The short answer
Your prompt states what you need this minute. Rules state what must be true every time. A Skill shows the agent how a kind of job gets done. MCP servers and tools reach systems beyond the chat. A subagent takes one piece of a large job and works on it separately. A plugin packs several of these into one install.
Seven layers, from a single prompt to a plugin
Each layer settles a separate question. The first three steer how the agent reasons, the next two set what it can touch, and the last two divide a job up and hand the setup to others.
| Layer | What it is for | Where it sits |
|---|---|---|
| Prompt | What you want in this turn | The message you type |
| Rules / instructions | Lasting rules, in view on every turn | CLAUDE.md, AGENTS.md, .cursor/rules/ |
| Skill | A method the agent can reuse | A folder containing SKILL.md |
| Tool | Acts where words fall short, such as running code | Part of the agent itself |
| MCP server | Data and actions from other systems | A server on your machine or online |
| Subagent | One narrow part, in a separate context | The agent's subagent settings |
| Plugin | Skills, MCP servers and more as one install | A plugin marketplace or a repository |
Anthropic sums up the split this way: MCP links Claude to the data, and a Skill tells Claude what to do once it has it.
A plugin is packaging, not an extra ability. In Claude Code one plugin can carry Skills, hooks, slash commands, MCP servers and subagents, and /plugin sets up the lot. Plugins are listed in plugin marketplaces, catalogs normally kept in a Git repository. Codex plugins work on the same principle, bundling MCP connections with Skills.
A Skills marketplace like The Frontier Brief sits one step lower: it hands you single Skill folders to place in Claude Code, Codex or Cursor. Anthropic releases Skills too, including ones that produce Word, PowerPoint, Excel and PDF files. Choose a plugin for a complete kit in one go, and a lone Skill when one method is all you need.
Choose by the job you have in front of you
Begin with the work, not the feature list. Most setups start with the top three rows and add the rest when a genuine need appears.
A question you ask once
A prompt. Add a tool when it needs fresh data.
“We always use pnpm”
A rule, written in a single line.
A process you run over and over
A Skill, plus MCP when it needs outside data.
Working with Jira, an API or a database
An MCP server, with a Skill that spells out the steps.
Researching, building and reviewing in parallel
Subagents that all draw on the same Skills.
Giving the full setup to your colleagues
A plugin, so each person installs it in one go.
One real job, followed through most of the layers
“Get Friday’s product review ready and point out anything that could hold up the launch.”
The prompt sets the week and the audience. Repository rules list what must stay inside the company. A Skill fixes the report's layout and the checks it must clear. MCP servers fetch finished issues from Linear and figures from analytics. A subagent works through customer feedback while the main agent drafts. A plugin then gives everyone else at the company that Skill and those connectors in a single step.
Claude Code can also run a Skill as a subagent of its own: put context: fork in the Skill’s frontmatter.
Plugins vs Skills, and other things people confuse
- “A plugin is simply a large Skill.” A plugin is the container. Skills go inside, and so can hooks, commands, MCP servers and subagents. On its own, a Skill is a single folder holding a single method, and it runs fine without a plugin.
- “My Skill can talk to the database.” Not on its own. An MCP server or a tool supplies the access; the Skill describes how to use it.
- “Skills and rules do the same thing.” Rules load into every conversation; a Skill loads only when its task appears. A long procedure kept in your rules costs context in every chat.
- “MCP and Skills are rivals.” Anthropic presents them as partners, and an official MCP extension now exists for delivering Skills.
Where to read next
MCP vs Skills
Reaching systems versus knowing the method, and when a job calls for both. See the comparison.
Skills vs prompts
What you want today versus a method you keep reusing. See the comparison.
Skills vs instructions
Rules that are always loaded versus Skills that load when needed. See the comparison.
Common questions
How is a Skill different from a subagent?
A Skill is guidance the agent reads in. A subagent is a second worker running in its own context. In Claude Code, adding context: fork makes a Skill run inside a subagent.
Is a Claude Code plugin just a Skill?
No. Think of a plugin as a bundle: besides Skills it may carry hooks, MCP servers, subagents and slash commands. Plugin Skills are called as /plugin-name:skill-name, so their names never collide with yours.
Do I have to use every layer?
Not at all. Begin with prompts and a handful of rules. Add a Skill when you catch yourself repeating a process, and an MCP server when the agent must get into another system.
Sources
Did this guide help?
