Commands, Skills, Plugins & Agents — Claude Code Recap Card C04

Comparison table

Mechanism	Scope	Persistence	Resources	Use case
Command	Single prompt	Session	No	Repeatable one-off task
Skill	Reusable capability	Cross-session	Yes	Shared expertise
Plugin	Marketplace	Global	Yes	Third-party ecosystem
Agent	Autonomous specialist	Cross-session	Via memory	Complex delegation

Commands: one-off tasks

A command is a Markdown file in .claude/commands/ that you invoke with /command-name. It is resolved at call time: its entire content is injected into the context as an instruction.

.claude/commands/
├── release.md        # /release
├── security-check.md # /security-check
└── sync.md           # /sync

When to choose: you have a sequence of instructions you type manually several times a week. A command codifies it once and for all.

Skills: reusable know-how

A Skill is a structured knowledge module (SKILL.md file with YAML frontmatter) that Claude can invoke when needed. Unlike a command that runs explicitly, a Skill is loaded based on context.

# .claude/skills/pdf-generator/SKILL.md
---
name: pdf-generator
description: Generates PDFs using the whitepaper-typst template
allowed-tools: [Read, Write, Bash]
---
Use quarto render --to whitepaper-typst...

When to choose: you have a technical process with variable parameters (PDF build, service deployment) that you want to fully delegate to Claude with guardrails.

Note on the 56% warning: Skills are invoked on demand, not automatically. If Claude does not invoke your Skill, use .claude/rules/ (systematic loading) or an explicit Command instead.

Plugins: third-party integrations

Plugins come from the marketplace and add external capabilities (Context7 for documentation, Linear MCP for tickets, Figma MCP for design-to-code). They are installed via /plugin marketplace add.

When to choose: an integration with an existing third-party service covers exactly your need. Do not build what already exists.

Agents: complex task delegation

An Agent is a specialized Claude with its own tools, its own scope, and potentially its own memory. It serves to isolate context, not to simulate a human role.

---
name: security-audit
model: opus
tools: Read, Grep, Glob
---
Analyze code for OWASP vulnerabilities...

When to choose: you have a long task that would pollute the main context, or work that can run in parallel while you continue elsewhere.

Quick decision rule

An instruction to repeat manually = Command. Know-how to encapsulate = Skill. A third-party integration = Plugin. A task to fully delegate with its own context = Agent.