Claude Code Skills vs Subagents vs Plugins

Start with the context problem

The question is not which feature is most powerful. The question is what should load, when it should load, and who should own it. CLAUDE.md loads every session. Skills load when relevant or invoked. Subagents run in separate context. Plugins package capabilities for distribution.

A long checklist that only applies to release work should not live in every prompt. A code review persona that reads hundreds of files should not bloat the main session. A team bundle that includes hooks and MCP config should not be copied by hand into every repository.

Primary-source anchor: Anthropic documents skills as SKILL.md-based capabilities that can load on demand, including supporting files and invocation controls.

Use skills for repeatable procedures

A skill is a durable instruction set for a task you repeat: review a PR, prepare a release, debug a flaky test, generate API docs, or run a deployment checklist. It can include supporting files and scripts, so the model does not need every detail in the prompt.

Skills are not just prompts. Good skills include the workflow, boundaries, required evidence, and any local scripts or references needed to perform the work.

• Good skill A release checklist with commands, required evidence, rollback notes, and a report template.
• Bad skill A vague personality instruction like be careful, with no commands or completion criteria.

Use subagents for isolated work

Subagents help when a task needs a separate context window or a specialized role. They are useful for broad investigation, independent review, security checks, migration analysis, or any task that would pollute the main implementation conversation.

They should return findings, not silently decide the merge. A subagent can review, explore, or implement a bounded slice, but the main session or human reviewer still integrates the result.

Primary-source anchor: Anthropic describes custom subagents as specialized assistants for task-specific workflows and context management.

Use plugins when distribution matters

A plugin is a packaging boundary. Use it when you want to share or version a set of skills, agents, hooks, and MCP configuration across projects. If the workflow is still experimental and local to one repo, standalone `.claude/` configuration is usually simpler.

Plugins are especially useful for platform teams: one package can carry the company review skill, approved hooks, code intelligence integration, and internal MCP server definitions.

Where to go for skill examples

This site stays focused on engineering practice. For a broader skill directory and authoring examples, use the sibling site Claude Has Skills. For memory-specific patterns, use Claude Memories. For context-window strategy, use Claude Context.

That boundary keeps this page useful: it helps you decide which extension to use before you go collect examples.