There is nothing magical about Skills

They are to AI agents what recipes are to cooks.

Before embedded agents like Codex or Claude Code became so popular, personalized assistants were the best way to feed your AI the context it needed for recurring tasks. If these terms are new to you, I wrote about them in the AI ladder.

Then, skills became a thing.

At first, they appeared hidden behind commands or plugins or wizards. You had to install or download a skill.

Very black-box-y, making them sound much more technical than they really are.

In practice, they’re easy to create.

Whenever you iterate on a task with your agent and arrive at a format, structure, pattern or workflow that works, you can ask the agent to save that process as a skill.

All the agent will do is create a text file that contains things like:

• its purpose and goal

• when the agent should use it (the trigger)

• what inputs it should look for

• the rules it should follow

• the actual steps of the workflow

• the expected output format

• a few examples or checklists

Pretty basic right? Because it is.

In more corporate words, think of it as your Standard Operating Procedure for that task or process.

And it does not need to be perfect.

Over time, you can improve it, add examples, include edge cases, include more rules, or, on the contrary loosen rules, remove vague instructions, restrict the format further, etc.

But the starting point is simple: when AI finally does the thing right, do not just keep the result. Say:

Save this a a skill.