Subagents Creator

This skill provides guidance for defining, using, and improving Claude subagents—the specialized agents that handle specific domains like explore, librarian, oracle, and frontend-ui-ux-engineer.

Quick Start

Delegating Work

When delegating to subagents, use the mandatory 7-section structure:

1. TASK: Atomic, specific goal (one action per delegation)
2. EXPECTED OUTCOME: Concrete deliverables with success criteria
3. REQUIRED SKILLS: Which skill to invoke
4. REQUIRED TOOLS: Explicit tool whitelist (prevents tool sprawl)
5. MUST DO: Exhaustive requirements - leave NOTHING implicit
6. MUST NOT DO: Forbidden actions - anticipate and block rogue behavior
7. CONTEXT: File paths, existing patterns, constraints

Choosing a Subagent

See subagent-types.md for detailed guidance on which subagent to use:

• explore: Contextual grep for codebases
• librarian: Reference search (docs, OSS, web)
• oracle: Deep reasoning for architecture/complex decisions
• frontend-ui-ux-engineer: Visual UI/UX changes

Defining New Subagents

Only create subagents when: The task domain has distinct tooling, expertise, or patterns that benefit from specialization.

See delegation-patterns.md for:

• Subagent definition templates
• When to create a new subagent vs using existing ones
• Naming and description guidelines

Common Pitfalls

See common-pitfalls.md for:

• Vague delegation prompts and why they fail
• Over-delegating trivial tasks
• Subagent misalignment with task type
• Anti-patterns in agent orchestration

Best Practices

1. One action per delegation: Combine tasks in parallel calls, not one call
2. Be exhaustive: "MUST DO" and "MUST NOT DO" sections prevent drift
3. Background everything: Use background_task for explore and librarian
4. Explicit tool lists: Prevent subagents from using unauthorized tools
5. Verify results: Check that delegated work meets expectations before proceeding

Delegation Example

# GOOD: Specific, exhaustive
background_task(
    agent="explore",
    prompt="""
    1. TASK: Find all authentication implementations
    2. EXPECTED OUTCOME: List of files with auth logic, patterns used
    3. REQUIRED SKILLS: explore
    4. REQUIRED TOOLS: Grep, Read
    5. MUST DO: Search for 'jwt', 'session', 'auth' patterns; identify middleware; list all endpoints
    6. MUST NOT DO: Don't modify any files; don't run build/test commands
    7. CONTEXT: Working in ./src directory, looking for Express.js patterns
    """
)

# BAD: Vague, implicit expectations
background_task(
    agent="explore",
    prompt="Find auth stuff in the codebase"
)

Reference Files

• subagent-types.md - When to use each subagent type
• delegation-patterns.md - Prompt templates and patterns
• common-pitfalls.md - Anti-patterns and how to avoid them