If you have explained the same job to AI more than twice…stop.

You do not need a bigger prompt. You need to teach the thing how you work. Once. And then hands off.

OpenAI has just added Skills to ChatGPT. I shot a Youtube guide on AI skills here.

Skills have been around for a while but they have broken confinement. They are coming to more and more surfaces - including now vanilla ChatGPT.

Anthropic invented them, Claude has had them for a while and they have already spread into coding tools (including Codex). Now they are inside the app used by about a billion normal humans.

That makes this a VERY good time to understand what they actually are.

And a great time to catch up if you haven’t been using.

Because yes, "Skills" sounds like yet another stupid AI noun to add to Projects, GPTs, agents, Plugins, apps, MCPs and whatever arrives next Tuesday.

But this one is useful. Promise! 😄

Here's the full video:

Never prompt twice

A prompt asks the AI to do one job.

And when you want to do that task again you need to … prompt again. Eww.

A Skill teaches it the method so the AI can do that job again.

Say you turn a livestream into a newsletter every week. Like I do!

You can keep pasting in the voice rules, structure, links, image requirements and all the annoying little things I complain about when the draft gets them wrong.

Or you can package that process once and let the AI pull it in whenever the job appears.

Much better. Far more efficient. And actually timesaving (which is what AI is for right??).

Models are pretty smart already. BUT they don’t know your process. Your workflow. How you put things together. That is the stuff Skills capture.

FYI: This follows the same logic (and complements) building one shared AI vault. Your useful context should not be trapped in one random chat. It should live somewhere the AI can reach again. Skills are another way to do this.

Anthropic launched Agent Skills in October 2025. Then it opened the format up. OpenAI has now adopted the same basic standard for ChatGPT and Codex.

Thank god. We don’t need even more standards.

It Is Basically A Folder

What actually is a Skill though?

Well…it’s a little underwhelming.

It’s a folder with a bunch of text files in it.

That’s….sorta it?

At minimum, a Skill is a folder with one file inside called SKILL.md.

That .md bit may be unfamiliar. It stands for Markdown and is basically a way to format text. That’s all. Here’s what markdown looks like (on the left) and how it displays (on the right). Don’t worry you don’t need to know this - just want to explain the .md part.

That text file SKILL.md tells the AI what the Skill does, when it should use it and how to carry out the work. You can then add reference documents, templates, examples, images or scripts if the job needs them.

Most of the time it’s just a written set of instructions. For most business jobs, a clear set of instructions and a couple of good examples will get you a long way. Add technical bits (like code) only when something needs to run in a precise, repeatable way.

This is where people think Skills are too complicated. They see folders and markdown and assume it is developer stuff.

Nah. It’s a folder with some text files.

Oh and ChatGPT or Claude can write the file for you anyway.

Build It By Doing The Job

OK how do we make a Skill? Well…get the AI to make it for you…

You do not need to sit down and write an operating manual from a blank page.

Do the job first with the AI. Correct it. Explain why something is wrong. Give it examples. Swear at it a bit (optional but traditional). Keep going until the output is actually good.

Then ask it to review the whole conversation and extract the process you just discovered into one clean Skill.

Do Not Build A Skill To Run Your Entire Business

When people realise how powerful Skills are there is a very natural tendency to want to get ALL their work into Skills.

I get it. They are great.

One important caveat early days.

The first instinct is to chuck your entire business into one giant Skill.

Don't.

It will be too broad, too heavy and very difficult to test. You will have no idea which instruction caused the result to go a bit shit.

Your first Skill should be boring, repeated and bounded….

Turn call notes into a client update. Check a proposal against the same risk list. Research five competitors using the same evidence rules. Turn a transcript into a newsletter with the same structure. Yadda yadda.

One job. Clear inputs. Clear output. A checklist at the end. Kinda dull stuff but important!

Then test it with three requests: one that should trigger it, one that should not, and one horrible messy real-world example.

If it survives all three…now we're talking.

If not: tweak. Again, do so with the AI. Run some more tests and tweak again. Until happy.

Then, when you are happy with the Skill keep it saved somewhere. Mine are in Github personally but you can have them locally on your computer no problem. Ask ChatGPT and it’ll save them somewhere smart.

You can then use those Skills anywhere - in Chat, in ChatGPT Work, in Codex. Or indeed in other tools like Claude. I highly recommend adding them to your AI brain if you have one so they can be accessed everywhere - but that’s a little more advanced.

For now just wrap one process up nicely and use it over the rest of the week to get the feel for how we build repeatable Skills.

To the Task,

Kyle