Claude Code Skills vs. Subagents vs. Plugins: What to Use and When

When you add capabilities to Anthropic’s Claude Code coding assistant, you can give Claude reusable instructions, delegate work to a separate assistant, or install a package of related capabilities. These options solve different problems: skills provide guidance, subagents handle delegated work, and plugins package components for installation and sharing.

A “context” is the information Claude uses while working on a task. A subagent gets its own context, while a skill supplies instructions that Claude can use in a conversation or in a subagent. A plugin can bundle skills, subagents, and other components. The Claude Code feature overview describes how these parts can work together.

Use a skill for repeatable guidance

A skill is a set of reusable instructions, usually stored in a SKILL.md file. It can describe a procedure, checklist, coding convention, or provide reference information. Claude can load a skill when its description matches a task. You can also invoke one directly as a slash command, such as /review.

For example, a code-review skill could tell Claude which issues to check and how to format findings. Use a skill when you keep giving Claude the same directions or want it to follow the same procedure each time. The skills documentation also explains how to prevent a skill from being invoked automatically, which helps when its instructions involve actions with side effects.

When Claude uses a skill, its instructions load and stay in context across later turns. Keep the skill focused so it does not take up more of Claude’s working space than the task requires. You can also set a skill to run inside a subagent, so the subagent follows reusable instructions while doing separate work.

Use a subagent to delegate focused work

A subagent is a specialized Claude worker with its own instructions, tools, permissions, and context. Claude Code can choose a subagent based on its description and delegate work to it, or you can invoke one directly. The subagent sends its results back to the main conversation while keeping its exploratory work separate.

For example, ask a research subagent to inspect a set of files and report likely causes of a bug. The main conversation receives the findings rather than every search result and file detail. Use a subagent for an investigation that would otherwise fill the main conversation, for tasks that can run independently, or for work that needs limited tool access.

Subagents make their own requests, and those requests count toward the same usage limits as the main conversation. You can configure a subagent to use a different model or a narrower set of tools, but delegation does not guarantee lower cost or faster completion. Anthropic’s subagent documentation describes these options and limits.

Use a plugin to install and share components

A plugin packages Claude Code components so you can install and load them together. A package can include skills and subagent definitions. It can also include hooks, commands that run in response to configured events, and MCP servers, which connect Claude Code to external tools or data. A marketplace is a catalog for finding plugins; the plugin itself is the package.

Choose a plugin when you want to install a prepared set of capabilities, share a setup across repositories, or distribute updates. Skills and subagents can also work without plugins, so a plugin adds packaging and distribution rather than a new kind of worker. The plugin overview covers how plugins group these components.

An enabled plugin makes its components available across sessions. Component names and descriptions can take up context even when you do not invoke those components. Hooks and external-tool connections can also affect a session. Review a plugin before installing it: plugin code can run on your machine with your privileges. Anthropic’s plugin security guidance recommends inspecting its source, hooks, MCP servers, and executable files.

Choose by the job, then combine them

You can combine these choices. For example, you could write a deployment checklist as a skill, ask a subagent with limited tools to review a deployment, and package both in a plugin others can install. The skill provides the procedure, the subagent handles isolated work, and the plugin distributes the setup.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com