Agent Skill.

An AI agent can do a lot, but it knows neither your processes nor your templates. Until recently people relied on long prompts pasted in again for every task. Agent skills solve this in a more structured way: each skill is a folder of its own. It contains a text file with a short header (name and a description of when the skill applies) followed by the actual instructions. Scripts, example files, fonts or checklists can sit alongside. Anthropic introduced the format for Claude in October 2025, and other agent tools now read skills built the same way.

The trick is loading in stages. At the start the agent only sees the name and description of every available skill, which is a few lines. Only when a task matches does it read the full instructions and, if needed, the accompanying files. This keeps the context window free for the actual work even when dozens of skills are installed. A skill is therefore not a new capability of the model but retrievable know-how plus tooling: 'This is how we do it here, and this script makes it faster.'

Skills should be distinguished from two related building blocks. Project files such as CLAUDE.md and AGENTS.md describe a single project and are always loaded. The Model Context Protocol connects an agent to external systems such as a CRM or a database. Skills sit in between: they are reusable across projects and describe how a task is done, not where data comes from. In practice all three work together.

Skills are code and instructions at once, so the same care applies as with any software. A skill from someone else's repository can run scripts or steer the agent towards actions you do not want. The risk resembles prompt injection: instructions in a file are taken seriously by the agent. Before using one, read what a skill does, check the licence (see open-source licence) and only use sources you trust.