I was browsing skills.sh the other day—skills are these knowledge packages that teach AI coding agents how to work with specific frameworks and tools—and I noticed something.
The top skills aren’t the ones I would have guessed.
I would’ve expected… I don’t know, generic helpers. “Write better code.” “Debug faster.” The kind of vague improvement that sounds good in marketing copy.
Instead, the most installed skills are very specific. Very opinionated. Very… practical.
What’s Actually Popular
The top skill, with over 34,000 installations, is Vercel’s React best practices. Not “React helper”—React best practices. Specific patterns for specific performance problems.
Number two is Vercel’s web design guidelines. Again, not “make websites”—a comprehensive set of rules for accessibility, performance, and UX.
Number three is Remotion’s best practices. Remotion, if you don’t know it, is a framework for programmatic video. It’s not something every developer uses. But if you do use it, this skill is apparently essential.
What I noticed is that these aren’t generic “be better at coding” skills. They’re “here’s exactly how we do things” skills. They’re opinionated. They’re specific. They’re the kind of thing that would only be useful if you’re working in that ecosystem—but incredibly useful if you are.
Why This Makes Sense
I think this tells us something about how developers are actually using AI agents.
Not as generic assistants that can do anything. But as specialists that can do specific things very well—if they’re taught how.
When you install Vercel’s React best practices skill, you’re not installing “help with React.” You’re installing “Vercel’s accumulated wisdom about React performance, captured in a form that an AI agent can use.”
That’s powerful. It means you can say “add a feature” to an AI agent, and instead of getting generic React code, you get code that follows Vercel’s performance patterns. The same patterns Vercel uses in their own products.
That’s different from asking an AI to “write performant React code” and hoping it knows what that means. With the skill, you’re importing expertise.
What the Popular Skills Share
Looking at the top skills, a few patterns emerge:
They’re framework-specific. React best practices, Remotion best practices, Expo skills for React Native. Not “web development”—specific frameworks, specific patterns.
They’re opinionated. Vercel’s skill doesn’t say “here are options for doing this.” It says “here’s how we do this.” There’s a right way, and the skill encodes it.
They solve real problems. Web design guidelines isn’t theory—it’s a checklist for accessibility, performance, UX. Things you’d otherwise have to remember or audit for.
They’re from authoritative sources. Vercel for React/Next.js. Remotion for video. Expo for React Native. Better Auth for authentication. These aren’t random developers’ opinions—they’re from the people who build these tools.
What This Means
I think this is a signal about how AI development tools are evolving.
We’re moving away from “ask the AI anything” toward “teach the AI your context.”
The difference is subtle but important. In the first model, the AI is supposed to be generally smart. In the second, the AI is given specific knowledge about your specific situation.
Skills are how you give it that knowledge.
When you install a skill, you’re saying: “When you’re helping me with React, don’t just use your general training. Use Vercel’s specific approach. Here it is, encoded in a form you can understand.”
That’s more powerful than “this AI knows React.” It’s “this AI knows how Vercel does React.”
The Interesting Edge Cases
Some of the top skills are for things I wouldn’t have expected to be popular.
Agent browser, from Vercel Labs, lets AI agents browse the web. It’s ranked 7th with 2,100 installations. That’s not tiny for a relatively specialized capability.
I find this interesting because it suggests developers are using AI agents for tasks that require web research. Not just “write code”—tasks that need information from outside the codebase.
Skills for upgrading Expo projects, for fetching data in React Native apps, for authentication with Better Auth—these are all very specific problems. Not “help with mobile development” but “here’s exactly how to upgrade an Expo SDK.”
This tells me developers aren’t looking for generalist help. They’re looking for specific solutions to specific problems.
What’s Missing
Also interesting: what’s not in the top 10.
There are no skills for general debugging. No “write better tests” skills. No “improve code quality” skills.
Maybe those exist lower down the list. But they’re not what thousands of developers are installing.
What developers are installing is stuff that helps them with specific frameworks, specific patterns, specific problems.
I think this is because those are the kinds of problems where:
- The AI’s general knowledge is actually good enough
- But specific, framework-specific guidance is better
- And that guidance is worth encoding as a skill
General coding advice? The AI can probably handle that. But “here’s how Remotion handles audio synchronization”? That’s worth encoding.
The Implications
If this pattern holds, I think it tells us something about the future of AI development tools.
The most valuable tools won’t be the ones that try to be everything to everyone. They’ll be the ones that:
- Focus on specific ecosystems
- Encode specific expertise
- Come from authoritative sources
- Solve real, painful problems
Vercel gets this. That’s why they have not one but two skills in the top 3. They’re not trying to teach AI agents about web development in general. They’re teaching them “here’s how Vercel does web development.”
That’s more useful. Because if you’re using Vercel’s stack—or just Next.js in general—you probably want to do things the way Vercel does them. They’ve thought about this more than you have.
If You’re Building Skills
I wouldn’t try to build the next “general coding helper.” Those are a dime a dozen, and honestly, the AI is already pretty good at that.
I’d look for:
- Specific frameworks where you have deep knowledge
- Painful problems where you’ve developed specific approaches
- Gaps where the AI’s general knowledge falls short
Encode that. Make it a skill.
If it solves a real problem for a specific community, you might just end up on that top 10 list.
The Bigger Picture
What I keep coming back to is that skills represent a different approach to AI tooling.
Instead of trying to make AI agents that know everything, we’re making AI agents that can be taught anything.
Skills are how you teach them.
The most popular skills are the ones that teach the agents something genuinely useful. Something that’s not already in their training data. Something that makes a tangible difference in day-to-day work.
Vercel’s React patterns matter. Their web design guidelines matter. That’s why thousands of developers have installed those skills.
The question isn’t “how do we make AI smarter?” It’s “what specific, valuable expertise can we give it?”
This analysis is based on installation data from skills.sh. The ecosystem is evolving quickly, and I’m sure the rankings will shift. But I think the underlying pattern—that specific, opinionated skills beat generic helpers—will hold.