What is a Skill?

If you’ve searched “what is a skill,” you probably just heard the word in a video or a thread about AI and weren’t sure what it meant. A skill is a capability you add to an AI assistant (like Claude) so it knows how to perform a specific task without you having to explain it every time. They’re packaged instructions, and sometimes additional files, that the assistant uses when the conversation calls for it.

You don’t need to know how to code to follow this post. All you need is to have used an AI assistant like ChatGPT or Claude via chat and asked it things by typing.

What is a skill, in one sentence?

A skill is a saved recipe that the assistant pulls out when it fits. Imagine you have a new coworker. The first week you explain in detail how your company writes reports: what sections go in them, the order, the tone. By the second week you don’t repeat it. They know. What they learned and saved is, more or less, a skill.

The same thing happens with an AI assistant. Without a skill, every time you want it to do a task with its own specific rules, you have to rewrite all the instructions. With a skill, those instructions live saved somewhere and the assistant retrieves them when it detects it needs them.

The key word is: reusable. A skill isn’t something you write that then disappears. It’s a capability that stays there, available, ready for the next time you need it.

How does a skill work internally?

In Claude’s case, a skill is literally a folder with an instructions file inside. Anthropic, the company that makes Claude, calls this feature Agent Skills. The folder contains a file called SKILL.md with two things: a name and a short description (what the skill does), followed by detailed instructions on how to do the task. Optionally it can include other files: small programs (scripts) that the assistant runs, or reference documents it consults only if it needs them.

Here’s the part that’s hardest to understand at first. The assistant doesn’t use all skills at once. First it only reads the name and description of each one. When you ask it something, it looks at those descriptions and decides which one fits what you’ve asked. Only then does it open the full file for that skill and read the detailed instructions.

This has a name: activation is probabilistic. Claude reads the skill’s description and decides if the task fits, based on the conversation context. It’s not a rigid rule that always triggers. It’s more like a judgment: “what they’re asking me to do looks like what this skill does, I’m going to use it”.

# A minimal SKILL.md: header (name + description) and instructions
---
name: redactar-informe-mensual
description: Redacta el informe mensual del equipo con nuestras secciones y tono.
---

# Cómo redactar el informe mensual

1. Empieza siempre por un resumen de 3 frases.
2. Usa estas secciones en este orden: Logros, Bloqueos, Próximos pasos.
3. Tono directo, sin adjetivos de relleno.

If you look closely, there’s no complicated code. It’s instructions in plain language, the same you’d give a person. That’s the idea: a skill is written knowledge that the assistant reads when it’s its turn.

That’s why a skill saves work. The description is the hint that tells the assistant when to look inside. The instructions inside tell it how to do the task well.

”Skill” doesn’t always mean the same thing

Here’s where the confusion comes in, because the word “skill” is used for several similar but different things. It’s worth separating them, because depending on where you read it, it means one or the other.

The meaning this post is about is the one for AI assistants: a capability you add to the assistant, packaged as instructions. In Claude it’s called Agent Skill. ChatGPT has also incorporated a feature called Skills (currently in beta and in its paid plans for businesses and teams), with the same underlying idea: turn a workflow you repeat into saved instructions that the assistant applies when they’re relevant. And in programming tools like Claude Code, a skill is this same idea applied to coding tasks: saved instructions for a specific development task, not a separate concept. I’ll clarify later how this fits with ChatGPT’s GPTs, which is a different thing.

There’s also the everyday sense of the word. In English, “skill” means “ability” or “competence.” When someone says “you need to improve your communication skills,” there’s no AI assistant involved. It’s the human ability you’ve always known. It’s the same word, but it has nothing to do with the technical feature.

And there’s a third place where you’ve probably heard “skill”: voice assistants like Alexa. An Alexa skill is a kind of mini voice app that adds a new capability (ordering a taxi, telling a joke, controlling lights). It shares the idea of “capability you add,” but it’s an Amazon product, before this wave of AI assistants, and it works differently.

If you were looking for the first meaning, let’s go on. What confuses beginners most isn’t the word itself: it’s mixing up a skill with other things that look similar.

A skill is not the same as a prompt, an agent, or an MCP

These three confusions are the ones people repeat most when they start with AI assistants. Let’s go through them one by one, at a basic level.

Skill versus prompt

A prompt is what you write to the assistant in the moment: your message, your request. “Summarize this text for me,” “write me an email to cancel a booking.” It’s a loose instruction that works for that conversation and then is gone.

The difference with a skill is permanence. The prompt is disposable. The skill is the saved, reusable version of instructions that otherwise you’d have to repeat in the prompt over and over. In other words: when you notice you’ve spent weeks pasting the same block of instructions into every chat, that block is begging to become a skill.

Skill versus agent

An agent is an AI assistant that doesn’t just answer you: it decides on steps and executes them on its own to meet an objective. You tell it “sort these files by date” and the agent thinks about what to do, does it, checks the result, and moves on, without you marking each step.

The relationship with the skill is about who commands and who obeys. The agent is the one that decides and acts. The skill is one of the capabilities the agent can use when the time comes. The agent is the chef; the skill is one of the recipes in their notebook. They don’t compete: they work together.

Skill versus MCP

An MCP (Model Context Protocol) is the standard “cable” that connects an AI assistant to external tools and data: your database, your calendar, your code repository. It’s like the USB of assistants: a single connector that plugs in many different tools without inventing a new connection for each one. If you want to understand it well, I explain it step by step in what is an MCP and what’s it for.

The difference with a skill? The skill is knowledge: instructions on how to do something well. The MCP is plumbing: the connection that gives the assistant access to an external tool. A skill can tell the assistant “for this task, check the sales database”; the MCP is what makes that database access possible. One explains the what and the how. The other opens the door.

Concept	What it is	In one sentence
Prompt	Your message in the moment	A loose instruction you write and use up in that conversation
Skill	Saved capability	Reusable instructions that the assistant applies when the task fits
Agent	Who decides and acts	An assistant that executes steps on its own to meet an objective, and can use skills
MCP	The standard connection to tools	The “cable” that plugs the assistant into external data and tools

What is it for and when to use a skill?

A skill fits when you repeat a task with its own rules and want it to turn out the same way every time: drafting a specific type of document, applying a process your team has defined, reviewing a PR with your checklist, formatting data the way your report needs it. If you do it just once, it’s not worth it. If you’re going to ask for it twenty times, saving it as a skill saves you from rewriting the instructions each time. And if you want to learn how to build agents that combine skills and tools, the course Agentic Patterns works through it step by step.

Now, “when to use a skill” has nuances, especially if you work with Claude Code, Claude’s version for programming. There, a skill isn’t the only option: there are other tools to configure the assistant’s behavior, and choosing wrong makes things work sometimes. That full decision, with all the alternatives and when each one fits, you’ll find in the guide on when to use skills in Claude Code. Here I’ll stick with the what; there’s the when.

What’s worth remembering is the danger of probabilistic activation. Because the assistant decides whether to use the skill based on the description, a vague description makes it not activate when it should, or activate when it shouldn’t. From experience, that’s exactly what gets neglected most when creating a skill: you pour yourself into the instructions inside and dash off the description in a weak line. Then I’ve run into skills that were perfect but the assistant wouldn’t pull them, because their description didn’t clearly tell it when to step in. If you ever create a skill, that description is what deserves the most care.

Frequently asked questions

Do I need to know how to code to use a skill?

To use skills that already exist, no. A skill’s instructions are written in plain language, the same you’d give a person. Creating a skill from scratch in some environments does require knowing how to work with files and folders, but understanding what it is and taking advantage of the ones already made doesn’t require coding.

Is a skill in Claude the same as in ChatGPT?

The idea is very similar: in both, a skill turns a workflow you repeat into saved instructions that the assistant applies when they’re relevant, and both even use a SKILL.md file with a name and description. Anthropic calls them Agent Skills in Claude; OpenAI launched its Skills feature in ChatGPT later and, for now, in beta and mostly in its paid plans for businesses and teams. Watch out not to mix up ChatGPT’s Skills with GPTs, which are a different and earlier feature: a GPT is a customized version of ChatGPT itself, and its “actions” are connections to external services, closer in spirit to an MCP than a skill.

Where do I find skills that are already made?

Because a skill is ultimately a folder with files, it’s shared like any file: people publish theirs in public repositories and on sites that collect community skills. For Claude, the official starting point is Anthropic’s repository github.com/anthropics/skills, with examples and a template for creating your own; there are also community catalogs where other users upload the ones they create. Before using one, it’s good to open its SKILL.md and read what it does, just like you wouldn’t run some random file without looking at it first.