Skills, Plugins & Agents | Claude Code Recap Card C04

Comparison table

Mechanism	Scope	Invocation	Resources	Use Case
Skill (user)	Codified workflow	`/name` (user)	Yes	Repeatable task triggered manually
Skill (auto)	Reusable knowledge	Auto by model	Yes	Shared expertise across agents
Plugin	Marketplace	Global	Yes	Third-party integration
Agent	Autonomous specialist	Agent tool	Via memory	Complex delegation

CC 2.1.3: Custom commands and skills share invocation behavior. Existing .claude/commands/*.md files remain supported. Add disable-model-invocation: true for user-triggered-only skills.

User-Invocable Skills (formerly “Commands”)

Since CC 2.1.3, manually triggered workflows live in .claude/skills/ with disable-model-invocation: true.

# .claude/skills/release/SKILL.md
---
name: release
description: Prepare a release (bump version, CHANGELOG, tag)
allowed-tools: [Read, Write, Bash]
disable-model-invocation: true
---
1. Read current version from package.json
2. Bump according to semver (patch/minor/major)
3. Update CHANGELOG.md
4. Create git tag

When to choose: You have a sequence of instructions you trigger intentionally several times a week.

Model-invocable skills

A skill without disable-model-invocation may be invoked by the model when its description fits the task.

# .claude/skills/security-guardian/SKILL.md
---
name: security-guardian
description: OWASP security review -- use when reviewing auth, routes, or user input
allowed-tools: [Read, Grep, Glob]
---
Analyze code for OWASP Top 10 vulnerabilities...

When to choose: Multiple agents need the same specialized knowledge, or you want Claude to automatically load the right expertise.

Plugins: Third-Party integrations

Plugins come from the marketplace and add external capabilities. Add a catalog with /plugin marketplace add <source>, then install a selected plugin with /plugin install <name>@<marketplace>.

When to choose: An existing third-party integration covers exactly your need.

${CLAUDE_PLUGIN_DATA} (v2.1.78+): Persistent directory for storing state between sessions.

Agents: Complex task delegation

An agent is a specialized Claude with its own tools and scope. It isolates context, not simulates a human role.

---
name: security-audit
description: Review authentication and input handling for security defects
model: opus
tools: Read, Grep, Glob
---
Analyze code for OWASP vulnerabilities...

When to choose: A long task that would pollute the main context, or parallelizable work.

Quick decision rule

Manually triggered workflow = User-invocable Skill. Knowledge to share across agents = Model-invocable Skill. Third-party integration = Plugin. Task to delegate with isolated context = Agent.