Why your skills never get better
Most people run their skills like this. Claude writes the script. You read it. You like some parts and hate others. You tweak it yourself, record it, and move on.
That workflow has a hole in it. All your fixes live in the final file. Claude never sees them. Next time you run the skill, it starts from the same blank slate and writes the same stuff you hated last week.
You end up fixing the same three things forever. Too long. Too formal. Wrong opening. Every single time.
The skill is not bad. It just has no memory of you.
Where this comes from
Andrej Karpathy, one of the founders of OpenAI, put a name on this back in May 2025. He called it system prompt learning.
His point was simple. Pretraining gives a model knowledge. Fine-tuning gives it habits. But a lot of how humans actually learn looks like neither. You hit a problem, figure something out, and write yourself a note for next time. "When this happens, do that." Learning by taking notes, not by rewiring your brain.
He argued models should learn the same way. Not by retraining. By editing their own instructions based on what worked.
You do not need to wait for the AI labs to build that. You can do it right now, inside your own Claude Code skills, with a couple of markdown files.
What preference learning actually is
Here is the whole change. One extra step.
That last part is the magic. Because you did the edits in the conversation instead of in a text editor, Claude can see the gap between what it wrote and what you actually wanted. It writes that gap down. Next run, it reads its own notes before it drafts.
That is preference learning. Claude building taste for you, one revision at a time.
The three files that do the work
You do not need code. You need three markdown files living inside the skill folder.
I run this inside my own email skill. The revisions log is now dozens of entries deep. Things like "revenue stat gets followed by the offer shape on the next line" and "hold the price for the release email, not the tease." I said each of those once. Claude wrote them down. It has not messed them up since.
How to add it to any skill
You have two options. The paste-it-yourself block, or the prompt that makes Claude do the whole thing and ask you the right questions first. Most people should use the prompt.
Option 1: The prompt RECOMMENDED
Paste this into Claude Code:
Claude will list your skills, make you pick, push back if you picked wrong, and build the whole loop.
Option 2: Paste the block yourself
Already know which skill and comfortable editing files? Open its SKILL.md and paste this at the end, then create the three empty files next to it.
Which skills this belongs on
- Content scripts, reels, YouTube outlines
- Emails to your list
- Ad copy and headlines
- Sales pages, landing pages, website copy
- Proposals, DMs, captions, anything a client or audience reads in your voice
- Research and summarizing skills. You want facts, not your flavor.
- Scrapers, data pulls, spreadsheets
- Coding skills. Correct beats "sounds like me."
- One-off tasks you will never run again. Nothing to learn.
One skill at a time. Get it working on the skill you run most, feel the drafts improve, then add it to the next one.
The one habit that makes it work
This only works if you give feedback in the chat instead of quietly fixing the file yourself.
I know. Editing the file is faster in the moment. But every edit you make outside the conversation is a lesson Claude never learns. You pay for that shortcut every single week after.
So say it out loud. "Cut the first paragraph, too slow." "Never call it a game changer." "Shorter sentences in the hook." Approve the final in the chat. Let Claude do the bookkeeping.
Two weeks in, you will notice the drafts need fewer fixes. A month in, you will approve some on the first pass. That is the compounding most people never get from AI because they never gave it a memory.