Skills (Agent Skills)
Skills are clearly defined capabilities that are additionally given to an AI assistant, such as checking an invoice or entering an appointment. Instead of guessing everything itself, the program calls up the appropriate instructions along with the necessary tools for a given task.
A text program like ChatGPT is inherently good at one thing above all: processing language and responding. It doesn’t know how a vacation request works at your company, and it can’t rename a file. Skills close exactly this gap. A skill is a self-contained package of instructions, examples, and access to a tool that is additionally given to the program. You can think of it like the handbook for a new employee: they bring the basic knowledge with them, but look up the specific procedure at the company. The term appears especially in connection with programs that carry out tasks independently in multiple steps.
Why assistants without skills reach their limits
A language model has a limited working memory for instructions. This area is called the context window and only holds a certain amount of text. If you tried to write all of a company’s rules into it at once, it would be full before the actual question even arrives. Skills solve this by loading only the instructions that are actually needed at the moment.
The second reason is reliability. Without fixed instructions, a model reinvents the procedure every time. Sometimes it asks for the cost center, sometimes it forgets to. A skill sets the procedure down in writing, so the same task produces a similar result each time. For companies, this is often more important than sheer cleverness.
There’s also a practical point: skills are interchangeable. A department can write down its process as a skill without touching the model itself. A new, better model simply continues to use the existing skills.
From trigger to tool call
A skill usually consists of three parts. First, a brief description of when it applies. Second, a detailed set of instructions with rules and examples. Third, often permission to use a specific tool, such as a search in the company database or a small program.
The process then looks like this: the user writes a request. The assistant compares it against the short descriptions of all available skills. If one matches, its full instructions are loaded and added to the prompt. Only then does the actual work begin. This step-by-step loading is also called progressive disclosure.
It’s important to distinguish this from two related terms. A tool is merely the technical interface, i.e. something like the command to send an email. A skill additionally contains the knowledge of when and how to do that. And unlike fine-tuning, where the model itself is retrained, a skill doesn’t change anything about the model. It sits alongside it and can be deleted at any time.
Skills in products and in the media
Anthropic made the term well known with Agent Skills for Claude. There, a skill is simply a folder with a text file and, optionally, a few helper programs. Similar concepts exist at other providers under names like Custom GPTs, Actions, or Plugins. The underlying idea is the same everywhere.
In everyday life, you usually encounter this invisibly. When a chatbot in an online shop tells you your order’s delivery status, there’s a skill behind it with access to the shipping system. Coding assistants like Cursor or Copilot use skills to stick to the style of a particular project.
In business news, skills come up when it’s about marketplaces and dependencies. Anyone who has written many of their own skills for one provider doesn’t switch to another easily. Another common misconception is assuming skills are inherently safe. A skill can trigger real actions, such as transferring money or deleting data. That’s why companies carefully review exactly what permissions an individual skill is given.