AI Skills: Stop Teaching the Agent the Same Job Twice

I’m slightly embarrassed by how recently I properly got around to using AI skills.

I had seen people talking about them for ages. I knew they were meant to be useful. But, like quite a few things in AI, I assumed there must be more technical complexity involved than I was seeing.

There really isn't that much.

At its simplest, a skill packages up how you want an AI agent to perform a repeatable job so you don't have to explain the process again every time.

If you regularly find yourself pasting the same instructions into ChatGPT, recreating the same role, or correcting the same behaviour, you probably already have the beginnings of a skill.

Prompt, project guidance or skill?

The distinction that helped it click for me is:

A prompt is what I want you to do this time.

Project guidance is what you need to know about this particular project.

A skill is how I want you to perform this kind of job whenever it comes up.

That last one was the bit I had effectively been recreating manually.

For example, I would often tell an LLM to take on a particular role, explain how I wanted it to question me, what I wanted it to challenge, and how the final output should be structured.

That's already very close to defining a reusable workflow.

Instead of rebuilding that behaviour each time, a skill lets you package it once and call on it again.

A skill can be simple without being simplistic

In current agent systems, a skill is commonly built around a Markdown instruction file — often called SKILL.md — describing what the skill does, what inputs it expects, the steps it should follow and what a good output looks like.

But it doesn't have to stop there. A skill can also include supporting examples, templates, reference material, scripts or tool instructions where the job requires them.

So saying a skill is “just an .md file” isn't quite technically right.

But for a beginner, the important point remains: the core behaviour can be described in normal written language.

That makes the barrier to creating your first one remarkably low.

Start with something you already repeat

I wouldn't begin by trying to design your perfect library of AI workers.

Pick something you already do regularly.

Maybe you want an agent that always reviews a piece of writing in a particular way.

Maybe you want one that interviews you thoroughly before producing an article.

Maybe you regularly analyse meeting notes, review code, turn research into a certain format or challenge an idea from the same set of angles.

Write down how you want that job done.

Use the skill.

Then improve it when the behaviour annoys you.

That last part matters. My own skills have needed touching up as I've used them. Instructions that sounded sensible on paper didn't always create the behaviour I actually wanted.

But that refinement is an investment. Once you fix the workflow, you benefit every time you reuse it.

Skills need maintenance too

I can already see another problem coming: skill clutter.

If you create enough of them, some will overlap. Some will become stale. Others will be replaced by better ways of working.

My current plan is to keep my own skills in a GitHub repository so I can review and version them over time.

Eventually, there is nothing stopping an agent helping with that maintenance too: analyse the library, identify duplication, flag outdated instructions and help work out which skills are still earning their place.

That feels like the natural lifecycle:

notice repeated work → turn it into a skill → use it → refine it → periodically clean up the library.

Just make one

If skills still sound like something for people who are far deeper into AI than you are, I would genuinely just make one.

Ask the AI system you use how skills work in that environment. Then tell it the repeatable task you want to package and ask it to help you create one.

The first skill doesn't even have to be important.

The point is to see how straightforward the idea actually is.

For me, that was the surprising bit.

Skills sound like another technical layer of AI.

In practice, much of the value comes from something much simpler: writing down a good way of working once, then not having to teach it again.