Chapter 4 of 14 · Understand · 5 min read
Skills vs prompts vs MCP vs subagents vs plugins: when to use which
Once you start customising an agent, you meet a crowd of overlapping ideas: system prompts, project instruction files, skills, MCP servers, subagents, plugins, slash commands, hooks. They can all change how an agent behaves, so it’s easy to reach for the wrong one. The good news is that each answers a different question. This chapter gives you a comparison table, a short description of each, and a rule of thumb for choosing.
The one-line version
- Prompt: what you want right now.
- Project instructions: facts and rules that apply to every task in this project.
- Skill: how to do a particular kind of job, loaded when that job comes up.
- MCP server: access to a system the agent couldn’t otherwise reach.
- Subagent: a separate worker with its own context for a self-contained subtask.
- Plugin: a package that bundles several of the above so they can be installed together.
Side by side
| Prompt | Skill | MCP server | Subagent | Plugin | |
|---|---|---|---|---|---|
| Answers the question | What should happen now? | How is this kind of job done well? | What can the agent reach? | Who should do this part? | How do I ship a set of these? |
| What it is | A message you type or paste | A folder with SKILL.md and optional files | A program exposing tools over a protocol | A separate agent instance with its own instructions | A package of skills, commands, agents, servers and settings |
| When it loads | When you send it | Description always; body when relevant | Tool definitions when connected | When the main agent delegates | When installed and enabled |
| Context cost | Whatever you write | Small until used | Every tool definition, every session | Runs in its own context; returns a summary | The sum of what it contains |
| Adds new access? | No | No, though its scripts use existing tools | Yes | Only what it’s given | If it includes servers |
| Reusable? | Only if you save and paste it | Yes, automatically | Yes | Yes | Yes, and shareable as a unit |
| Best for | One-off requests | Repeatable procedures and expertise | Integrations with apps and data | Parallel or noisy work: searching, reviewing | Distributing a team’s setup |
Prompts
A prompt is the right tool when the request is a one-off or changes every time. If you notice yourself pasting the same long prompt into conversation after conversation, that’s a sign it wants to become a skill. The prompt engineering basics guide covers writing good ones.
Project instructions
Most coding agents read an instructions file from the project, such as CLAUDE.md for Claude Code or AGENTS.md for several others, at the start of every session. Put things there that are true for all work in the project: how to run the tests, the coding conventions, which folders not to touch. Because it is loaded every time, keep it short. If a section only matters for one kind of task, such as how to write a database migration, move it into a skill so it loads only when needed.
Skills
Skills are for know-how: the steps, standards and pitfalls of a particular kind of job. They shine when a task recurs, when doing it well needs more than common sense, and when you’d like it done the same way every time. Because only the description is loaded up front, you can have many of them without slowing the agent down.
MCP servers
Reach for an MCP server when the agent needs to see or change something it currently can’t: your issue tracker, a design file, a company database. MCP gives abilities; it doesn’t teach judgement. The two pair naturally. An MCP server lets the agent read and update tickets; a skill tells it how your team writes a good bug report.
Subagents
A subagent is a helper the main agent can hand a task to. It starts with a fresh context, its own instructions and often a restricted set of tools, does the work, and returns a summary. Use subagents when a subtask would flood the main context with material you don’t need to keep, such as searching a large codebase or reading dozens of pages, or when independent pieces can run in parallel. The trade-off is that the subagent doesn’t see your whole conversation, so it needs a clear, self-contained brief.
Plugins
A plugin is a distribution format rather than a new capability. In Claude Code, for example, a plugin can bundle skills, slash commands, subagents, hooks and MCP server configuration so a team can install a whole setup in one go. Other agents have similar packaging. Use a plugin when you want to share a collection; use a plain skill folder when one skill is enough.
Not sure? Ask in this order: Is it a one-off? Use a prompt. Does the agent lack access? Add an MCP server. Is it a repeatable way of working? Write a skill. Would it flood the context or benefit from running in parallel? Use a subagent. Do you need to ship several of these together? Package a plugin.
Worked examples
| Need | Best fit | Why |
|---|---|---|
| “Summarise this article for me” | Prompt | One-off, nothing to reuse |
| “Always use our brand voice in marketing copy” | Skill | A repeatable standard that only matters for some tasks |
| “Look up the customer in our CRM before replying” | MCP server, plus a skill for the reply | Needs access the agent doesn’t have, then know-how |
| “Check all 40 services for the deprecated API” | Subagents | Lots of reading that can run in parallel and be summarised |
| “Tests are run with make test in this repo” | Project instructions | True for every task in the project |
| “Give the whole team our review setup” | Plugin | Bundles skills, commands and settings for one-step install |
They work best together
These aren’t rivals. A mature setup usually has a short project instructions file, a handful of well-described skills, one or two MCP servers for the systems you use every day, and perhaps a plugin that packages it for teammates. Subagents get used when a task is large. Each piece does one job, and none of them crowds the others out of the context.
Dispatching Parallel AgentsAgent skillA skill for splitting work across parallel subagents.Agent skill · MIT
Context EngineeringAgent skillPractical guidance on what to put in an agent’s context.Agent skill · MIT
MCP Server PatternsAgent skillPatterns for designing MCP servers well.Agent skill · MITKey takeaways
- Prompts are for now, project instructions for always, skills for “when this kind of job comes up”.
- MCP servers add access; skills add method. Many workflows need both.
- Subagents protect the main context and enable parallel work.
- Plugins package everything else for sharing.
Put it into practice
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Written by the AI Tool Garage editors. Tool details link to their catalogue pages, which are refreshed weekly; spotted something out of date? Tell us.