A skill isn't a document. It's a drill.

Researchers recently catalogued how developers actually configure their AI coding agents — eight mechanisms across 2,853 open-source repositories.¹ The headline finding is a quiet indictment: most people configure almost nothing. One context file, and that’s it.

The mechanism they found most neglected is the one I’ve become most interested in: skills. Of those 2,853 repositories, just 158 used skills at all — and of the skills that did exist, 85.5% had no executable component. Notes pinned up for the model to read. Almost nobody builds the other kind.

I’m just starting to use skills myself. But the distinction that made them click is this: a static skill is a note to the AI. An executable skill is a procedure the AI runs. One reminds. The other does.

Here’s a real one. I have a skill called scoping. It doesn’t say “remember to scope the project.” It runs the scoping — an adaptive interview that grills me question by question, saves my answers after every round, and ends by generating the project’s foundational documents. I don’t re-improvise that procedure every time I start something. I invoke it, and it executes the same disciplined drill — whether I’m sharp that day or not.

That’s the difference between a checklist taped to the wall and a checklist someone actually runs before takeoff.

The reason most skills stay static is that the static version is easy and the executable one takes work — you have to encode the procedure, not just describe it. But that work is where the leverage is. A documented procedure depends on a human remembering to follow it. An executed procedure doesn’t. In a regulated build, that’s the difference between “we have a process” and “the process ran.”

Here’s what’s faintly embarrassing about that. I’ve been building software and managing content across a stack of projects for a while now — the exact repeated, multi-project work that skills exist to streamline. And I only just built my first set. Not because I didn’t know they existed. Because the static version — write yourself a note — is always the easy thing in front of you, and you don’t reach for the executable one until you’ve re-run the same procedure by hand enough times to resent it. That’s the trap the numbers describe. Not ignorance. Inertia.

The point is the one I keep coming back to: the value in AI-assisted work isn’t the model. It’s everything you build around it so the model does the right thing every time — not just the time you happened to be paying attention. Skills turn a procedure you keep repeating into one you never have to repeat again.

I’m early in this. But I already know which way the leverage runs.

---

¹ Galster, Mohsenimofidi, Lulla, Abubakar, Treude & Baltes. “Configuring Agentic AI Coding Tools: An Exploratory Study.” arXiv:2602.14690.