A Skill Is a Correction You Kept
Stop taking the work back off Claude. Correct it, keep the correction, and it does the job your way from then on.
Start Here, Paste This
Here from the video? This is the line. Next time Claude Code gets something wrong, correct it the way you would a new starter, and once it has done the job the way you wanted, paste this in. The conversation is the interview, so it has nothing to ask you.
In a hurry, "save that as a skill" on its own does the same job. The longer version is there so it merges into a skill you already have rather than making a second one, and shows you what it wrote.
What you'll walk away with
- The line that turns a correction into a file Claude reads before the next go
- The five steps of building one, and the line for keeping it sharp afterwards
- The version for the Claude or ChatGPT app, and the three jobs worth saving first
Your first one takes about ten minutes, most of it the correcting you were going to do anyway.
How One Gets Built
You do not sit down and write a skill. You do a job, correct it, and keep the correction, which means the first one is about ten minutes on top of work you were doing anyway.
1Do the job in a normal chat
Ask for the thing the way you would anyway, the email, the summary, the report, and let it get it wrong. The wrong version is the raw material, so do not steer it into being right first time.
2Correct it like a new starter
Say what you wanted and why, in your own words, rather than pasting in a fixed version. "Shorter, and never open with a compliment" teaches it something. A rewritten email teaches it nothing.
3When it is right, keep it
Paste the line above. Claude writes the file, names it, gives it the description it will match jobs against, and puts your corrections in as rules.
4Read what it wrote
It is a plain text file, and it is short. Open it and read it. If a rule is not how you would say it, change the words, and if it has kept something you only said once, take it out. This is the bit most people skip, and it is the bit that makes the skill yours rather than Claude's guess at you.
5Keep correcting it
Saving it is not the end. The next time it gets that job wrong, the correction goes into the skill rather than the chat, so the file gets sharper and you never repeat yourself. That is the whole loop, and it is your job to keep it turning.
For step five, this is the line. Same idea as the first one, except the correction lands in the file.
What's Actually Going On
When Claude gets something wrong, most of us take the work back and do it ourselves, and that is the one move that guarantees it never gets better. Your job is to correct it, the way you would a new starter, and a new starter remembers. Claude does not. Every new chat is day one, so whatever you told it yesterday is not in the room today unless you gave it somewhere to keep it.
A skill is that somewhere. A plain text file with a name, a one line description of which jobs it is for, and the steps as you want them done, saved where Claude looks before it starts. The description is the clever bit, because Claude reads it and reaches for the skill itself whenever a job matches, so you do not have to remember it exists. Save it in .claude/skills inside a project and everyone on that project gets it, or in ~/.claude/skills and it follows you everywhere.
None of it is clever. I have thirty-odd of these and nearly every one started with Claude getting something wrong. That is all an agent ecosystem is, a pile of corrections that never leave, and it takes on more of the job every month.
If You're in the App
No Claude Code, no folder to save into? The same idea works in the Claude or ChatGPT app, it just needs to interview you first, because there is no conversation behind it to learn from. Paste this in, answer its questions, and correct what it produces until it is right.
What it writes up goes into a Project's instructions in Claude, or a Custom GPT in ChatGPT, both of which read it at the start of every chat. Without either, keep it in a note and paste it in first, which is less tidy and works just as well.
The Three Worth Saving First
The rule is anything you have corrected twice. If you want a place to start, these are the three almost everyone has.
1The email you write most
The client update, the chaser, the reply to an enquiry. You have corrected the tone of it more times than anything else you ask for, and every one of those corrections is a rule waiting to be kept.
2The summary you always reformat
Meeting notes, a call, a long thread. You get a summary back and rearrange it into the shape you actually send, which means the shape is the skill and you have been doing it by hand.
3The report in your company's format
Anything with a layout somebody else expects. The headings, the order, what goes in the first line. Once saved, you ask for the report and get the format for free.
One Honest Bit
It is slower the first few times, because you are training someone, and the payoff is the tenth time you ask rather than the first. It is your job to keep correcting it through that, and the people who give up at the second go are the ones still doing the work by hand a year later.
Two things to know. Claude only reaches for a skill when the job matches its description, so if one is not getting used, the description is the line to fix. And do not save one-offs. Keep the corrections you have made twice, because a drawer full of instructions for things you asked for once is a drawer Claude half reads.
Where This Goes Next
A skill is how one job gets done. How you like things done in general, the tone, the rules, the things it should never do, lives in your claude.md, which Claude reads before every prompt. If you want five finished skills to copy before you write your own, the five skills are mine. And the folder structure is where all of this lives, one place your AI works out of rather than a chat you start again each time.
Want to get better at this?
Let's set up your agent ecosystem.
This page is one job. The whole thing is a set of them running off the same foundation, the folder your AI works out of, the brief on how you like things done, the accounts it can reach and the skills that run off all of it, so your admin gets done the way you would do it. It is what I teach one to one, on your own machine and against your own work, and the quickest way to find out what yours would look like is a proper chat about it.
Not ready for a call? Start with the free agent series, what an AI agent actually is, and build up from there.