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.
Three Man Team gives you three agents with distinct jobs, clear handoffs, and rules that prevent the most expensive failure modes. The Architect plans and deploys. The Builder builds exactly what the brief says. The Reviewer doesn't pass work that isn't right.
DeepMind's multi-agent research shows teams of 3-5 with structured artifact handoffs outperform both solo agents and larger groups. Three is not arbitrary — it is the minimum for meaningful review and the maximum before coordination overhead eats the gain.
The roles map to how real software ships:
- Someone who understands the whole system and owns the deploy
- Someone who builds fast and clean
- Someone who catches what the builder missed
Open your terminal and run:
git clone https://github.com/russelleNVy/three-man-team.git ~/.claude/skills/three-man-team
cd ~/.claude/skills/three-man-team && ./setupThen tell Claude Code:
Install three-man-team: add a three-man-team section to CLAUDE.md listing the available agents: /architect, /builder, /reviewer. Set token-optimizer rules as always active.
git clone https://github.com/russelleNVy/three-man-team.git .claude/skills/three-man-team
cd .claude/skills/three-man-team && ./setup# Pick a template — named personas (recommended) or blank slate
cp -r ~/.claude/skills/three-man-team/templates/project-folder/* /path/to/your/project/
# Copy the handoff files
cp ~/.claude/skills/three-man-team/handoff/* /path/to/your/project/Copy this into your Claude Code session:
You are the Architect on this project. Read CLAUDE.md, then ARCHITECT.md.
Report project status in one paragraph, then wait for me.
Every unit of work follows the same path. Architect plans and writes the brief. Builder reads it, shows a plan, builds, and hands off to Reviewer. Reviewer clears it or sends it back. Architect deploys with the Project Owner's go-ahead. Nothing skips a step.
Three agents. Three distinct jobs. Built to work together.
Architect, Builder, Reviewer are the defaults. Rename them to anything — see config/team.yml.example and docs/customizing-your-team.md.
Every session starts with five rules baked into CLAUDE.md:
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.
Install your token-optimizer skill globally at ~/.claude/skills/token-optimizer or
reference it from your project's .claude/skills/ directory.
For bash output compression on top of these rules, see RTK —
a separate tool that compresses find, ls, grep output before it reaches Claude's context.
Not required, but recommended for heavy Claude Code CLI users. The combination of RTK (bash layer)
- token-optimizer (behavior layer) is where real savings compound.
See docs/token-optimization.md for the full discipline.
templates/project-folder/— Start here. Named personas (Arch, Bob, Richard), fully written and ready to use. Customize the Who You Are sections and rename to fit your team.templates/generic/— Blank slate with[CUSTOMIZE]placeholders. Use this if you want to build your own personas from scratch or install globally across all projects.
See docs/customizing-your-team.md for the full walkthrough.
MIT. Free forever.
Russell Aaron — 20+ years building and supporting software the right way. He built this team in production shipping a real SaaS platform. It works because it was used before it was published, fine tuned, and will continue to get better over time as AI models and tools evolve.