Agent skills
This blogpost promotes the use of agent skills in an agentic workload system.
With agent rules, we have an established way to guide our coding agents towards predictable outcomes. Over time, rules can increase in size and complexity, polluting our context window with information the agent might not always need. Remember that rules are enforced universally, regardless of the context of the prompt.
This is where agent skills come to the rescue. They are reusable, packaged capabilities that a coding agent can call upon to perform specific tasks...well, actually they are plain old markdown files! For example, at skateparks.gr I have set up a dedicated minitest skill that my tester subagent can invoke when needed.
The important thing here is that skills are not loaded by default into our context window. The agent reads the title and description of each skill and decides if it needs to use it to complete the task at hand.
TL;DR: Keep your rules lean and universal, and let your agent reach for the right skill only when it's needed.
