What are skills,
and how do they work?

The 30-second version

For most of AI's short history, getting a model to do good work meant one of two things: write a really clever prompt, or pay someone to fine-tune the model. Both are exhausting, and neither sticks — your clever prompt vanishes into chat history the moment you close the tab.

A skill is the third option. It's a small, reusable file that tells an AI how to do a specific job well — written once, used forever, and shared with humans or other agents. Think of it less like a prompt and more like an app for your AI: give your assistant a skill called "respond to angry customer email," and from then on, it knows your tone, your refund policy, and your escalation rules without being asked.

What a skill actually is

Strip away the jargon and a skill is just a text file. Usually Markdown. It has a name, a short description, some instructions, and often a few "fill-in-the-blank" slots called {{variables}}. That's it. No code. No infrastructure. No special format.

Here's a real one — short enough to read in 20 seconds:

That's a complete, working skill. An AI agent given this file will produce dramatically better, more consistent customer responses than the same model with a generic prompt — because it now knows the shape of the job, not just the goal.

Why skills exist now

Skills aren't a new idea so much as the natural answer to three things that all happened in the last 18 months. None of them was a single product launch — they're shifts in how AI actually works.

1. Models stopped being the bottleneck

For a while, the difference between a useful AI app and a useless one was which model you used. That gap has mostly closed. The frontier labs all ship comparable models, and the interesting questions moved one layer up: what do you tell the model to do, and how do you organize that knowledge so it's reusable?

2. Context windows got huge — but you still can't paste your whole company in

Modern models can read hundreds of pages at once. That sounds like it solves everything until you try to use it: stuffing every document you own into every conversation is slow, expensive, and usually worse, because the model loses the plot. You need a way to load just the right context for the job at hand. That's a skill.

3. Agents started doing real work in loops

The new generation of AI tools — Claude Code, Cursor, Devin, ChatGPT's agent mode — don't just answer one question. They run for minutes or hours, taking dozens of steps, opening files, calling tools, course-correcting. An agent doing a 40-step task can't be guided by a 200-word system prompt. It needs reference material it can pull up when the situation calls for it. That's a skill.

Skills are what you write when "tell the AI in chat" stops scaling and "fine-tune a model" is overkill.

The three eras of working with AI

It helps to see where this fits. Each era didn't replace the last — it added a new layer.

That third layer is what people mean by "skills." It's the part that's yours — the part that doesn't reset when the model changes or the harness is upgraded.

The "context filesystem"

Here's the picture that makes it all click. Imagine your AI assistant has, sitting next to it, a small folder of reference cards. Each card is a skill — a recipe for one kind of task. The AI doesn't read every card before every reply (that would be exhausting and expensive). Instead, when a request comes in, it scans the names and descriptions of the cards and pulls only the one or two it needs.

That folder is the "context filesystem." It sits between you and the model, and it's portable — the same library can serve a coding agent in Cursor on Monday and a writing assistant in Claude on Tuesday. The skills are yours; the harnesses are interchangeable.

How agents find and load skills

This part feels like magic until you see the steps. There's no special protocol — agents pick skills the same way a person picks a book off a shelf:

That's the whole mechanism. The cleverness isn't in the discovery — it's in the writing. A skill with a clear name and a sharp description gets picked up at the right moment; a vague one gets ignored.

Skills vs. prompts vs. system prompts vs. RAG

These words get used interchangeably and it makes everything harder to learn. Here's a clean cut:

Why portability matters

Here's the catch nobody warned you about: every AI tool implements skills slightly differently. Same idea, different file shapes, different folder names, different conventions for how variables are written.

The idea is the same in all of them: a reusable text file with instructions for a job. But if you write a great skill in one tool, you currently can't drop it into another without some surgery — different headers, different variable syntax, different folder conventions.

This is the same place the early web was in before HTML standardized. The instinct says "just pick one tool" — but in practice, most people use two or three (a coding agent, a writing assistant, something on their phone), and your best skills should follow you across all of them.

Bringing it home

That's the whole picture. A skill is a small Markdown file. A skill library is the new layer of working with AI. And the trick to actually using one is keeping it portable across whatever tools you'll be using a year from now.