# Why I Built Claude Agents Team: Stop Hand-Writing Your .claude Directory

> I kept hand-building agent teams for Claude Code. Claude Agents Team is what I extracted from three of them, with a linter that checks every file.

*2026-10-19T12:49:03Z*

# Why I Built Claude Agents Team: Stop Hand-Writing Your .claude Directory

I run three large agent teams in Claude Code, with about 14, 42 and 40 agents. I built each one by hand, from scratch. Claude Agents Team is what I got when I stopped and wrote down the patterns that worked.

![Why I built Claude Agents Team](/uploads/sites/3b93f650-28ec-4093-99aa-e7b16c2a2e2f/media/claude-agents-team-founder-en.png)

It is an open-source Claude Code plugin (MIT). You give it a product description, and it writes an orchestrator, specialist agents, skills, rules and hooks into `.claude/`. Then it lints every file before the run exits.

## The problem I kept hitting

A good `.claude` directory takes hours. You need an orchestrator, the right specialists, skills, rules and hooks, and every file has to be right: valid frontmatter, a description that tells Claude when to delegate, sensible tools, a fitting model. It is easy to get one of those subtly wrong, and nothing tells you.

Doing that three times made the pattern obvious. The same archetypes kept coming back, so I turned them into templates.

## What I built

The repo's first commit is dated 29 April 2026, and v0.1.0 was tagged the same day. It has three pillars:

- `/team-gen` turns a product brief into a team from a library of 8 agent archetypes, 4 skill archetypes, 13 rule templates and 4 hook templates.
- `/evaluate-agent` scores an agent or skill file against a rubric where each rule cites a source. A static linter is free and deterministic. `--deep` adds an LLM judge.
- `/meet` sends a topic to every agent in parallel and writes the minutes.

The design choice I care about most is that the generator lints its own output. If any file gets a `reject` verdict, the run fails. A team that was never checked is just a pile of prompts.

## Honest limits

- A passing grade means the files meet the rules. It does not prove the team behaves well on your codebase.
- The behavioral harness is a first version: it replays a transcript after the fact.
- `--deep` calls the Anthropic API and needs `ANTHROPIC_API_KEY`.
- The defaults are my opinions: plan first, small wins, no push without permission, a QA agent and a security agent. Delete what you do not want.
- It is not Claude Code's experimental agent-teams feature. It generates the definition files and does not change how Claude Code runs them.

## FAQ

**Who is it for?** Claude Code users starting a new project, and teams that want the same agent setup across repos.

**Is it free?** It is open source under the MIT license.

**What should I tell you if it gets something wrong?** Tell me what your generated team got wrong. That is how the rule set improves.

## Next steps

- Read the full guide: [Claude Code agent team generator](https://agents-team.fadymondy.com/en/blog/claude-code-agent-team-generator)
- Source: [github.com/fadymondy/agents-team](https://github.com/fadymondy/agents-team)
- Product Hunt: [Claude Agents Team](https://www.producthunt.com/products/claude-agents-team)
- Arabic version: [النسخة العربية](https://fadymondy.com/ar/blog/why-i-built-claude-agents-team)

— Fady Mondy

---

Source: https://fadymondy.com/en/blog/why-i-built-claude-agents-team
