Choose the Right Claude Extension Surface

TL;DR

Use the smallest Claude extension surface that fits the job:

• Skill: reusable expertise, standards, or process knowledge.
• Slash command: a user-triggered workflow shortcut.
• Hook: deterministic automation around a Claude Code lifecycle event.
• Subagent: delegated specialist work with its own focus and context.
• MCP server: live tools, resources, or external systems.
• Settings: project or user policy that should apply consistently.

When two surfaces could work, choose the one with less runtime authority. A static skill or slash command is usually easier to review than a hook or MCP server because it does not add live access by itself.

Prerequisites & Requirements

• {"task": "Workflow is named", "description": "You can describe the repeated task in one sentence"}
• {"task": "Sharing scope is known", "description": "The extension is personal, project-local, team-shared, or organization-managed"}
• {"task": "Access needs are mapped", "description": "You know whether the workflow needs prompts only, repository files, commands, or external systems"}
• {"task": "Risk is understood", "description": "You know whether the workflow can read data, write files, call tools, or run automatically"}

Core Concepts Explained

Skills package reusable expertise

Choose a skill when Claude needs durable guidance: domain conventions, review criteria, migration steps, testing heuristics, or a playbook that should travel between sessions. Skills are a good fit when the main value is knowledge and instructions rather than a live integration.

Slash commands package repeatable prompts

Choose a slash command when a human should intentionally start a workflow with a short command. They are useful for recurring tasks such as review checklists, release notes, triage prompts, migration planning, or issue analysis.

Hooks package deterministic automation

Choose a hook when something should happen at a lifecycle point without a user typing a prompt each time. Hooks are best for narrow, predictable automation: notifications, formatting checks, logging, guardrails, or context reminders.

Subagents package delegated specialist work

Choose a subagent when Claude should hand a task to a focused specialist. A subagent can carry a narrower role, review lens, or domain focus than the main conversation. Use this when separation of attention matters.

MCP packages live capability

Choose MCP when Claude needs to interact with tools, resources, APIs, databases, files, or services through a protocol boundary. MCP is the right surface for live data and actions, but it also deserves the most access review.

Settings package policy

Choose settings for behavior that should apply consistently across the user or project environment: allowed tools, model preferences, permissions, or team defaults. Settings are not a workflow by themselves; they shape the environment where workflows run.

Step-by-Step Decision Guide

1. Start with the output. If the workflow produces advice, a checklist, a review, or a plan, start with a skill or slash command. If it must interact with live systems, consider MCP.

2. Decide who starts it. If a human should opt in each time, use a slash command. If the behavior should apply automatically at a lifecycle point, use a hook.

3. Decide whether expertise should travel. If the workflow is mostly reusable know-how, package it as a skill. A skill can include process, examples, rubrics, and references without creating a live integration.

4. Decide whether work needs delegation. If the task benefits from a specialist role or separate focus, use a subagent. This is useful for code review, test planning, security triage, documentation review, or domain analysis.

5. Decide whether Claude needs live capabilities. If Claude must read a resource, call an API, query a database, inspect an issue tracker, or invoke a tool, use MCP and review the exposed tools/resources.

6. Put policy in settings. If the question is "what should Claude be allowed to do here?" instead of "what workflow should Claude run?", use settings.

7. Combine surfaces only when needed. A slash command can call on a skill; a subagent can use MCP tools; a hook can remind Claude about policy. Keep each layer small enough that a teammate can review it.

Decision Matrix

Review Checklist

• {"task": "Smallest surface", "description": "The extension does not add automation or live access unless needed"}
• {"task": "Clear owner", "description": "A person or team owns updates and review"}
• {"task": "Sharing scope", "description": "Personal, project, team, or organization scope is documented"}
• {"task": "Access review", "description": "Hooks and MCP tools are reviewed for files, commands, services, and side effects"}
• {"task": "Discoverability", "description": "Teammates can find and understand the extension"}
• {"task": "Rollback path", "description": "The extension can be disabled, removed, or replaced safely"}

Troubleshooting

• The workflow is becoming too broad: split it into a skill for reusable knowledge plus a slash command for the entry point.
• A hook is doing too much: move judgment-heavy work into a command or skill and keep the hook deterministic.
• An MCP server is used only for static text: replace it with a skill or command unless live resources are actually required.
• Subagents overlap heavily: merge roles or define sharper handoff criteria.
• Settings hide project behavior: document the policy in the repository so teammates understand what applies automatically.

Duplicate Check

This guide is a decision framework for choosing between Claude Code extension surfaces. Existing entries cover individual commands, hooks, agents, skills, and MCP servers, but they do not provide a source-backed comparison that helps teams choose the smallest appropriate surface for a workflow.

References

• Claude Code subagents - https://code.claude.com/docs/en/sub-agents
• Claude Code skills - https://code.claude.com/docs/en/skills
• Claude Code slash commands - https://code.claude.com/docs/en/slash-commands
• Claude Code hooks - https://code.claude.com/docs/en/hooks
• Claude Code MCP - https://code.claude.com/docs/en/mcp
• Claude Code settings - https://code.claude.com/docs/en/settings