I kept telling AI the same things every session. So I started writing them down as skills, and testing them like code.
Every session with an agent starts from zero. It doesn't know how we name layers, what a good component doc looks like, or how I write. So I kept explaining it.. again and again.
That's not a workflow though. It's a habit waiting to drift.
A skill is a standard the agent reads every time
A skill is a short file. It says when to use it, then what to do, in order. The agent picks it up whenever the task matches, so the standard shows up without me having to bring it.
At ecoPortal I've written 13 of them. One writes component docs, so we get them done faster. Another runs design QA, so we catch defects sooner than we would by eye. And one turns a finished design into a working prototype. It reads every panel first, writes the spec, and stops for my OK before it builds anything.
Test it like code
A skill that reads well isn't the same as one that works. So I test them: 18 cases across 6 skills, a blind A/B comparison, and tuning so each one comes in when it should and stays quiet when it shouldn't. A skill nobody tests is just a nicer prompt. You get?
Then let it keep learning
The newest one is personal. It's how I write. I built it from 2,686 messages I'd typed to Claude, and every time it runs, it reads what I've typed since, so it keeps up with me. It keeps numbers, never my messages. When I say a line doesn't sound like me, that correction goes straight in and outranks everything else.
This post went through it too.
I think that's the shift for me. Write the standard down once. Test it. Then keep teaching it.