📊 Install count is biased
Install count is the most visible signal on skills.sh—and the most misunderstood. It measures cumulative adoption, not current quality. It’s a cumulative number that favors those who arrived early, has brand and became a prerequisite for other things. find-skills has 1.8M not because it’s "twice as good" as frontend-design (488k), but because it’s installed as a dependency in almost every workflow.
✗ What install count does NOT tell you
- ✗That the skill is still maintained and up to date
- ✗That it’s better than a newly released one
- ✗That the people who installed it actually use it
- ✗That it serves YOUR specific use case
✓ What it really signals
- ✓That many people backed it at some point
- ✓That it’s probably not obvious junk
- ✓First-mover advantage
- ✓Strength of the brand behind the source
💡 Treat it as reputation, not a score
Install count is like followers on Twitter: a high number correlates with relevance, but doesn't prove the quality of today's content. Use it as an initial filter ("worth a look"), never as the final decision.
📈 Find rising skills and untapped niches
The power law is also an opportunity map. Remember: 60,2% of skills have <100 installs, 30,8% between 100-999, and only 131 skills exceed 100k. The top is saturated and locked in. The opportunities — and undervalued gems — are in the middle range and smaller groups.
Look in the middle range
Skills with 100–10k installs that are growing are often better bets than the fossilized ones at the top. Look for a recent update date + rising install counts = a skill on the rise.
Compare group vs. installs/skill
Groups with low supply and real demand are untapped niches: Office, Finance & Business, Productivity & PM have hundreds of skills, not thousands. Searching there pays off more than competing in AI/Agents (12.693 crowded skills).
Look for the absence
The best opportunity is what you DON’T find in search. If you look for your specific pain point and can’t find anything good, that’s the empty niche—and the signal to create something (go back to Track 1.4).
🏷️ Evaluate the source
Before looking at install count, look at who published it. The source matters a lot because you'll give a skill the chance to change your agent's behavior — and sometimes to run things. There are three source profiles, and each calls for a different level of trust.
supabase, stripe, microsoft, vercel-labs — the product owner
anthropics/skills — who defined the format
obra/superpowers, coreyhaines31 — a name with a track record
account with no history, few installs, no README
💡 A skill is an instruction you adopt
A skill from a questionable source may contain bad (or malicious) instructions that the agent will follow. Open the SKILL.md before installing from unknown sources—it’s text and takes 30 seconds. You can install from an official vendor with confidence.
⌨️ Get the most out of the CLI
You don’t need to open the site for anything. The npx skills handles search, installation, and maintenance from the command line—and works even better with the find-skills skill installed, which lets the agent operate the CLI for you.
the commands that matter:
$ npx skills find "postgres rls" # buscar por termo $ npx skills add supabase/supabase # instalar de um repo $ npx skills add ./skills/minha-skill # instalar local $ npx skills list # ver o que está instalado $ npx skills update # atualizar tudo para a última versão
get the most out of find
Search for the pain point, not by name ("validate stripe webhook," not "stripe"). Compare 2-3 results by description before installing—the description already reveals whether the skill targets your use case.
let the agent operate
With find-skills installed, ask in natural language: "find and install the best skill for e2e testing". The agent runs find, evaluate and add alone.
💡 updates aren't free
npx skills update pulls the latest version—great for fixes, but it can change behavior you relied on. Run it periodically and test afterward, especially for process skills (like TDD) that affect your entire workflow.
🛑 When NOT to Install
The most advanced skill is restraint. Every installed skill consumes context (the name + description are always loaded) and adds a possible unwanted trigger. More skills aren’t always better—sometimes the best move is to don’t install.
✗ Do not install when...
- ✗It’s something you do only once—a prompt gets it done
- ✗A similar skill is already active (descriptions collide)
- ✗The source is unknown, and you haven’t read the SKILL.md
- ✗It’s fixed project context—better suited to CLAUDE.md
- ✗You’re accumulating things "just in case," without a use case
✓ Install when...
- ✓It’s a behavior you want to be repeatable and automatic
- ✓Solves a specific, recurring pain point of yours
- ✓The source is trusted, or you’ve read the content
- ✓Doesn’t conflict with what’s already installed
- ✓You’ll use it in the coming weeks, not "someday"
the 3-question test before installing:
[ ] vou usar isto mais de uma vez? [ ] confio na fonte (ou li o SKILL.md)? [ ] não tenho nada que já faça isto?
💡 Do periodic cleanup
Run npx skills list once in a while and remove what you haven't used in weeks. A lean set of relevant skills triggers better and uses less context than an bloated collection competing for attention.
🧠 The pro mindset: actionable pro tips
Bringing it all together, reading the ecosystem “like a pro” is less about numbers and more about applied skepticism. Six habits that separate people who just look at the leaderboard from those who get real value from the catalog.
Read the description, not the number
The description reveals more about whether it fits your use case than any install count. If it doesn't clearly say WHAT and WHEN, be wary of the entire skill.
Prefer recent + trending over old + large
A skill with 5k installs updated last week tends to beat one with 200k that hasn’t changed in a year. Activity > size.
Search for the problem, install the minimum
Look for the problem in natural language; install only what passes the 3-question test. Resist accumulation.
Treat the catalog as a map of niches
Where searches for your pain point come up empty, you have an opportunity to create something. The catalog tells you both what to install and what’s missing from the world.
Update and review in cycles
Run update e list periodically; test after updating; remove dead code. Maintenance is part of using it well.
Know when not to install
Restraint is a competitive advantage. Fewer, more relevant, better-tuned skills — that's what gets the agent to perform for real.
💡 End of Track 1 — you have the map
You already know what a skill is, what the market looks like, what to install first, how to create your own, and how to read everything critically. Track 2 dives deeper into what separates the top from the long tail: quality.
✅ Module Summary
Next:
Track 2 — 🎯 Quality & The Best. What sets the 131 top skills apart from the 60.2% in the tail — and how to recognize (and write) a high-quality skill.