Claude Code Mechanism Selector

Claude Code Mechanism Selector

This skill helps you choose the right extension mechanism for Claude Code based on your requirements.

Quick Decision Guide

Ask these questions to determine the best mechanism:

1. Does it need to run deterministically (guaranteed, not AI-decided)?

• YES → Use Hooks (shell commands at lifecycle events)
• NO → Continue to question 2

2. Does it need a separate context window or specialized AI behavior?

• YES → Use Subagents (isolated context, custom system prompts)
• NO → Continue to question 3

3. Should it activate automatically based on context?

• YES → Use Skills (model-invoked, automatic discovery)
• NO → Use Slash Commands (user-invoked with /​command)

Comparison Matrix

When to Use Each

Slash Commands

Best for quick, repeatable prompts you want explicit control over.

Examples:

• /​review - Review code for bugs
• /​explain - Explain code in simple terms
• /​commit - Generate commit message

See commands.md for details.

Skills

Best for complex capabilities that should activate automatically.

Examples:

• PDF processing with scripts and templates
• Data analysis with reference documentation
• Code review with checklists and style guides

See skills.md for details.

Subagents

Best for specialized tasks requiring isolated context or custom AI behavior.

Examples:

• Code reviewer with specific review checklist
• Debugger with systematic debugging process
• Data scientist for SQL/​BigQuery analysis

See subagents.md for details.

Hooks

Best for deterministic actions that must always happen.

Examples:

• Auto-format code after edits (PostToolUse)
• Log all bash commands (PreToolUse)
• Block edits to sensitive files (PreToolUse)
• Custom notifications (Notification)

See hooks.md for details.

Decision Examples

Combining Mechanisms

These mechanisms complement each other:

• Skill + Subagent: Skill provides knowledge, subagent provides isolated execution
• Slash Command + Hook: Command triggers workflow, hook ensures formatting
• Subagent + Hook: Subagent does work, hook validates output