The short answer
A prompt says what you want right now. Rules say what must always hold. A Skill teaches the agent how a kind of work is done. MCP and tools let it reach other systems. A subagent takes one slice of a big job in its own context. A plugin bundles several of these into a single install.
What each layer does, from prompts to plugins
There are seven layers, and each one answers a different question. The first three shape how the agent thinks, the next two decide what it can reach, and the last two are about splitting up work and shipping it.
| Layer | Its job | Lives in |
|---|---|---|
| Prompt | Says what you want this turn | Your message |
| Rules / instructions | Keep standing rules in front of the agent | CLAUDE.md, AGENTS.md, .cursor/rules/ |
| Skill | Teaches a method you can reuse | A folder with SKILL.md |
| Tool | Does what text alone can’t, like running code | Built into the agent |
| MCP server | Connects outside data and actions | A local or remote server |
| Subagent | Handles one focused part in its own context | The agent’s subagent config |
| Plugin | Bundles Skills, MCP servers, and more into one install | A plugin marketplace or repo |
“MCP connects Claude to data; Skills teach Claude what to do with that data.” (Anthropic)
A plugin is packaging, not a new capability. In Claude Code, one plugin can bundle Skills, slash commands, subagents, hooks, and MCP servers, and /plugin installs the whole set. Plugins are listed in plugin marketplaces, which are catalogs that usually live in a Git repo. Codex plugins follow the same idea for shipping Skills and MCP connections together.
A Skills marketplace such as Skillry works one level down. It gives you individual Skill folders to drop into Claude Code, Codex, or Cursor. Anthropic publishes its own Skills too, like the ones that create Word, PowerPoint, Excel, and PDF files. Install a plugin when you want a whole toolkit at once, and a single Skill when you only need one method.
Pick by what you’re trying to do
Start from the job in front of you, not from the feature list. Most setups begin with the first three rows and add the others when a real need shows up.
A one-off question
A prompt. Add a tool if it needs live data.
“Always use pnpm”
A rule, one line long.
A workflow you keep repeating
A Skill. Add MCP if it needs outside data.
Reaching Jira, a database, or an API
An MCP server, plus a Skill that explains the procedure.
Research, build, and review at the same time
Subagents, all sharing the same Skills.
Handing the whole setup to your team
A plugin, so everyone installs it in one step.
One real task that uses most of the layers
“Prepare Friday’s product review and flag launch risks.”
The prompt names the week and who’s reading. Repository rules say what can’t leave the company. A Skill sets the report’s structure and the checklist it has to pass. MCP servers pull closed issues from Linear and numbers from analytics. While the main agent writes, a subagent reads through customer feedback. And a plugin lets everyone else at the company install the same Skill and connectors in one step.
In Claude Code, a Skill can even run as its own subagent. Add context: fork to its frontmatter.
Plugins vs Skills, and other common mix-ups
- “A plugin is just a bigger Skill.” A plugin is the box. It can hold Skills, but also commands, subagents, hooks, and MCP servers. A Skill on its own is one folder with one method, and it works without any plugin.
- “A Skill can connect to my database.” Not by itself. Access comes from a tool or an MCP server; the Skill explains how to use it.
- “Rules and Skills do the same job.” Rules load in every chat, while a Skill loads when its task comes up. Put a long procedure in your rules and every conversation pays for it.
- “MCP and Skills compete.” Anthropic describes them as complements, and MCP now has an official extension for serving Skills.
Where to go next
MCP vs Skills
Access versus know-how, and when you need both. Read the comparison.
Skills vs prompts
Today’s goal versus the method you reuse. Read the comparison.
Skills vs instructions
Always-on rules versus Skills that load on demand. Read the comparison.
Common questions
What’s the difference between Skills and subagents?
A Skill is a set of instructions the agent loads. A subagent is a separate worker with its own context. In Claude Code you can run a Skill inside a subagent with context: fork.
Are Claude Code plugins the same as Skills?
No. A plugin is a package that can hold Skills along with slash commands, subagents, hooks, and MCP servers. A plugin’s Skills run as /plugin-name:skill-name, so they never clash with yours.
Do I need all of these?
No. Start with prompts and a few rules. Add a Skill once you notice you’re repeating a workflow, and MCP when the agent has to reach another system.
Sources
Did this guide help?