I used to search for Agent Skills the obvious way.
If I needed a skill for analyzing stocks, I'd search for stock analysis.
If I needed one for making a PowerPoint, I'd search for PPT.
Then I'd look at the download counts and pick one that seemed popular.
It sounds reasonable. But there’s a problem:
I wasn't really searching for a skill. I was searching for a name.
Take PPT.
A keyword search can easily return skills for reading PPT files, extracting content, parsing slides, compressing presentations, and generating presentations.
They're all related to PowerPoint.
But if what I actually mean is:
I need a skill to make a PowerPoint presentation for my product launch.
those aren't equally useful.
So I started searching in a different way.
Instead of telling the search engine what the skill is called, I tell it what I actually need to get done.
And the difference is surprisingly large.
For the PPT example, a simple PPT search returned a bunch of literal matches, including tools for parsing and extracting PowerPoint files.
When I described the actual task — making a PowerPoint presentation for a product launch — the results became much more focused. The search returned five skills specifically oriented toward presentation creation, each with information about what it can actually do, what it requires, and where its limitations are.
That distinction matters.
Because when I'm looking for a skill, I don't really care whether its name contains the right keyword.
I care whether it can help me finish my task.
Popularity, relevance, quality
This also changed how I think about download counts.
A skill with 50,000 downloads is obviously popular. That's useful information.
But popularity answers a different question:
How many people have downloaded this?
It doesn't necessarily answer:
Is this the right skill for what I'm trying to do?
And even after finding a skill that fits the task, there's another question:
Does it actually work well?
I've been testing Agent Skills across different categories, and that last question is where things get interesting. In a previous experiment, I ran several code-comment skills against the exact same Python file and found meaningful differences in the actual outputs.
So I've started thinking about skill discovery as three separate layers:
Popularity → Relevance → Quality
Download counts can tell you something about popularity.
Task-based search can help with relevance.
And actual testing is what tells you about quality.
You shouldn't confuse the three.
The search problem isn't just finding more skills
There are already plenty of places where you can browse Agent Skills.
The harder problem, at least for me, is figuring out:
Which of these are actually relevant to the thing I'm trying to accomplish?
That's why I think searching by task is more useful than searching by skill name alone.
You don't need to know the vocabulary of the Skill ecosystem.
You don't need to guess which keywords the author used.
You can just describe the job you need to get done.
And ideally, the search should do the filtering for you — not hand you 50 vaguely related results and make you figure out the rest.
Top comments (0)