Trust the person, not the catalog

Ask your agent to "find a frontend skill," and it reaches into a directory of 900,000 of them (skills.sh alone indexes that many), grabs whatever ranks highest, and runs it inside your project.

Stop and look at that. You just let a stranger's instructions drive your tools because they were popular. You'd never do that with code.

And you already know you wouldn't, because you have a model for it: GitHub. You don't install the top search result and run it blind. You look at who's behind the repo. Is it Vercel? Is it someone whose other work you've shipped against? A package with a lot of downloads and no name you recognize is exposure, not trust. Downloads tell you something is popular. The maintainer tells you whether to trust it.

A skill is code for your agent: instructions that run inside your tools, deserving the same scrutiny. The catalog model throws that scrutiny away. A leaderboard ranks by install count, and install count tells you what's popular, not what's good, and never what's good for you. Vercel's React best-practices skill has 556,000 installs, and most of those people don't deploy on Vercel. It was written for someone else's stack and someone else's taste.

Here's the part most people miss. The cure for slop is a stronger perspective, not a bigger model, and a perspective belongs to a person.

The reason AI output slides into slop is that the model has no point of view. Ask for "a landing page" and you get the average of every landing page: gray, hedged, forgettable. The model will generate forever in any direction. What it lacks is a reason to choose one.

That's what a skill actually is: a point of view, not a checklist. It's how Maya writes, how Grace reviews code, how Priya sizes a launch to its risk. It's taste, made portable: a person's judgment, packaged so your agent can borrow it. That's the thing the model doesn't have and can't invent for itself.

Which is the whole point. A skill is only as good as the taste behind it, so the question that matters is whose judgment you want in the room, not what ranks first. You don't trust a frontend skill with forty thousand installs. You trust the engineer whose work you'd copy on purpose.

So follow people, not catalogs. Subscribe to the few whose taste you'd vouch for, and let their best work land in your tools and stay current, as a diff you approve, not a surprise. Reputation here is how many people run your skills, not how many stars you collected. You earn it the way you earn any reputation: by doing good work in the open, with your name on it.

The catalog treats skills as interchangeable. They're the opposite. A skill is somebody's taste, and taste is the one thing a model can't average its way into.

You'll hear the other prediction: that all of this gets absorbed into the model. That a year from now skills are "batteries included, no instructions necessary," the way prompt engineering quietly became just talking to your agent. Half right. The generic skill, the one that wraps a thing the model will soon do on its own, does evaporate. But the model ships generic. It never ships yours: your taste, your team's conventions, the judgment you'd vouch for. And the faster the generic stuff gets absorbed and the faster models turn over, the sharper the only two questions that survive get: whose skill is this, and does it still work today? A catalog can't answer either. A person can.

Trust the person, not the catalog.