跪拜 Guibai
← Back to the summary

Top-10 AI Skills by Installs: Only 3 Actually Work — and #2 Is 157 Bytes of Nothing

Yesterday I came across a Skills leaderboard sorted by install count. I stared at the top 10 for half a minute, then pulled down the whole page and installed every single one into an empty directory.

They all installed — all ten exited with code 0. But when I ran the first command from each skill's documentation, only 3 actually produced something. The #2 skill, 1.5M installs, has a skill file of 157 bytes.

How I define "usable"

The criteria were fixed before I ran anything, not adjusted to fit the results. Three of them:

Pass all three and it's usable. Miss one, I count it as stuck.

#1: 4.2M installs, and its job is to teach you how to find skills

find-skills, 5.4KB, one file. Running npx skills find react --owner vercel-labs actually returned a result. 868.4K installs.

Usable. But it produces nothing itself — it teaches you how to find skills, judge quality, and install them. The leaderboard's #1 is a manual.

A navigation app having the most installs doesn't mean it can drive you around.

#2: 157 bytes

grill-me, 1.5M installs. The entire SKILL.md:

Call the Skill tool with "grilling".

That's it.

It tells you to call another skill called grilling. grilling is in the same repo, 2,077 bytes, and looks reasonably well-written — but it was not installed. What you get after installing grill-me is a room key with the wrong room number.

The leaderboard calls this a "skill." I followed the docs and couldn't reach anyone on the first step.

#3, #4, #6: three entries, one dependency

hyperframes-cli, hyperframes, and media-use take up three spots, with installs of 1.4M, 1.4M, and 1.3M.

I tried installing only hyperframes-cli in a clean directory. Three references in its SKILL.md point to ../hyperframes/references/, and that directory doesn't exist. All three siblings have to be installed together for anything to work.

Once together, hyperframes actually works: npx hyperframes init demo generated a complete HTML video project with index.html, package.json, everything.

media-use is the most substantial of the three — 117 files, 69 scripts. Running its first documented command to fetch background music gives:

heygen CLI not found — it's the free path for bgm/image/voice/avatar-video.
Install the CLI from ..., then run: hyperframes auth login

The skill with the most scripts demands you install an external CLI and log in before the first step. It's in the docs, but it's outside the scope of "installing this package."

#5: smallest, but the smoothest

agent-browser, 3.2KB, one file, no scripts.

npx agent-browser open https://example.com actually launched a browser and returned the page title, "Example Domain."

In the smallest tier by size, it's the only one I can't find fault with. How well a skill works has less to do with how many files and scripts it has than you might think.

#7 to #10: each with its own gate

#7 grill-with-docs, 247 bytes, full text:

Call the Skill tool twice, for "grilling" and "domain-modeling".

Needs both. Neither is installed.

#8 reddit-automation, 1.2M installs. One line from its docs:

Reddit has no "post a comment" API — the user pastes the final reply by hand.

A skill named "automation" whose final step is you copy-pasting. No code blocks, no scripts. To its credit, the doc is honest — it doesn't deceive you. But install count isn't something it earned on merit.

#9 improve-codebase-architecture has real content, 6KB, plus an HTML report template. The opening line: first call the codebase-design skill to unify architecture vocabulary. That skill isn't present locally; it's in the same repo, just not installed.

#10 design-mobile-apps, 8.5KB, clearly documents REST endpoints, error codes, and pagination. Its frontmatter has a line: requires-env: SLEEK_API_KEY. I hit it with curl and got 401.

This isn't a trap — it's a clearly priced paywall. The leaderboard just doesn't tell you.

The full view

Rank Skill Verdict Where it gets stuck
1 find-skills Usable It's a manual, produces nothing
2 grill-me Shell 157 bytes, points to uninstalled skill
3 hyperframes-cli Conditional Must be installed with hyperframes
4 hyperframes Usable None
5 agent-browser Usable None
6 media-use Conditional Needs separate CLI install and login
7 grill-with-docs Shell 247 bytes, both dependencies absent
8 reddit-automation Conditional No posting API, manual paste
9 improve-codebase-architecture Broken link Depends on uninstalled codebase-design
10 design-mobile-apps Conditional Requires paid key

3 usable directly, 4 conditional, 3 shells or broken links.

Six of the top 10 come from two accounts. Not my selection — that's how the install ranking shakes out.

Counter-arguments

"Did you deliberately pick the bad ones?" I took the top 10 by installs, missed none, installed all.

"Are the shells because you can't read docs?" I read them. The #2 doc was one line.

"Is your environment wrong?" agent-browser launched a browser, hyperframes generated a project. The environment is fine.

Before installing, do three things

npx skills add owner/repo --skill <name> -a qoder -y

wc -c .qoder/skills/<name>/SKILL.md              # under 500 bytes, don't get excited yet
grep -rn "\.\./" .qoder/skills/<name>/           # cross-skill references, install the siblings first

Then run the first command in the docs. If it fails, delete the skill — it costs not disk space, but the extra twenty minutes you'll burn the next time you hit a similar problem.

The installer's final line is more honest than any leaderboard: Review skills before use; they run with full agent permissions.

Of the Skills you've installed, how many have you opened once and never opened again? Name them in the comments.

Comments

Top 2 of 4 from juejin.cn, machine-translated. The original thread is authoritative.

Wanvideo

The biggest problem with many open-source Skills is the lack of clear environment constraints and pre-validation of tool calls; switching the execution host or network environment causes them to fail outright. What actually works well are practical scripts with tightly bounded scopes.

kyriewen

Like.

汪汪队首席上单

How do you use it? What does it do?

kyriewen

Just install it per the documentation.