By now you have built context. The AI reads who you are before every task. That solves a huge part of the problem, the part where it forgot you. But there is a second kind of repeating you do that context alone does not fix, and this chapter is about that.
Context is who you are. A skill is how you do a specific job.
Let me draw the line clearly, because people mix these up. Context is standing facts: my customers are these people, my voice is this, my rates are that. A skill is a procedure: here is how I turn a rough idea into a finished thing, step by step, with all my particular rules about how it has to come out. Context tells the AI who it is working for. A skill tells it how to do one specific kind of work the way you do it.
Here is the tell that you need a skill, and it is a feeling you already know. You have handed the AI a task, gotten a result, and given it the same five corrections you gave it last time. Move that to the top. Cut the intro, you always write too much intro. Keep it under this length. Link it to that. And then next time, the same five corrections again, because the AI does not carry your corrections forward on its own. Those five corrections are not really corrections. They are a procedure you have never written down. That procedure is a skill waiting to exist.
What a skill actually is
A skill is a written-down procedure plus the house rules for doing it your way. It is a set of instructions the AI can follow, saved and reusable, so that instead of re-teaching the steps and re-giving the corrections every time, you invoke the skill and the steps and the rules come with it.
The key word is repeatable. A skill pays off when the work has a fixed shape you do again and again. Turning research into a summary in your format. Turning a rough idea into a structured outline with your constraints. Producing a monthly document that always has to look a certain way. Anything where the shape is stable and only the input changes, that is skill territory. If a task is genuinely one-off, something you will never do again in this form, a skill is overkill, just do it. The question is not "is this complex," it is "will I do this shape of work again." If yes, the effort of writing the skill once buys you every future run for free.
Building your first one
You do not write a skill by sitting down to write a skill. You write it by noticing you already have one, unwritten, living in the corrections you keep repeating.
Take the outline example. Say you turn rough ideas into structured outlines all the time, and every time you catch yourself saying the same things. Start with a strong opening but not a gimmick. Keep the whole thing to this length. Structure it in these sections. Never use these three words I hate. Always end by connecting it back to the larger theme. Those instructions are the skill. You write them down, once, as the procedure for "make an outline my way," and from then on you invoke that instead of re-listing the rules. The AI reads the procedure, follows the steps, applies the rules, and hands you something that already respects the corrections you used to give by hand.
That is the whole arc of a skill. Notice the repeated corrections. Write them down as a procedure. Stop giving them by hand. The job changes from "explain and correct, explain and correct" to "invoke and review."
Let me show you two real shapes of this.
Example: the content skill [from my own system]. Turning a rough idea into a publish-ready outline is a job I do constantly, and it has real, specific rules. Title length limits. Description length limits. Which larger theme the piece has to connect to. Tone constraints. Which of my context files it should consult before writing, so it knows the audience and the voice. Early on I specified all of that in the prompt every time, or worse, accepted output that broke the rules and fixed it by hand.
Now that whole procedure is a skill. Read the brand and audience context, apply my structural rules, follow the post structures I know work, produce the outline. I invoke it in one step and the rules travel with it. I am describing the shape, not handing you the file, because the shape is the lesson. What changed is that the house rules stopped being something I police and became something the skill guarantees. I do not check whether the title is the right length anymore. It is the right length, because the skill will not produce it otherwise.
Example: the monthly invoice skill. Same invoicing you do every month, same format, same tax lines, same way the files have to be named. The naive way is re-explaining the format each month and re-checking that it came out right. The built way is a small skill that knows the format, reads your hours file, and produces the invoices named correctly. What changed is that a recurring chore became a one-line request, and the format is guaranteed instead of re-verified.
The bridge from skills to a team
Notice what just happened across these last two chapters. You gave the AI standing knowledge of who you are, context. Then you gave it standing procedures for how you work, skills. Put those together and you have something new. You have an AI that knows your world and knows your methods. That is no longer a chat assistant. That is starting to look like a colleague.
And once you have a colleague who knows your world and your methods, the natural next question is whether you can give it a whole job to run, not just a task to answer. That is an agent, and it is the next chapter.