Write it once
As a beginner, the frustration with AI was not that it got things wrong. It was repeating the same instructions in every chat. Skills changed that.
When I started using AI for my investing research, the frustration was not that it got things wrong now and then. It was that I had to keep telling it the same things.
Every chat, I wrote the same instructions again. Use the latest numbers. Check the period. Show me the source. And still, numbers went missing. There was no structure, only me repeating myself. There were other moments too, like being told a web domain was available when it was not.
The models have improved a lot at not making things up. But as a beginner, part of the problem was mine. My rules lived in my head and in old chats, nowhere the AI could use them again.
From prompts to skills
Then I learned about skills. A skill is a written set of instructions the AI loads whenever that kind of task comes up. You write the rules once, and it follows them every time.
It is how people learn a skill too. A new salesperson in their first week needs telling everything, every day: which shops to visit, what to say, which price to quote. Nobody can keep repeating that. So you write the playbook, train them on it, and after a while they do it without being told.
A skill is that playbook, written for AI. The one difference is memory. A person eventually carries the playbook in their head. AI mostly starts each chat fresh, unless the rules are written down somewhere it can find them.
My first one was for investing research. The rule that changed the most: every number has to carry four things, its value, its source, the period it covers and when it was retrieved. One line in that skill sums it up. A figure without a period end is not data. It is a rumour with a decimal point.
A second skill has only one job: check the numbers before they are used, and stay out of the conclusion.
The same idea now runs my writing, with my voice rules and what I must never disclose, and the weekly log of how my posts perform.
Scheduled, not remembered
Because the rules are written down, they can run on a schedule. A review can run on its own and land in Slack for me to look at, instead of waiting for me to remember to ask.
Same loop
It is the same model I wrote about for teams.
Learn. Where does it keep failing me?
Define. What must always be true?
Build. Write it down as a skill.
Execute. Every session starts from it.
Skills do not make AI perfect. They make mistakes rarer and easier to spot. I still check.
What is the instruction you keep typing again and again?