Find your first skill
Everyone is talking about AI agents right now. If you are just getting started, you do not need one. What you need is to stop re-explaining yourself every single time you open the thing.
Jump to the promptWhat a skill actually is
A skill is instructions you write down once so the AI does a task your way every time, instead of you explaining the whole thing from scratch on every attempt.
That is genuinely all it is. It is not an agent, it is not automation in the scary sense, and there is no code involved. If you have ever typed the same paragraph of context into a chat window for the third time and thought "I have said this before," you have already found the gap that a skill fills.
Different tools call it different things. Claude calls them skills, ChatGPT has custom GPTs and projects, and most tools let you save a set of instructions somewhere. The names change and the idea does not, so nothing below depends on which one you use.
The other reason this is worth your time: it is boring. Agents are the exciting thing to talk about and this is the thing that actually changed how I use AI day to day. Those are usually different things.
The three-part test
Not every task is worth teaching. Writing the instructions takes real effort, and if you only do the task twice a year you will never get that time back. A task is worth turning into a skill when all three of these are true.
You do it more than once a week
Repetition is what pays back the setup cost. Once a week is roughly the floor for me. Below that I forget the skill exists and just do the task by hand anyway, which tells you something.
The tell to listen for is the sentence "I have typed this before." Every skill I have built started with me noticing that sentence in my own head and then ignoring it about four more times.
You want the same result every time
This is the one people get backwards. A skill is for the tasks where consistency is the whole point, not the ones where you want something fresh.
If you want a different answer each time, that is a conversation and it should stay a conversation. If you want the same shape of output every time, with the same rules applied, that is a skill.
You already know what good looks like
This is the one almost nobody mentions and it is the one that matters most. You have to be able to look at the output and know whether it is right.
If you cannot judge it, you have not saved yourself work. You have just automated being wrong faster, and you will not find out for a while. Teach AI the tasks where you are already the expert, not the ones where you are hoping it knows better than you.
The one I actually built
Here is mine, and it is deliberately unglamorous.
Every post I make needs a cover image. I make several of these a week, and I want them to look like they came from the same person each time, which means the same font, the same colors, the text sitting in the same place, and the whole thing still readable once the app crops it. Doing that by hand is fifteen minutes of fiddling, every time, forever.
So I taught it. It pulls a frame out of the video, drops the headline in my font and my colors, puts one word in the accent color, and checks that nothing important gets cut off by the crop. The cover on the post that sent you here was made that way.
Look at the three-part test against it. I do it several times a week, I want it identical every time, and I can tell instantly when it looks wrong. That is why it worked as a first skill and why "help me come up with post ideas" would have been a bad one.
Worth saying plainly, since I am a nurse and this is an AI page: there is honestly not a lot in my actual hospital job that AI can do for me. Everything I have automated is on the content side. I would rather tell you that than pretend otherwise.
Find yours in one conversation
If you cannot think of a task off the top of your head, do not sit there staring at it. Hand the problem to the AI and let it interview you. Paste this in:
I want to find a repeated task in my work or life that would be worth turning into a reusable set of AI instructions, so I stop explaining the same thing to you over and over. Interview me to find it. Ask me one question at a time and wait for my answer before asking the next one. Ask no more than six questions total. Ask me about the things I actually do in a normal week, especially anything I do more than once. Dig into the boring repetitive parts, not the interesting ones. Then judge each candidate against these three rules and tell me honestly which ones fail: 1. I do it more than once a week. 2. I want the same kind of result every time, not a different one. 3. I already know what a good result looks like, well enough to spot a bad one. When you have a candidate that passes all three, say so plainly and tell me why. If nothing I gave you passes all three, tell me that instead of picking the best of a bad set. Then write me a first draft of the instructions for that task, written so I could paste them back to you next week and get the same result. Keep it plain. No jargon. Once I confirm that draft is good enough, turn it into a saved, reusable capability in whatever form this tool supports: a skill, a custom GPT, a project with saved instructions, or whatever the equivalent is here. Give it a short trigger description that says plainly when it should be used, written in the words I would actually use for the task. Where this tool supports it, that description is what lets you reach for it on your own instead of me hunting for it every time. If this tool cannot do that, tell me plainly and tell me how I start it by hand instead. Then tell me where it now lives and what I say to set it off.
The reason it asks one question at a time is that it will otherwise fire off six at once and you will answer two of them. I use that same rule in almost every prompt I write.
The last part of that prompt is the bit I would not skip. A saved skill you have to remember to go and find is a skill you will stop using. The trigger description is what turns it from a document you own into something that actually shows up when you need it, and writing it in your own plain words matters more than making it sound technical.
How well that works depends on the tool, and the prompt asks it to tell you the truth about that. Some tools will read the description and reach for the skill on their own. Others need you to pick it from a list first. Worth knowing which one you have before you build the habit around it.
Then teach it, one fix at a time
Do not try to write perfect instructions up front. You will not manage it, and it is not how this works anyway.
Do the task with the AI manually once, and do it properly. Then ask it to write down what it just did as instructions you can reuse. It has the whole conversation in front of it, so it writes a much better first draft than you would from a blank page.
Then run it on a real job and fix what comes out wrong. One thing at a time. If you change five rules at once and the result gets worse, you have no idea which one did it. If you change one, you know exactly.
Mine took a handful of rounds and it still gets a note added to it whenever something comes out slightly off. That is not a failure of the setup, that is the setup. You are training a way of working, not installing software.
If you want it to sound like you as well as work like you, that is a different job and it is covered in teaching AI to write like you. And if the task involves facts you are going to publish, run it past how to check AI’s work before you trust it.
What this does not do
It does not make the AI reliable at things it is bad at. A skill makes a task repeatable. It does not make the underlying answer more correct. If it gets the task wrong by hand, it will get it wrong consistently once you write it down, which is arguably worse.
It is not an agent and it will not go do things unsupervised. You still start it, and you still look at what comes out. I consider that a feature.
Your first one will probably be too complicated. Mine was. Pick something small enough that you would be mildly annoyed rather than genuinely set back if it came out wrong.
Guides go out to the newsletter before they go anywhere else. No hype, no daily emails, just the things that held up when I actually used them.