Notre avis

Guide pour définir et utiliser efficacement les sous-agents Claude avec des modèles de délégation structurés.

Points forts

Structure de délégation en 7 sections obligatoire qui élimine l'ambiguïté
Prise en charge des tâches en arrière-plan pour les sous-agents explore et librarian
Liste blanche d'outils explicite empêchant les dérives
Modèles réutilisables pour créer de nouveaux sous-agents spécialisés

Limites

Nécessite un investissement initial pour définir chaque sous-agent
Peut être excessif pour des tâches simples qui ne justifient pas une délégation
Dépend de la qualité des définitions de sous-agents existantes

Quand l'utiliser

Lorsque vous devez déléguer des tâches complexes et multi-étapes à des agents spécialisés nécessitant un contexte précis.

Quand l'éviter

Pour des requêtes simples ou des actions triviales qui peuvent être effectuées directement sans sous-agent.

Exemples

Create a new subagent for database migrations

I need to define a new subagent called 'migration-engineer' that specializes in database schema changes. Use the subagent definition template and include required tools: Read, Write, SqlRun. The subagent should focus on SQL migrations and rollback planning.

Delegate code exploration with structured prompt

Background the explore subagent to find all API routes in the project. Use the 7-section structure: TASK, EXPECTED OUTCOME, REQUIRED SKILLS (explore), REQUIRED TOOLS (Grep, Read), MUST DO (search for 'router.get', 'router.post', etc.), MUST NOT DO (modify files), CONTEXT (working in ./src/routes).

Debug a subagent delegation that failed

My subagent delegation for 'oracle' on the architecture decision keeps returning vague results. Check the common pitfalls and suggest improvements to the MUST DO and MUST NOT DO sections of my prompt.

name: subagents-creator description: Guide for defining and using Claude subagents effectively. Use when (1) creating new subagent types, (2) learning how to delegate work to specialized subagents, (3) improving subagent delegation prompts, (4) understanding subagent orchestration patterns, or (5) debugging ineffective subagent usage.

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

One action per delegation: Combine tasks in parallel calls, not one call
Be exhaustive: "MUST DO" and "MUST NOT DO" sections prevent drift
Background everything: Use background_task for explore and librarian
Explicit tool lists: Prevent subagents from using unauthorized tools
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

Créateur de Sous-agents Claude

Recommandé pour