skills/memory and avoiding speculative tool calls.
• Structured Workflow: Enforces a strict, step-by-step process: Architect plans -> Builder builds -> Reviewer checks -> Architect deploys.
• Production-Hardened: Developed and refined through production use in a real SaaS platform, ensuring practical applicability.
• Customizable Roles: Agent names (default: Architect, Builder, Reviewer) can be easily customized during the setup process.

Maintenance & Community

The project is maintained by its creator, Russell Aaron, who brings over 20 years of software development experience. It is presented as a continuously improving system, refined from real-world application. No specific community channels (e.g., Discord, Slack) or roadmap links are provided in the README.

Licensing & Compatibility

The project is released under the MIT license, a permissive open-source license allowing for free use, modification, and distribution, including in commercial and closed-source projects.

Limitations & Caveats

The system's effectiveness is dependent on the capabilities and integration points of supported AI environments like Claude Code, VS Code, and Cursor. Users must be comfortable executing shell scripts and interacting with AI via specific prompt structures. While token optimization is a core feature, achieving maximum savings may benefit from the additional setup of the optional RTK tool.