AI Development Framework

Three Man Team

Structure for teams that ship.

By: Russell Aaron

See how it works ↓ View on GitHub →
v1.2.3 813 Stars MIT License Built in production
The Problem

AI tools are powerful.
And undisciplined.

AI coding tools are powerful but undisciplined. They read entire codebases when they need one function. They add features nobody asked for. They drift mid-task. They burn tokens on every session doing work that didn't need to happen.

"The solution isn't a better prompt. It's a process."

The Research

Three is not arbitrary.

Solo agents drift — there's no one to catch a wrong turn. Large teams generate coordination overhead that eats the productivity gain. Three is the minimum for meaningful review and the maximum before the team starts managing itself instead of the work.

Role 01
Architect
Owns the plan and the brief. Nothing gets built without a written specification. The single fixed point everything orients around.
Role 02
Builder
Reads the brief. Confirms the plan. Builds exactly what was specified — no more, no less. Hands off clean.
Role 03
Reviewer
Reviews against the brief, correctness, and standards. Will not pass work that isn't right. Nothing ships without a review.
The Sprint

Every step. Every time.

💬
Plan
Architect
Project Owner
📋
Brief
Architect
Architect
⚙️
Build
Builder
Builder
🔍
Review
Reviewer
Reviewer
🚀
Deploy
Architect
Architect
📝
Log
Build Log
Architect
HANDOFF 01

Architect plans the work and writes the brief. The brief is the contract — specific, scoped, unambiguous. Builder doesn't start without it.

HANDOFF 02

Builder reads the brief, posts a plan, and builds. When done, Builder writes the review request and hands off to Richard. Clean, documented, traceable.

HANDOFF 03

Reviewer clears it or sends it back with findings. Nothing skips this step. When Richard signs off, Architect deploys.

How the team runs

One Claude Code session. Three roles. You don't open three windows. Arch runs in your main session. When work is ready, Arch spins up Bob and Richard as subagents via Claude Code's Agent tool. All three roles happen in one place — handoff files on disk are how they communicate.

The Team

Three roles. One process.

Arch
Architect
Arch
Plans the work, writes the brief, owns the deploy. The fixed point everything orients around.
Bob
Builder
Bob
Builds exactly what the brief says. Fast, precise, no ego. Hands it to Richard clean.
Richard
Reviewer
Richard
Quiet until he speaks. Will not pass work that isn't right. Has seen what happens when corners get cut.
Token Efficiency

Less noise. More signal.

Three Man Team bakes token discipline into every role. Each agent follows five rules that kill waste before it compounds across a session. The token-optimizer skill ships with every install and auto-loads via CLAUDE.md — no manual setup required.

Is this in a skill or memory? → Trust it. Skip the file read. Is this speculative? → Kill the tool call. Can calls run in parallel? → Parallelize them. Output > 20 lines you won't use? → Route to subagent. About to restate what user said? → Delete it.

For bash output compression, RTK is a separate tool that compresses find, ls, grep output before it reaches Claude's context. Learn more at github.com/rtk-ai/rtk.

Goes Well With

Not required. Just good.

ccusage ↗
Tracks your Claude Code token usage and costs by session, day, or billing window. Know exactly what you're spending.
claude-historian-mcp ↗
An MCP server that lets Claude search your past conversation history. Find solutions you've already solved without starting from scratch.

Three Man Team works without either. These just make the experience better.

Quick Start

Up and running in two steps.

Per-project install
bash 
cd your-project
git clone https://github.com/russelleNVy/three-man-team .claude/skills/three-man-team
cd .claude/skills/three-man-team && ./setup
Global install
bash 
git clone https://github.com/russelleNVy/three-man-team ~/.claude/skills/three-man-team
cd ~/.claude/skills/three-man-team && ./setup
Clone on GitHub →

Arch checks for new releases automatically at the start of every session. See releases →