Metaskill

One skill to create them all.

Metaskill is a Claude Code skill that creates AI agent teams, individual agents, and custom skills — all through a single /metaskill command.

# Generate a full agent team for a project
/metaskill ios app with SwiftUI
/metaskill fullstack web app with React and PostgreSQL
/metaskill data science pipeline with PyTorch
/metaskill game dev with Unity and C#

# Create a single agent
/metaskill a security reviewer agent for Go microservices
/metaskill a code reviewer agent

# Create a single skill
/metaskill a deploy-to-staging skill with Docker
/metaskill a lint-and-format skill

One command. It detects your intent and routes to the right flow automatically.

---

What It Does

Team Generation (default)

When you describe a project type, Metaskill runs a 4-phase process:

Phase 1: RESEARCH
  ├── Web search: real-world team structures for your domain
  ├── Web search: existing Claude Code agent configs on GitHub
  ├── Web search: MCP servers relevant to your tech stack
  └── Web search: best practices, linters, testing frameworks

Phase 2: BUILD
  ├── CLAUDE.md          — routing table + orchestration protocol
  ├── .claude/agents/    — 4-6 agents (tech-lead + specialists + reviewer)
  ├── .claude/skills/    — 2-4 workflow automation skills
  ├── .claude/rules/     — coding standards for the primary language
  └── .mcp.json          — MCP server configuration

Phase 3: CREDENTIALS
  └── Asks for any API keys/tokens needed by MCP servers
      (always offers "skip / configure later")

Phase 4: VERIFY
  └── Lists all created files, validates .mcp.json, prints summary

The research phase is what makes each generated team genuinely useful — it's not a fixed template. It adapts to what actually matters in your domain right now.

Single Agent Creation

When your request mentions "agent", "reviewer", or a specific role, Metaskill creates a single .claude/agents/<name>.md with:

• Expert persona design
• Frontmatter configuration (model, tools, permissions)
• System prompt with self-verification and Workflow Discipline built in

Single Skill Creation

When your request mentions "skill", "command", or "slash command", Metaskill creates a single .claude/skills/<name>/SKILL.md with:

• Scope and invocation model configuration
• Dynamic context injection
• Argument handling and tool access

---

What Gets Generated (Team Mode)

ios-app-agents/
├── CLAUDE.md                     ← routing table + orchestration protocol
├── .mcp.json                     ← MCP servers, auto-discovered by Claude Code
└── .claude/
    ├── agents/
    │   ├── tech-lead.md          ← Opus model, routes all tasks
    │   ├── ios-engineer.md       ← SwiftUI, platform APIs
    │   ├── ui-designer.md        ← layouts, animations, design system
    │   ├── test-engineer.md      ← XCTest, UI testing
    │   └── code-reviewer.md      ← quality gate for all code changes
    ├── skills/
    │   ├── build-and-test/SKILL.md
    │   └── run-simulator/SKILL.md
    └── rules/
        └── swift-standards.md

After generation:

cd ios-app-agents
claude
> "Build a SwiftUI todo app with iCloud sync"

The tech-lead agent breaks the task down and delegates to specialists. You supervise, they build.

---

Install

curl -fsSL https://raw.githubusercontent.com/xvirobotics/metaskill/main/install.sh | bash

This installs /metaskill to ~/.claude/skills/metaskill/.

Or manually:

mkdir -p ~/.claude/skills/metaskill/flows

curl -fsSL https://raw.githubusercontent.com/xvirobotics/metaskill/main/skill/SKILL.md \
  -o ~/.claude/skills/metaskill/SKILL.md

curl -fsSL https://raw.githubusercontent.com/xvirobotics/metaskill/main/skill/flows/team.md \
  -o ~/.claude/skills/metaskill/flows/team.md

curl -fsSL https://raw.githubusercontent.com/xvirobotics/metaskill/main/skill/flows/agent.md \
  -o ~/.claude/skills/metaskill/flows/agent.md

curl -fsSL https://raw.githubusercontent.com/xvirobotics/metaskill/main/skill/flows/skill.md \
  -o ~/.claude/skills/metaskill/flows/skill.md

Requirements: Claude Code CLI installed and authenticated.

---

The Idea

Modern AI coding agents are powerful, but setting up a well-structured multi-agent team takes hours: designing roles, writing system prompts, choosing MCP servers, configuring routing. Metaskill automates that entire setup step.

The name is a reference to the philosophical concept of a meta-skill — a skill that makes you better at acquiring other skills. In this case: a skill that generates the skills (and agents, and rules) you need to work on any kind of project.

One prompt. One team. Start building.

---

Agent Architecture

Every generated team follows this pattern:

| Role | Model | Responsibility | |------|-------|---------------| | tech-lead | Opus | Routes tasks, coordinates agents, never implements directly | | <specialist-1> | Sonnet | Domain expert #1 (e.g. frontend, iOS, data engineering) | | <specialist-2> | Sonnet | Domain expert #2 | | <specialist-3> | Sonnet | Domain expert #3 (optional) | | code-reviewer | Sonnet | Quality gate — all code passes through here |

Orchestration protocol (in every generated CLAUDE.md):

• Tech-lead is the routing authority
• Main Claude never implements directly for multi-step tasks — it delegates
• Structured handoff documents between agents
• Code reviewer is a mandatory quality gate
• All agents follow built-in Workflow Discipline: plan-first, re-plan on failure, verify before done, autonomous execution

---

MCP Server Catalog

Metaskill selects MCP servers based on your domain. Verified catalog:

| Server | Package | Purpose | Transport | |--------|---------|---------|-----------| | context7 | @upstash/context7-mcp@latest | Up-to-date library docs | stdio | | playwright | @playwright/mcp@latest | Browser automation & e2e testing | stdio | | filesystem | @modelcontextprotocol/server-filesystem | Enhanced file operations | stdio | | postgres | @modelcontextprotocol/server-postgres | Database queries | stdio | | sequential-thinking | @modelcontextprotocol/server-sequential-thinking | Structured multi-step reasoning | stdio | | memory | @modelcontextprotocol/server-memory | Persistent knowledge graph | stdio | | github | https://api.githubcopilot.com/mcp/ | GitHub API access | HTTP |

All servers are written to .mcp.json and auto-discovered by Claude Code on launch. No claude mcp add needed.

---

Examples

The examples/ directory contains real, usable agent teams generated by Metaskill. Each is a complete project you can copy and use directly:

| Example | Stack | Agents | Skills | |---------|-------|--------|--------| | fullstack-web/ | React + Node.js + PostgreSQL | tech-lead, frontend-engineer, backend-engineer, devops-engineer, code-reviewer | build-and-test, deploy-preview, api-test | | ios-app/ | SwiftUI + Swift | tech-lead, ios-engineer, ui-designer, test-engineer, code-reviewer | build-and-test, run-simulator | | data-science/ | Python + PyTorch | tech-lead, data-engineer, ml-engineer, analyst, code-reviewer | run-pipeline, evaluate-model, generate-report |

Try one:

cp -r examples/fullstack-web my-project
cd my-project && claude

---

Project Structure

metaskill/
├── skill/
│   ├── SKILL.md              ← /metaskill entry point (intent routing)
│   └── flows/
│       ├── team.md           ← full agent team generation flow
│       ├── agent.md          ← single agent creation flow
│       └── skill.md          ← single skill creation flow
├── install.sh                ← installer
└── examples/                 ← example generated outputs

---

Contributing

PRs welcome. The skill is modular — edit the specific flow file for what you want to improve:

1. Fork the repo
2. Edit the relevant file in skill/flows/
3. Test in Claude Code with /metaskill
4. Submit a PR with a before/after example

---

Citation

If you use Metaskill in research, please cite:

@software{sung2025metaskill,
  author  = {Sung, Flood},
  title   = {Metaskill: A Meta-Skill for Autonomous AI Agent Team Generation},
  year    = {2025},
  url     = {https://github.com/xvirobotics/metaskill},
  license = {MIT}
}

A CITATION.cff file is also provided for GitHub's "Cite this repository" button.

---

License

MIT © XVI Robotics