Agent skills vs prompts vs MCP servers vs subagents: which one do you need?
Four words get used for “making the AI do more”, and they are not the same thing. Here is what each one is, when it’s the right choice, and how they work together — explained with one real job built all four ways.
The short answer
- Prompt — what you say to the AI, once. 380 free prompts.
- Skill — how to do a job, saved so the AI uses it by itself every time. 1828 free skills.
- MCP server — a tool the AI can call: GitHub, a browser, a database. MCP servers.
- Subagent — a specialist the AI hands a whole task to, with its own workspace. Subagents.
Prompts and skills are knowledge. MCP servers are hands. Subagents are extra workers.
Side by side
| Prompt | Skill | MCP server | Subagent | |
|---|---|---|---|---|
| What it is | Text in one message | A folder with SKILL.md | A program exposing tools | A Markdown file defining a role |
| Who starts it | You, every time | The AI, when the task matches | The AI, when it needs the tool | The AI (or you) delegates to it |
| Lasts | One chat | Every chat | While connected | Every session |
| Can carry files | No | Yes — templates, scripts | It is code | Instructions only |
| Reaches outside the chat | No | Only through tools | Yes | Through the tools it's allowed |
| Own context window | No | No | No | Yes |
| Where | Any chat app | Claude app, Claude Code, API; others via the open format | Any MCP client | Claude Code |
| Setup | Paste | Upload or copy a folder | Paste an address or a command | Drop a file in .claude/agents/ |
Each one in plain words
Prompt
The instruction you type or paste. Perfect for one-off jobs and for trying something out. Its weakness is memory: the next chat starts from zero, so anything you want every time has to be pasted every time.
Skill (agent skill)
A folder whose SKILL.md starts with a name and a description of when to use it. The AI keeps only that description in view and loads the full instructions — and any files next to them — when your request matches. It is how you stop re-explaining your process. Anthropic published the format as an open standard, which is why “agent skills” and “Claude skills” mean the same thing. More in what Claude skills are.
MCP server
The Model Context Protocol is a shared plug. An MCP server exposes tools — “search issues”, “open this URL”, “run this query” — and any AI app that speaks MCP can call them. It gives the model reach, not judgment: it can now open your pull requests, but it doesn’t know what a good review looks like. That part is a skill.
Subagent
A separate worker with its own instructions, its own tool permissions and its own context window. Your main session hands it a task and gets back a result, not the whole transcript. Use one when a job is big enough to crowd your conversation, or needs tighter permissions.
One job, four ways: a weekly SEO check
| Built as | What you do | What you get |
|---|---|---|
| A prompt | Paste an SEO-audit prompt and the page text every Monday | A decent audit — if you remember the prompt and paste the page |
| + a skill | Install an SEO-audit skill once | Claude runs the same checklist every time you ask “check SEO on this page” |
| + an MCP server | Connect a crawling server | Claude fetches the live pages itself instead of you pasting them |
| + a subagent | Add an SEO specialist subagent in Claude Code | The whole crawl-and-audit runs in its own context and returns a short report |
Each layer removes a step you used to do by hand. The real pieces, all free on AgentAlley:
- 1Claude opus as SEO auditorThe prompt version — paste it with your page for a one-off audit.by Musa Toktas · CC0-1.0 · full source on the page
- 2SEO auditThe skill version — install once, and Claude audits the same way every time.by coreyhaines31 · MIT · full source on the page
- 3Firecrawl MCP ServerThe tool — lets Claude crawl and read the live site itself.by firecrawl · MIT · full source on the page
- 4SEO SpecialistThe specialist — a subagent role that owns the whole SEO job.by msitarzewski · MIT · full source on the page
How to choose, in four questions
- Will I do this once? Use a prompt.
- Will I do it again, the same way? Make it a skill.
- Does the AI need to see or change something outside the chat? Connect an MCP server.
- Is the job big enough to bury my conversation? Hand it to a subagent.
FAQ
What are agent skills?
Agent skills are folders of instructions — a SKILL.md file plus optional scripts and reference files — that an AI agent loads when a task matches. Anthropic introduced them for Claude and published the format as an open standard, so other agents can read the same folders.
Is a skill better than MCP?
Neither is better; they do different jobs. MCP gives the model a tool it can call. A skill tells it how to do a job, including how to use its tools well. Most real workflows use both.
Is a skill just a saved prompt?
Close, with two differences. Claude decides by itself when to load a skill, based on its description, so you don't paste anything. And a skill can bundle files — templates, examples, scripts — that a pasted prompt can't carry.
When should I use a subagent instead of a skill?
Use a subagent when the job would flood your main conversation — a long code review, a search through a large codebase — or needs different tool permissions. A subagent works in its own context and hands back a summary. A skill works inside the conversation you're in.
Do I need to code to use any of these?
No for prompts, skills in the Claude app, and hosted MCP connectors — all are paste or upload. Local MCP servers and subagents are easiest in Claude Code or Cursor.
Which one should a beginner start with?
Start with a prompt. When you catch yourself pasting the same one repeatedly, turn it into a skill. Add an MCP server when the AI needs to reach something outside the chat. Add subagents last, when sessions get crowded.
Keep going
Updated 24 September 2026.