What a Skill is
Your CLAUDE.md sets rules that apply to everything in a project. But some jobs are not always-on. They are recurring: the way you write a weekly update, the steps you follow to prep a client call, the exact shape of a good invoice. You do not want those rules cluttering every conversation. You want them to show up only when the job comes up. That is a Skill.
A Skill is a job you teach the AI once so it can do it the same way every time, without you re-explaining the steps. Where CLAUDE.md is the standing law, a Skill is a saved recipe the model reaches for when the moment matches. Teach it once, reuse it forever.
The cleanest way to hold the difference: CLAUDE.md is pushed at the model, always present in the background. A Skill is pulled in, only when it is relevant. You would not want every recipe in your kitchen taped to the fridge door at once. You want them filed, and the right one in your hand when you are cooking that dish.
Here is the part nobody tells beginners. You do not write a skill by hand. You build it by talking. Claude has a skill that helps you make skills, and it interviews you: what is the job, when should it trigger, what are the steps. You answer in plain words and it writes the recipe with you.
You do not need to know the format. You just describe the job clearly. Here is the kind of thing you say to start, for a consultant who writes the same style of weekly client update every Friday.
One quiet lesson worth keeping from how skills work: the description of when a skill should trigger matters as much as the steps inside it. A skill with a vague trigger never fires when you need it. A clear trigger (the exact phrase, the exact situation) is what lets the model reach for the right recipe at the right time. Good descriptions are not decoration. They are what makes the thing reliable.
Skills in this exact form are one of the more Claude-specific things in the course, so here is the honest portability note. The other tools have cousins, not copies: ChatGPT has custom GPTs and Gemini has Gems, both reusable configured assistants you set up once. They are not identical to a pull-in-when-needed skill, but the instinct transfers: spot a job you repeat, teach it once, reach for it by name. If you move tools, you keep the instinct and learn the local version.
Pick one job you genuinely repeat: a weekly update, a standard reply to a common question, the way you outline a video, your invoice format. Ask Claude to help you make a skill for it. Let it interview you. Name a clear trigger phrase and the steps you actually follow. Then test it on a real instance and fix one thing.
Show the worked solution
- A Skill is a recurring job taught once, so the model does it the same way every time without you re-explaining the steps.
- CLAUDE.md is pushed (always on); a Skill is pulled in only when the job appears.
- You build a skill by talking. Claude interviews you about the job, the trigger, and the steps, then drafts the recipe with you.
- The trigger description matters as much as the steps. A clear trigger is what makes a skill fire reliably.
Find one job inside your thread project that you already repeat, and turn it into your first skill. Now your project has three layers: a constitution (who you are), a CLAUDE.md (how this project behaves), and a skill (a job it can do on command). That is a real system, built by you, in plain language. Next: how to hand it bigger work and stay in control.
You used to re-explain the same job every week. Now you teach it once and reach for it by name. That is the difference between using a tool and owning one.