agent-factory

Agent Factory

A comprehensive system for generating production-ready Claude Code agents and sub-agents. This skill provides templates, standards, and generation tools to create custom agents that seamlessly integrate with Claude Code's agent system.

What This Skill Does

This skill helps you create custom Claude Code agents for any domain or workflow. It generates properly formatted agent files that Claude Code can automatically discover and invoke when relevant.

Capabilities

1. Generate Custom Agents - Create specialized agents for any domain (frontend, backend, testing, product, etc.)
2. Enhanced YAML Frontmatter - Rich metadata including color coding, field categorization, expertise levels
3. Tool Access Guidance - Recommends optimal tool configurations based on agent type
4. MCP Integration - Suggests relevant MCP server tools for enhanced capabilities
5. Execution Pattern Assignment - Ensures proper parallel/sequential execution for safety
6. Validation - Checks agent configuration against best practices

Agent Types Supported

Strategic Agents (Lightweight, Parallel-Safe)

• Purpose: Planning, research, analysis
• Tools: Read, Write, Grep only
• Execution: 4-5 agents can run in parallel
• Color: Blue
• Examples: product-planner, market-researcher, architect

Implementation Agents (Full Tools, Coordinated)

• Purpose: Code writing, feature building
• Tools: Read, Write, Edit, Bash, Grep, Glob
• Execution: 2-3 agents coordinated
• Color: Green
• Examples: frontend-developer, backend-developer, api-builder

Quality Agents (Heavy Bash, Sequential Only)

• Purpose: Testing, validation, review
• Tools: Read, Write, Edit, Bash, Grep, Glob
• Execution: 1 agent at a time (NEVER parallel)
• Color: Red
• Examples: test-runner, code-reviewer, security-auditor

Coordination Agents (Lightweight, Orchestration)

• Purpose: Manages other agents, validates integration
• Tools: Read, Write, Grep
• Execution: Orchestrates others
• Color: Purple
• Examples: fullstack-coordinator, workflow-manager

Enhanced YAML Frontmatter

Every generated agent includes rich metadata:

---
name: agent-name-kebab-case
description: When to invoke this agent
tools: Read, Write, Edit  # Comma-separated
model: sonnet  # sonnet|opus|haiku|inherit
color: green  # Visual categorization
field: frontend  # Domain area
expertise: expert  # beginner|intermediate|expert
mcp_tools: mcp__playwright  # MCP integrations
---

Field Categories

Development: frontend, backend, fullstack, mobile, devops Quality: testing, security, performance Strategic: product, architecture, research, design Domain: data, ai, content, finance, infrastructure

Color Coding

• Blue: Strategic/planning agents
• Green: Implementation/development agents
• Red: Quality/testing agents
• Purple: Coordination/orchestration agents
• Orange: Domain-specific specialists

Expertise Levels

• Beginner: Simple, focused tasks
• Intermediate: Moderate complexity workflows
• Expert: Advanced, complex operations

How to Use

Quick Start

1. Open the prompt template: templates/AGENTS_FACTORY_PROMPT.md
2. Scroll to bottom - Find template variables
3. Fill in your details:

   AGENT_NAME: my-custom-agent
   DESCRIPTION: What this agent does and when to invoke it
   DOMAIN_FIELD: frontend
   TOOLS_NEEDED: Read, Write, Edit, Bash
   

4. Copy entire prompt - Include filled variables
5. Paste into Claude - Claude.ai, Claude Code, or API
6. Receive agent file - Complete .md file ready to use
7. Install agent - Copy to .claude/agents/ or ~/.claude/agents/

Example Invocation

@agent-factory

Create a custom agent:
Name: api-integration-specialist
Type: Implementation
Domain: backend
Description: API integration expert for third-party services
Capabilities: OAuth, REST clients, error handling
Tools: Read, Write, Edit, Bash
MCP: mcp__github

Output: Complete .claude/agents/api-integration-specialist.md file

Generated Agent Structure

Each generated agent is a single Markdown file:

---
name: custom-agent
description: Triggers auto-invocation
tools: Read, Write, Edit
model: sonnet
color: green
field: backend
expertise: expert
mcp_tools: mcp__github
---

You are a [role] specializing in [domain].

When invoked:
1. [Step 1]
2. [Step 2]
3. [Step 3]

[Detailed instructions]
[Checklists]
[Best practices]
[Output format]

Integration Workflows

Workflow 1: Feature Development

1. product-planner → Creates requirements
2. frontend-developer + backend-developer → Build (parallel)
3. test-runner → Validates (sequential)
4. code-reviewer → Reviews (sequential)

Workflow 2: Bug Fix

1. debugger → Analyzes issue
2. [appropriate-dev-agent] → Fixes
3. test-runner → Validates fix

Workflow 3: Code Review

1. code-reviewer → Quality review (can run solo)
2. security-auditor → Security scan (can run solo)

MCP Tool Integration

Common MCP servers to integrate:

• mcp__github: PR reviews, issues, repo operations
• mcp__playwright: E2E testing, screenshots, browser automation
• mcp__context7: Documentation search, knowledge queries
• mcp__filesystem: Advanced file operations
• Custom MCP servers: Any user-configured MCP tools

Agents automatically reference MCP tools in their capabilities when configured.

Safety & Performance

Process Monitoring

Agents consume system resources. Monitor with:

ps aux | grep -E "mcp|npm|claude" | wc -l

Safe ranges:

• 15-20: Strategic agents (parallel)
• 20-30: Implementation agents (coordinated)
• 12-18: Quality agents (sequential)

Warnings:

• 30: Reduce parallelization

• 60: Critical - restart system

Execution Rules

✅ Safe: 4-5 strategic agents in parallel ✅ Safe: 2-3 implementation agents coordinated ❌ Unsafe: Quality agents in parallel (crashes system)

Best Practices

1. Keep agents focused - One clear responsibility per agent
2. Use descriptive descriptions - Enables auto-invocation
3. Follow tool access patterns - Match tools to agent type
4. Specify execution pattern - Prevents performance issues
5. Leverage MCP tools - Enhance agent capabilities
6. Test agents incrementally - Start simple, add complexity
7. Version control agents - Check project agents into git

Limitations

• Agents are templates - customize for your specific needs
• Tool suggestions are guidelines, not requirements
• MCP tools require servers to be configured
• Performance depends on system resources
• Generated agents need testing in your environment

Installation

Generated Agent Files:

Place in one of these locations:

Project agents (shared with team):

.claude/agents/custom-agent.md

Personal agents (available everywhere):

~/.claude/agents/custom-agent.md

When to Use This Skill

Create custom agents for:

• Domain-specific workflows (data science, ML, finance)
• Team-specific conventions (your code style, testing approach)
• Specialized tools or frameworks (Shopify, AWS, Kubernetes)
• Custom MCP server integrations
• Rapid prototyping of agent ideas

Use the AGENTS_FACTORY_PROMPT.md template when:

• You need multiple related agents
• You want consistent agent patterns
• You're building an agentic framework
• You want to test agent concepts quickly

---

Version: 1.0.0 Last Updated: October 22, 2025 Compatibility: Claude Code (agents system) Template Location: templates/AGENTS_FACTORY_PROMPT.md

Review checklist

Before shipping a generated agent:

• Frontmatter has name (kebab-case), description with trigger phrase, and appropriate tools restriction
• Description states WHAT the agent does AND WHEN to invoke it (third person)
• At least one concrete "When to invoke" example is included in the body
• A "When NOT to invoke" section excludes 2–3 adjacent use cases
• Tool allowlist is the minimum needed; broad * access is justified or removed
• The agent file is the only artifact created — no orphan scripts or reference files

When NOT to Use

• When you need a multi-file skill (SKILL.md + scripts/resources) — build a skill directly instead; agents are single-file specialists.
• When you want an automated slash command or event hook — use generating-slash-commands or generating-hooks respectively; agents are invoked by Claude's delegation logic, not by commands or events.
• When you just need a one-off prompt for the current session — write it inline rather than generating a persistent agent file.