Skill file

Skill file

A skill file is a text document with instructions that an AI assistant opens and follows when needed. This makes it possible to teach an assistant new workflows without retraining it.

A skill file is a text document containing instructions for a computer program that understands and responds to language. Such programs are called assistants or chatbots. The file describes, in plain sentences, how a particular task is to be carried out. One example: “Here’s how to create a monthly report from our sales figures.” The assistant reads this instruction when it needs it and works through it. You can picture it like a recipe book next to the stove: it’s only opened when that particular dish is on the menu.

Adding capabilities without retraining

An AI model learns its knowledge during training. That’s a computing process that takes weeks and costs millions. Company-specific knowledge doesn’t appear in it: internal abbreviations, in-house forms, the sequence of steps in an invoice review. This exact kind of knowledge can be supplied afterward via skill files. Anyone who can write a text file can use it to teach an assistant a new task.

The second advantage is traceability. Anyone can read and change what’s written in the file. If the assistant does something wrong, you correct a sentence in the instructions. With a trained model that’s impossible, because no one can single out individual rules there. Skill files can also be versioned like program code, meaning they can be saved with a change history and shared within a team.

It’s important to distinguish this from a prompt. A prompt is the single question you’re typing right now. A skill file, by contrast, is permanently available and applies to all matching requests. So it’s less a question than a standing instruction.

The assistant decides when to consult it

At the start of every skill file is a short description. It states the name of the skill and says when it should be applied. Only these few lines are constantly visible to the assistant. The long remainder of the instructions stays unread at first.

If a request matches a description, the assistant loads the full file. Experts call this progressive disclosure. The reason for this is the limited working memory of such systems, known as the context window. Only a certain amount of text fits in there. If a hundred instructions were carried along permanently, there would be no room left for the actual conversation.

A skill file can also reference additional files. Typical examples are sample documents, templates, or small programs that the assistant is allowed to run. A skill for spreadsheets, for instance, might come with a ready-made script for converting numbers. The assistant then doesn’t have to invent the calculation method itself, which significantly reduces errors.

Skills in chatbots and enterprise software

The format became well known in 2025 through Anthropic, the maker of the assistant Claude. There, the files are called Agent Skills and usually carry the filename SKILL.md. Other providers have similar concepts, for example under the name Custom Instructions or project instructions. The shared idea is the same everywhere: knowledge lives in readable text files instead of in the model.

In companies, skill files are most commonly found where the same processes recur again and again. A legal department stores how a contract is to be reviewed. A marketing department specifies the tone in which texts must be written. Employees don’t notice any of this; they simply ask their question.

A common misconception is that a skill file makes the assistant smarter. It doesn’t. It merely hands it instructions and material. If the instructions are imprecise or contradictory, the assistant works just as imprecisely. And because skills are also allowed to launch programs, caution is warranted with unfamiliar files from the internet.

Subscribe free. Unsubscribe the second it sucks.

High-signal news across AI, business, UX, and tech. Every morning.