best-skills: nine daily agent-skill rankings over per-registry counts

Every row of the headline Top 10 is rated C

The Best 100 list is ranked by a Worth-Installing Score, and each row carries a Cov column where A means a score resting on all three registries and C means one. Every visible row in the Top 10 preview is C. `agent-browser` from vercel-labs leads with 81.4, `find-skills` from the same vendor follows at 73.7, `frontend-design` from anthropics is third at 72.2, and `microsoft-foundry` closes the preview at 64.9.

Vendors cluster at the top: vercel-labs takes four of the ten slots, anthropics and mattpocock two each, remotion-dev and microsoft one apiece. The full list of one hundred lives in `data/2026-10-02/rankings/best-100.csv`, so the preview is a slice and the coverage rating may change further down.

Still, the visible shape of the headline list is a ranking drawn from single sources, which is the part worth holding in mind when the project's stated purpose is to merge ecosystems into one cross-ecosystem picture.

Nine of the ten install leaders have data in exactly one column

The Top Installs table is the clearest statement of the problem this project is trying to solve. Its columns are skills.sh, ClawHub and SkillHub CN, and an empty cell means no data rather than zero. `find-skills` tops it with 3,661,126 installs on skills.sh and nothing elsewhere. `dev-expert` has 2,182,437 from SkillHub CN only, `parenting-expert` has 1,602,315, `tencent-docs` has 1,365,832, and `grill-me`, `grill-with-docs`, `improve-codebase-architecture`, `agent-browser` and `tdd` all carry skills.sh numbers between 1,000,365 and 1,265,692 and nothing in the other two columns.

One row breaks the pattern. `self-improving-agent` reports 482,140 on ClawHub and 1,294,264 on SkillHub CN, which makes it the only visible entry with two populated registries and the only place where a reader can compare an OpenClaw download count against a China install count directly.

No visible row fills all three columns, which is the concrete reason the percentile composite exists rather than a plain sum.

Trending lists the same skill names twice, from two vendor paths

The seven-day trending table exposes a duplication problem that the other lists do not. `ai-image-generation` appears once under the 101-skills vendor path and again under magentosh, and `twitter-automation` appears under both. Their install counts differ sharply, with the 101-skills rows clustered around 615,000 and the magentosh rows around 211,000.

Four of the 101-skills entries carry install counts of 615,716, 615,236, 614,949 and 615,241 with weekly changes of 12.8, 12.8, 12.7 and 12.7 percent. Sitting within a few hundred of each other, with identical growth rates, they look more like one pack of skills counted repeatedly than four independent hits, though the documents do not say which it is. The magentosh duplicates move faster, at 30.4 and 30.7 percent, from a lower base.

The first row is also shaped differently: `ui-taste` shows 326,977 installs with an empty weekly change column, so there is no growth figure to compare against the rest.

Percentile composite, preserved raw counts, and published judgement probabilities

The methodology commits to three things that make the numbers checkable. Cross-platform figures are never added together; they are shown side by side and ranked by a within-platform percentile composite, so a large China count cannot outrank a large global count by arithmetic alone. Every CSV keeps the original per-platform counts next to the score, which is what lets a reader verify a rank or recompute one differently.

Judgement of listings is delegated to `jev`, identified as TypeSafe's decision model at `jev-1.13` on OpenRouter, and not to keyword counts. It checks whether a social post is really about the skill, whether a listing is a real maintained skill, and which category it belongs to, then leaves placeholders, deprecated and harmful listings out of the rankings. Every probability it produces is published, and the details live in `docs/methodology.md` under a judgements section.

That combination is the strongest argument for the project: the aggregation is reproducible, and the subjective step is quantified rather than hidden.

A dated CSV directory per day, and a README generated between markers

The data layout is per-day rather than per-release. Both visible full lists sit under `data/2026-10-02/rankings/`, one file per list, and the README line above them records the same date with a UTC marker and notes that the tables are a Top 10 preview while the CSVs hold the full Top 100.

The tables inside the README are machine-written between an HTML comment marker named RANKINGS:START and a matching end, which is what makes a daily regeneration possible without hand-editing nine tables. The repository root holds seven entries: `.github/`, `LICENSE`, `README.md`, `SKILL.md`, `data/`, `docs/` and `llms.txt`. There is no releases history, so the daily commit is the versioning.

Translations are handled by separate files rather than by a build: English at the root, and Chinese, Japanese, Korean, Spanish, German and Russian under `docs/`. Whether the generated ranking tables appear in all seven is not stated.

Two consumption paths: llms.txt for a live answer, npx skills add for standing knowledge

The project is built to be read by an agent rather than by a person browsing tables, and it offers two ways in. The first is a single line pasted into Claude Code, Codex, Cursor, OpenClaw or any agent that can fetch a URL:

Read https://linkly.ai/skills/llms.txt

That file is meant to tell the agent where the CSVs live, what every column means, how to get from a question to the right file, and the rules for using the numbers.

The second path makes the rankings part of an agent's standing knowledge instead of a one-off lookup:

npx skills add https://github.com/LinklyAI/best-skills --skill best-skills

`SKILL.md` at the root is what makes that install work, and the README also links a rendered version of the rankings at linkly.ai/skills for human browsing.

The stated gap is that every registry only sees its own ecosystem

The premise is stated in the project's own terms: every skills registry sees only its own ecosystem. skills.sh counts Claude and Vercel CLI installs, ClawHub counts OpenClaw downloads, and Tencent SkillHub counts installs from China. None of them see social buzz. Merging those views is the stated contribution, with global installs, China installs and social mentions placed next to each other for each skill.

The bullets behind that claim are nine rankings refreshed daily, raw numbers preserved, no adding of apples to oranges, and the jev judgements. The description line goes further than the mechanism, naming skills.sh, ClawHub, Tencent SkillHub, GitHub, X and more than ten communities as sources, while the coverage column only spans three registries and the visible tables only show install counts from those three.

So the social and community signal is real to the pipeline and largely invisible in the published columns. That gap is where a reader should look before treating a rank as settled.

Editorial conclusion

Use best-skills if you want to see install counts from several registries side by side rather than a single leaderboard, and if you are willing to read the methodology before trusting a rank. Do not treat the headline order as a settled verdict: every row in the visible Best 100 preview is rated C, meaning its score rests on one registry. Before citing a number, open the dated CSV it came from, check the Cov column, read docs/methodology.md for how the percentile composite and the jev judgements are produced, and note that repository metadata records no detectable licence.

Frequently asked questions

Official sources

1. Issues
2. LinklyAI/best-skills on GitHub
3. Project website
4. README