Skills vs MCP vs Subagents vs Prompts vs Hooks

You installed a skill, and now you're wondering: how is this different from an MCP server? A subagent? Just a good prompt? Short answer — they solve different problems and compose together.

One-line version: A prompt is what you say once. A skill is reusable know-how the model loads when relevant. An MCP server gives the model new actions and live data. A subagent is a separate worker with its own context. A hook is automation that fires on an event. A skill needs no API, no server, no code — it's just a structured instruction file.

The 60-second comparison

	What it is	Gives the model…	Needs code/server?	Best for
Prompt	Text you send	Instructions for this turn	No	One-off asks; global persona
Skill (`SKILL.md`)	A structured file the model loads on demand	Reusable expertise — how a pro produces a specific artifact	No	Repeatable structured deliverables (PRDs, postmortems, rubrics)
MCP server	A running service (Model Context Protocol)	New tools & live data (DB, email, GitHub)	Yes	Acting on real systems & data
Subagent	A separate agent with its own context	Parallelism + isolation	No	Big or separable sub-tasks
Hook	A script that runs on an event	Automation — before/after a tool, on prompt submit	A script	Guardrails, auto-formatting, auto-loading context

How to choose

Consistent, professional output for a recurring task? → Skill. (Every incident gets a proper postmortem.)
Read/write a real system — DB, GitHub, Slack, files? → MCP server. Pair it with a skill so structure comes from the skill and data from the server.
Just need this one thing, once? → Prompt. If you keep pasting the same prompt, promote it to a skill.
Task is huge / want it off your main context? → Subagent.
Want something automatic every time? → Hook.

They stack: a hook can auto-load a skill, which structures work that an MCP server grounds in live data, optionally delegated to a subagent.

Why skills are so cheap: progressive disclosure

A skill loads in two stages, which is why you can have hundreds installed without blowing up the context:

Always in context (tiny): just each skill's name + description — ~100 tokens each. This is how the model decides what's relevant (and why the description is the most important line in a SKILL.md).
Loaded on demand: the full body — a few thousand tokens — enters context only when that skill is picked.

So 207 skills cost almost nothing until one is actually used. Broad capability, narrow cost.

See it in practice

▶ Run a skill free 🧩 Chain skills into a workflow ⚡ Ground in 500+ apps (Composio) ✍️ Write your own skill

Full written version with links: docs/concepts.md.