Subagent Builder

Subagent Builder

Overview

Create custom subagents—specialized AI assistants with isolated context windows, custom system prompts, specific tool access, and independent permissions. Subagents delegate specialized work while keeping verbose output out of your main conversation.

Quick Start

Create via /agents Command (Recommended)

1. Run /agents in Claude Code
2. Select Create new agent
3. Choose scope: User-level (all projects) or Project-level (current project only)
4. Select Generate with Claude and describe the subagent
5. Review/edit the generated configuration
6. Select tools, model, and color
7. Save and use immediately

Create Manually

Create a .md file in the appropriate directory with YAML frontmatter and system prompt:

• Project subagent: .claude/agents/agent-name.md
• User subagent: ~/.claude/agents/agent-name.md

---
name: my-agent
description: What this agent does and when to use it
tools: Read, Grep, Glob
model: sonnet
---

You are a specialized agent for [specific task].

When invoked:
1. First step
2. Second step
3. Third step

Subagent Creation Workflow

Step 1: Define Purpose

Identify the subagent's specific role. Answer:

• What task will this subagent handle?
• What problem does it solve?
• When should Claude delegate to it?

Write a clear description field—this is Claude's primary signal for when to delegate.

Good description: "Expert code reviewer. Proactively reviews code for quality, security, and maintainability. Use immediately after writing or modifying code."

Poor description: "Reviews code" (too vague, no trigger guidance)

Step 2: Choose Scope

Decide where the subagent should be available:

Step 3: Configure Capabilities

Set tool access based on what the subagent needs:

Read-only analysis (code reviewers, researchers):

tools: Read, Grep, Glob, Bash

Code modification (debuggers, optimizers):

tools: Read, Edit, Bash, Grep, Glob

Restricted access (database readers with validation):

tools: Bash
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "./scripts/validate.sh"

Full access (rare, use cautiously):

# Inherits all tools from main conversation
# No tools field needed

Step 4: Write System Prompt

The markdown body becomes the subagent's system prompt. Keep it:

• Task-focused: Specific to the subagent's role
• Actionable: Clear steps and checklists
• Concise: Under 5k words (progressive disclosure for complex agents)

Structure:

1. Role definition
2. When invoked behavior
3. Process/workflow
4. Output format

Example structure:

You are a ROLE specializing in DOMAIN.

When invoked:
1. First action
2. Second action
3. Third action

Process:
- Step 1 with details
- Step 2 with details

For each TASK:
- OUTPUT 1
- OUTPUT 2

Focus on OUTCOME.

Step 5: Choose Model

Select the appropriate model for the task:

Step 6: Set Permissions (Optional)

Use permissionMode to control permission handling:

permissionMode: acceptEdits  # Auto-accept file edits

For dynamic control, use hooks—see configuration.md for details.

Common Patterns

Pattern 1: Read-Only Analyst

Subagents that analyze without modifying. Use for code reviews, research, documentation generation.

tools: Read, Grep, Glob, Bash

Examples: Code reviewer, documentation analyzer, security auditor

Pattern 2: Fixer/Modifier

Subagents that analyze and modify code. Use for debugging, optimization, refactoring.

tools: Read, Edit, Bash, Grep, Glob

Examples: Debugger, optimizer, refactoring specialist

Pattern 3: Domain Specialist

Subagents for specialized domains with explicit model selection for capable analysis.

model: sonnet
tools: Bash, Read, Write

Examples: Data scientist, database analyst, API integrator

Pattern 4: Restricted Access

Subagents with tool access but validation via hooks.

tools: Bash
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "./scripts/validate.sh"

Examples: Database reader (read-only queries), API caller (rate-limited)

Pattern 5: Isolated Execution

Subagents for verbose operations to keep output out of main context.

model: sonnet
tools: Bash, Read

Examples: Test runner, log analyzer, documentation generator

Best Practices

Write Effective Descriptions

The description field is critical—Claude uses it to decide when to delegate. Include:

• What the subagent does
• When to use it (specific triggers)
• Use proactively to encourage automatic delegation

Example:

description: Security auditor for finding vulnerabilities. Use proactively after code changes or when working with authentication, authorization, or user input handling.

Limit Tool Access

Grant only necessary permissions. This:

• Improves security
• Keeps subagents focused
• Prevents accidental modifications

Default to read-only tools unless modification is essential.

Design Focused Subagents

Each subagent should excel at one specific task. Multiple focused subagents are better than one general-purpose subagent.

Good: code-reviewer, security-auditor, performance-profiler Avoid: do-everything-agent

Use Proactive Delegation

Include "use proactively" in descriptions to encourage automatic delegation:

description: Test runner that executes test suites and reports failures. Use proactively after code changes.

Check Into Version Control

For project subagents (.claude/agents/), commit them to version control so:

• Team members can use them
• Changes are tracked
• Knowledge is shared

Progressive Disclosure

For complex subagents with extensive documentation:

1. Keep SKILL.md lean (<5k words, <500 lines)
2. Move details to references - Create separate files for:
   • Advanced configurations
   • Domain-specific patterns
   • Extensive examples
3. Link from main prompt - Reference when relevant

This skill demonstrates this pattern:

• SKILL.md: Core workflow and common patterns
• references/examples.md: Concrete subagent examples
• references/configuration.md: Complete configuration reference

Resources

references/examples.md

Concrete examples of subagents for common patterns:

• Code reviewer (read-only)
• Debugger (read/write)
• Data scientist (domain-specific)
• Database reader (restricted access)
• Test runner (isolated execution)
• Documentation generator
• Performance optimizer

Use these as starting points or inspiration.

references/configuration.md

Complete reference for subagent configuration:

• All frontmatter fields and their purposes
• Tool access control patterns
• Permission modes and when to use each
• Model selection guidelines
• Skills in subagents
• Hooks configuration (PreToolUse, PostToolUse, Stop, SubagentStart, SubagentStop)
• Subagent scope and priority rules
• Disabling specific subagents

assets/templates/

Template files for common subagent patterns:

• readonly-analyst.template.md - For read-only analysis tasks
• readwrite-fixers.template.md - For agents that modify code
• domain-specialist.template.md - For domain-specific expertise

Copy and customize these templates to quickly create new subagents.

Next Steps

After creating a subagent:

1. Test it: "Use the [agent-name] agent to [task]"
2. Iterate: Refine the prompt and configuration based on usage
3. Share: Commit project subagents to version control
4. Chain: Use multiple subagents in sequence for complex workflows

For more subagent usage patterns, see the main Claude Code documentation on working with subagents.