The most-downloaded AI skills aren't the best

TL;DR. We sorted the SkillHub catalog two ways — by download count and by automated review score (aiScore) — and the two leaderboards barely overlap. Of the twenty most-downloaded skills, almost all score under 70/100, and the single most-downloaded skill of all hasn’t passed review at all. For five common jobs — humanizing text, building slides, generating images, trading, and driving a browser — a higher-rated alternative already exists. Download count tracks age, bundling, and discovery; it is not a quality signal. Shop by the score, then read the skill.

Two leaderboards that don’t agree

SkillHub indexes 72,761 skills. About 10,000 of them have been through our automated review pipeline, and across that reviewed set the highest aiScore anyone has earned is 89/100 (Anthropic’s official hook-development). So when you sort by quality, the top of the list sits in the 80s.

Sort by downloads instead and you get a completely different list. The most-downloaded skill in the entire catalog — a self-monitoring “debugging-agent” with over 3,200 installs — hasn’t passed review at all. The second, a security “skill-vetter” with 3,255 installs, scores 46. Walk down the top twenty and you will pass humanizers, search wrappers, and weather tools long before you reach anything that cleared 70.

The headline number. Of the 20 most-downloaded skills on SkillHub, only a handful score 70 or above, and none reaches the high-quality bar of 76 that the top ~485 skills clear. The download chart and the quality chart are nearly disjoint.

Why popularity and quality diverge

This is not a scandal — it is what download counters always measure. A download is a lagging, cumulative vote cast before anyone uses the skill. Three forces inflate it independently of quality:

Age. A skill published months ago has had months to accumulate installs. A better skill shipped last week starts at zero.
Bundling and ecosystems. Several of the top entries belong to one prolific publisher whose skills are installed together as a set. One decision to adopt the bundle counts as many downloads.
Discovery. Skills with broad, generic names (“search”, “weather”, “humanizer”) surface for more queries than precise, specialized ones — so they get clicked more, regardless of how well they do the job.

None of those forces inspects what the skill actually instructs the model to do. The review score does — it reads the skill body and judges clarity, scope, and whether the guidance is distinctive or generic. The two metrics answer different questions: downloads ask “how many people found this first?” and the score asks “is this any good?”

Five popular skills and their higher-rated alternatives

For most popular-but-mediocre skills, a higher-scored skill that does the same job already sits in the catalog — sometimes from the very same author. Here are five swaps worth knowing.

The job	Popular pick (downloads · aiScore)	Higher-rated alternative (aiScore)	What the gap buys you
Humanize AI-written text	`openclaw/humanizer` — 1,306 · 63	`dparedesi/…/humanize` — 84	Sharper, less generic rewriting guidance instead of a thin “remove AI tells” prompt.
Build a slide deck	`openclaw/ai-ppt-generator` — 838 · 49	`hermes-agent/powerpoint` — 85	Reads, edits, and writes real `.pptx` with notes and templates — not a one-shot wrapper.
Generate images	`openclaw/nano-banana-pro` — 838 · 57	`openai/skills/imagegen` — 84	A maintained, first-party generate-and-edit loop with clearer triggers.
Trade / analyze markets	`openclaw/stock-analysis` — 996 · 69	`openclaw/alpaca-trading` — 85	Same publisher, higher bar: execution, portfolio, and risk — not just a quote lookup.
Drive a browser	`openclaw/agent-browser` — 2,547 · 67	`mrgoonie/…/chrome-devtools` — 84	Automation plus real debugging and performance tracing through the DevTools protocol.

The pattern repeats well past these five. The lesson isn’t “avoid popular skills” — it’s that the install count told you nothing about which of two skills with the same name is the better one.

In fairness to the popular skills

A low score is not a verdict of “bad.” Many of these skills work fine for the common case. A 63 often means “competent but generic,” not “broken.” And a high download count is genuinely useful information — it tells you a skill is real, maintained enough to spread, and battle-tested by other users. The argument here is narrower: popularity should not be your primary filter, because it systematically favours whatever shipped first over whatever works best.

There is also a sharper reason to read before you install, and it has nothing to do with the score. Earlier this year we documented malicious skills hiding obfuscated payloads — with over a thousand combined downloads before they were caught. Download count cannot tell a useful skill from a dangerous one. The review can, and so can you.

How we measured this

Both leaderboards come straight from the public SkillHub API. We pulled the catalog sorted by downloadCount and again sorted by aiScore, both on the same day, and lined the two up. aiScore is a single automated reviewer’s 0–100 read of the skill’s instructions — clarity, scope, trigger quality, and how distinctive the guidance is — not a measure of runtime success. Treat it as a strong prior, the way you would treat a senior reviewer’s first pass, not as ground truth. Numbers drift as new reviews land and installs accumulate; the shape of the gap does not.

What to do with this

Sort by score first, then read. Use aiScore to build a shortlist, then open the top two or three and read the actual skill body before you commit.
When two skills share a name, the score breaks the tie. “humanizer” at 63 and “humanize” at 84 are not the same product.
Read every skill before install — popular or not. A skill is instructions your agent will follow with your tools and your keys. Popularity is not a safety check.

Browse the full reviewed catalog at skills.palebluedot.live, sorted by score rather than installs. And if you want to see what a top-of-the-chart skill actually looks like in practice, our Skill Spotlight series takes one apart at a time.

FAQ

Does a low aiScore mean the skill is broken?
No. It usually means the instructions are generic or loosely scoped, not that the skill fails. It is a reason to compare alternatives, not to assume the skill doesn’t work.

Why is the highest score only 89, not 100?
The rubric reserves the top of the range for genuinely exceptional, distinctive guidance. Most solid, useful skills land in the 70s and low 80s — that is the “good” band, not a disappointment.

Are download counts faked?
No — they are genuine installs. Each one is counted server-side only when the SkillHub CLI finishes installing a skill, and a given IP address can add at most one count per skill every five minutes, so page refreshes and retries do not inflate it. The counting code is open source. The skew toward older, bundled, and generically named skills comes from those ordinary forces — not from fabricated numbers.

So should I ignore popular skills?
No — popularity is useful secondary evidence that a skill is real and maintained. Just don’t let it be the only filter. Lead with the score, confirm by reading.

Where do the “485 high-quality skills” come from?
That is the set scoring 76 or higher — the top of the reviewed catalog. None of the twenty most-downloaded skills is in it.

About this Report. Reports are data-driven field notes from the SkillHub auto-review pipeline — open, reproducible, and published as the numbers land. Read more on the about page.

The most-downloaded AI skills aren’t the best ones. Here’s the data.

On this page