Explainer
Skills vs agents, plugins, MCP and commands
Five different mechanisms, one question behind all of them: what should the agent know or do that it does not know or cannot do today?
The short version
| Mechanism | What it is | When it loads | Write it when |
|---|---|---|---|
| Skill | A folder with SKILL.md: instructions for one job |
On demand, when the description matches the task | You want repeatable behaviour the agent should apply itself |
| Subagent | A separate agent with its own context and tools | When you or the main agent delegates | The work needs isolation, a different model, or parallel effort |
| Plugin | A package that can bundle skills, commands, agents and hooks | Installed into the client | You are distributing several capabilities together |
| MCP server | A process exposing tools and data over a protocol | Connected at session start | The agent needs connectivity: a database, an API, a browser |
| Slash command | A named prompt you invoke explicitly | Only when you type it | You want to trigger a workflow yourself |
| CLAUDE.md | Project memory read at session start | Always | The agent must know project conventions every time |
Skills are behaviour
A skill tells the agent how to do a job: the steps, the order, the checks. It is text, so it is cheap to write, review and version. Claude Code, Codex and Cursor all support some form of it.
The key property is triggering: the frontmatter description is what the agent matches against your request. A skill with a vague description is a skill that never runs.
Subagents are isolation
A subagent gets its own context window and usually its own tool permissions. Use one when the work is long, parallel, or needs a different model. A skill cannot isolate context; a subagent cannot be installed as a file.
Plugins are packaging
A plugin is how you ship a bundle: several skills, slash commands, agents and hooks in one installable unit. If you are distributing one skill, ship a skill. If you are distributing a workflow, a plugin is the delivery vehicle.
MCP is connectivity
An MCP server exposes tools: query a database, call an API, control a browser. It does not teach the agent a method. In practice you often want both: an MCP server for access, a skill for the procedure.
Commands and CLAUDE.md are explicit or always-on
A slash command runs only when you type it — good for deliberate workflows. CLAUDE.md is read every session — good for conventions, bad for procedures that only matter occasionally.
A decision rule
- Does the agent need access to something? → MCP server.
- Does it need a method it should apply by itself? → skill.
- Does the work need isolation or parallelism? → subagent.
- Do you want to trigger it manually? → slash command.
- Must it be true every session? → CLAUDE.md.
- Shipping several of the above together? → plugin.
Community notes
Found a broken install? Know a skill we missed?
One line is enough: the skill name, what happened, and the command you ran. We read every note and fix the index.