SKILL.md

SKILL.md is a simple text file in which you store a set of working instructions for an AI assistant to carry out a particular task. The assistant only loads these instructions once it actually needs them, and then works through the task according to fixed guidelines.

SKILL.md is a text file containing instructions for an AI program. In it, you write down how a particular task is to be carried out: for example, how a company report should be structured or which steps are required when checking an invoice. The name is meant literally. The file is actually called that, and the .md extension stands for Markdown, a very plain way of writing text with headings and lists. You can open and read it in any text editor; no programming code is required. The concept comes from the company Anthropic, which introduced it in 2025 for its assistant Claude.

Why a simple text file replaces programming

An AI assistant can do many things in general, but few things in particular. It doesn’t know what your company’s proposal template looks like or which review steps your department requires. Until now, such knowledge either had to be typed into the chat anew each time or a dedicated program had to be written for it. Both are tedious, and the result was a little different each time.

A SKILL.md file solves this at a very low threshold. Anyone who can write instructions for a new colleague can also write a skill. This shifts who is actually allowed to build such extensions: no longer just the software department, but also accounting or the legal department. That is precisely what makes the term interesting for companies and for business news.

A second point is reliability. Because the instruction is fixed in a file, the assistant does the task the same way every time. The file can also be versioned like normal text, meaning changes can be tracked, and it can be shared within a team. A skill is thus less a clever trick and more a document that someone is responsible for.

The folder, the header lines, and step-by-step loading

A skill usually consists of a folder. Inside it is the file SKILL.md, and additional files can be included: sample documents, templates, or small helper programs. At the very top of the file are a few header lines with the name of the skill and a short description of what it’s for. Below that follows the actual instructions in plain sentences.

What matters is when the assistant reads all of this. At startup, it only knows the name and description of each available skill, that is, just a few lines. If a description matches your request, it loads the full text. Only once the instructions refer to another file is that file also opened. This step-by-step approach is called progressive disclosure.

The reason for this is a technical limit. A language model can only take in a limited amount of text at once per request, the so-called context window. If a hundred skills were loaded in full, the space would be used up before the actual work even began. A common misconception, by the way, is that a skill trains the model. That’s not true: the model remains unchanged; it merely reads along an instruction.

Skills in developer tools and everyday office work

Today, SKILL.md files are most commonly found in programming tools where an AI works alongside you within the project folder. There, skills exist for recurring tasks: the style of code review, the format of changelogs, the process of a release. Because the files live within the project, they automatically apply to everyone on the team.

Outside of development, skills are emerging for office tasks, such as generating presentations or spreadsheets in a company’s house format. Anthropic has made the format open, and other providers are adopting it. In news reports, the term usually appears in connection with AI agents, that is, programs that carry out multiple work steps on their own, one after another.

Skills should be distinguished from two neighboring concepts. A system prompt is a basic instruction that always applies; a skill only applies for a matching task. And tool connections via the Model Context Protocol give the assistant access to external systems such as a database. A skill, by contrast, mainly explains how to proceed. Caution is nevertheless warranted: anyone who installs a skill from someone else is bringing that person’s instructions into their assistant unchecked.

Subscribe free. Unsubscribe the second it sucks.

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