A lot of people I talk to are struggling with the new Skills feature that ChatGPT and Claude have rolled out. They have read the announcements. They have watched the demos. They still walk away unsure of what a skill actually is and why the feature is supposed to matter. The thing they keep getting hung up on is the word itself.
I have come to think the word is the whole problem.
What People Actually Mean by "Skill"
When you say someone has a skill, you mean a specific thing. Riding a bike. Playing the guitar. Swimming. Touch typing. Driving a car. Knife technique in the kitchen. The way a tennis player's body moves before they consciously decide what shot to hit. The way an experienced salesperson reads a room. The way a doctor recognizes a familiar pattern in a patient's symptoms before the labs come back. You built the skill through repetition over a long period of time. You do not reference anything when you use it. It fires the moment the situation calls for it. You can forget about it for fifteen years, pick up a guitar at a friend's house, and your fingers will find a chord you have not played since high school. That is what a skill is. It lives in you.
Humans have another kind of know-how, and we do not call it a skill. We call it a reference. We have the ability to recall information from outside our head. A playbook. The chapter in a book you remember reading. Grandma's recipe card in the kitchen drawer. The notebook you used for a class five years ago. The napkin you jotted a thought down on while you were at the bar. The information was captured once, lives outside your head, and you walk to the shelf when you need it. You do not pretend to know the recipe from memory. You open the card and follow the steps. It is useful, often essential, but nobody would call that a skill. Nobody says "I have the skill of looking up grandma's pie recipe." They say they know where the recipe is.
The distinction is clean in everyday language. Skills are one thing. References are a different thing. We do not confuse them when we are talking about ourselves.
What Got Shipped
Look at the feature that ChatGPT and Claude both call Skills. A skill in this context is a short text file. It has a name. It has a description that tells the AI when this procedure is relevant. It has steps to follow. Sometimes it has supporting materials in the same folder, like templates or examples. That is the entire format. The file sits on a shelf with the rest of the AI's skills library. The AI grabs it when the moment calls for it, reads the steps, and follows them.
That is not a skill in the way humans use the word. That is a reference.
The AI companies took a word that means "something wired into you after long periods of practice" and applied it to "a procedure you keep on a shelf and look up when needed." Both kinds of know-how are useful, and both have a place in how work gets done. They are not the same thing, though, and calling them the same thing is the reason this feature feels confusing the first time you encounter it.
Why the Naming Causes Real Confusion
When you say "Claude got a new skill," people picture the bike. They imagine the AI internalizing a capability the way a person internalizes the feel of a perfectly seasoned pan. They expect the AI to know the thing now, the way you know how to swim once you have learned. That is not what happened.
To see why, look at how a language model is built. The model has two layers of know-how. The first layer is its training. Everything the model learned from the data it was trained on lives in the model itself: the ability to write a clear email, summarize a document, follow instructions, reason about a problem. You could fairly call that layer the model's actual skills. It is the embodied stuff, built through long periods of training, that the model already had before you typed anything in. It comes forward fast because it does not have to be looked up. Those are the model's real skills.
The second layer is the skills library, which is the part the AI companies named "Skills." This is where the misnaming gets concrete. These are not skills in any human sense of the word. They are references. Recipe cards on the shelf. The model does not internalize them. It looks at the names and descriptions, decides which one applies to what you asked for, opens the right card, and follows the steps. The references do not replace the training. They direct it at a specific dish for a specific occasion.
In humans, those two layers are fluid. We move back and forth across them all the time. Cook grandma's pie enough times and you stop pulling the card. The recipe becomes part of you. Go ten years without making it and the embodied version fades back. You walk to the shelf for the card again. Knowledge becomes skill, skill becomes knowledge, and the line keeps moving over a lifetime.
AI does not work this way right now. The model's training is fixed at a moment in time. Once a model ships, what it knows is what it knows. The recipes on the shelf stay recipes. No matter how many times the AI reaches for the same skill file, it never internalizes it. There is no quiet promotion from the shelf into the model. The next conversation, the AI walks back to the shelf and reads the card again. There is no fading on the other side either, because nothing was wired in to begin with.
Maybe future systems will close this gap. Models that learn from the skills they reached for most. Training runs that fold an organization's most-used recipes into the next version's baseline capability. That is not where we are today. The training stays fixed. The skills library stays external. The two never blend.
What This Means in Practice
If you have been trying to understand AI Skills through the lens of human skills, you have been holding a mental model that does not match what the feature actually does. That is not your fault. The word was chosen poorly. The right mental model is the recipe.
A skill is a recipe card for your AI. A team writes down how they do something well. The card goes on a shelf. The AI pulls it when the moment calls for it. The work comes out the way the team does it. The model is still the cook. The model's training is still the general kitchen know-how that makes a cook a cook. The skill is just the card that tells the cook which dish to make tonight, in your house, the way your family makes it.
That reframe makes a lot of practical questions easier. You can see why a skill works even though it is just a text file. The model already knew how to cook. The card just told it what to cook tonight. You can see why a recipe with outdated steps will produce outdated food, which is why keeping the cards current matters more than people expect. It also explains why the AI companies keep warning people about installing skills from public registries. Following a recipe from a stranger that calls for weird ingredients is a bad idea. You would not do that for dinner. You should not do it with your AI either.
Working With the Wrong Word
The naming is not going to change. ChatGPT and Claude are not going to rename the feature, and Anthropic's open standard already calls it Agent Skills. The word is in the air now.
That is fine. You do not need to fight the name. You just need to know what is actually under it. Once you swap "skill" for "recipe" in your head, the whole feature gets simple. The mystery evaporates. The practical questions get easier. Should we write one of these for our team? What goes in it? Who maintains it? Where does it live? Who else can use it? Those are sensible questions to ask about a recipe. They are confusing questions to ask about a skill in the bike-riding sense, because skills in that sense are not something you write down and hand to someone else.
The word will be wrong for a while. The feature is still useful. Treat it as what it actually is, and the rest of it gets easier.