LinklyAI/best-skills: a cross-registry ranking for Agent Skills, distributed as daily CSV
Daily-updated Top 100 Agent Skills rankings — installs, growth, and social buzz aggregated from skills.sh, ClawHub, Tencent SkillHub, GitHub, X and 10+ communities. Open data (CSV).
At a glance
- What is it?
- Best Skills is not a skill registry. It is a daily aggregation of nine rankings that pulls install counts from skills.sh, ClawHub and Tencent SkillHub alongside social mentions, then publishes the raw per-platform numbers as CSV. The value and the risk both come from the same design choice: nothing is normalised away.
- Who is it for?
- Adopt best-skills if you need to see a skill's install counts across skills.sh, ClawHub and Tencent SkillHub side by side, or if you want the raw CSV to re-rank yourself. Do not adopt it if you need a guaranteed schema, a stable historical series, or a licence you can hand to a legal team without a conversation.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap Best Skills is trying to fill between three registries
Each Agent Skills registry measures its own population. The README states the problem directly: skills.sh counts Claude and Vercel CLI installs, ClawHub counts OpenClaw downloads, Tencent SkillHub counts installs from China, and none of them observe social discussion. A skill that is heavily installed through one CLI and discussed on X is invisible to a ranking built on the other registry's numbers.
Best Skills merges those views. For a given skill the reader can see global installs, China installs and social mentions next to each other. The audience is narrow and specific: someone choosing which skill to install or evaluate, who wants to know whether a name that looks popular on skills.sh is also popular on ClawHub, or whether it is only popular in one region. That is a different question from "is this skill good," and the project does not claim to answer the second one.
Nine rankings, raw counts kept, and a percentile composite instead of a global total
The repository publishes nine rankings, refreshed daily. The README's preview shows four of them: Best 100 by a Worth-Installing Score (WIS), Top Installs across all ecosystems, Trending over 7 days, and Social Buzz drawn from X, HN, Bluesky and GitHub.
The mechanism that matters is how cross-platform numbers are combined. The project states it never adds apples to oranges. Cross-platform counts are displayed side by side and ranked by a within-platform percentile composite, with the methodology in docs/methodology.md. In practice that means a skill's position is derived from where it sits inside its own platform's distribution, not from a sum of installs across platforms with different scales. The Top Installs table makes the consequence visible: find-skills shows 3,332,825 on skills.sh with dashes in the ClawHub and SkillHub CN columns, while self-improving-agent shows dashes on skills.sh and 478,504 on ClawHub plus 1,183,196 on SkillHub CN. No single column is treated as the whole truth.
Every CSV keeps the per-platform original counts, so a reader can verify a rank or compute a different one. The Best 100 table also carries a Cov column, shown as C for every row in the preview, which appears to indicate coverage across sources. The README does not explain the letter in the excerpt available, so treat it as undocumented until docs/methodology.md is checked.
Getting the data: CSV paths, the live site, and what the README actually shows
There is no install step. The README points to a live interactive view at linkly.ai/skills and to dated CSV files under data/. The preview tables link to data/2026-09-10/rankings/best-100.csv, data/2026-09-10/rankings/top-installs.csv and data/2026-09-10/rankings/trending-7d.csv. A fourth list, Social Buzz, appears in the preview without a linked CSV path in the supplied excerpt.
A reasonable workflow is to fetch the dated directory for the day you care about and read the ranking you need, for example data/2026-09-10/rankings/best-100.csv. The date is part of the path, so a consumer has to know which day to request. The repository also ships localised READMEs under docs/ (README.zh-CN.md, README.ja.md, README.ko.md, README.es.md, README.de.md, README.ru.md) and a methodology document at docs/methodology.md. The README advertises a daily refresh and a badge labelled data updated that points at data/latest, which suggests a stable alias for the newest snapshot. That alias is the one thing worth confirming before writing a scheduled job against the repository, because the dated paths alone force you to track the calendar yourself.
Where the ranking breaks down: sparse coverage, negative trends, and a licence mismatch
The Top Installs table is the clearest limitation. Most rows have a number in exactly one column and dashes in the others. A skill with 3.3 million installs on skills.sh and nothing recorded elsewhere is not necessarily absent from ClawHub or SkillHub CN; it may simply not be tracked there. The README does not state whether a dash means zero, not applicable, or not measured. Anyone treating the dash as zero will draw the wrong conclusion about regional reach.
The Trending table has a second issue. Several high-install entries carry negative weekly deltas: video-edit at -2.2, ai-music at -2.1, reddit-automation at -2.3, image-to-video at -3.2, ai-image-generation at -3. The list is still labelled Trending. A ranking that surfaces declining skills under a growth heading needs the reader to check the delta column rather than the row position. Several other rows show a dash in the delta column instead of a number, which again leaves the meaning of a missing value unstated.
The licence is the third problem. The README badge says CC BY 4.0 and the LICENSE link points to the repository's licence file, but the repository metadata reports NOASSERTION, meaning the platform could not classify it automatically. Those two signals disagree. For a dataset that a company might redistribute internally, that disagreement is the thing to resolve before anything else. This is not legal advice; it is a reason to read the LICENSE file rather than the badge.
How it differs from a single-registry leaderboard
The obvious alternative is to use skills.sh's own listings, which the README names as one of its sources. The difference in approach is structural. A single registry ranks skills by that registry's install counter, so the ordering reflects one CLI's user base and one distribution curve. Best Skills takes that same counter as one input, adds ClawHub and Tencent SkillHub counters as separate inputs, and ranks by where a skill falls within each platform's distribution before combining. The output is a position on a percentile composite, which is why the Best 100 table shows compressed scores (79.6 down to 65.2 across the top ten) rather than raw install totals.
The trade-off is interpretability. A raw install count is a number anyone can reason about. A within-platform percentile composite requires reading docs/methodology.md before you can explain a ranking to a colleague. The project accepts that cost in exchange for not letting one platform's scale dominate the merged view. If you only care about skills.sh, the single-registry listing is simpler and needs no methodology document.
Maintenance cost for a consumer of the CSVs
The repository itself is a data pipeline, not a library, so there is nothing to upgrade in the dependency sense and no releases are listed. The maintenance burden lands on the consumer. Rankings shift daily, which the README states plainly, so a pinned snapshot goes stale within a day. The dated directory layout means a scheduled fetch has to construct or discover the current date path, and the data/latest alias is the only candidate for a stable URL in the material provided.
Schema stability is unverified. The preview tables show a Cov column in Best 100 with no definition, and dashes in Top Installs and Trending whose meaning is not specified. Any parser written against these files should treat missing values as unknown rather than as zero, and should tolerate columns appearing or changing meaning between snapshots. Because no release history was retrieved, there is no changelog to consult when a column shifts.
Who should use this, and what to check before depending on it
Use it if your question is comparative across ecosystems: whether a skill that dominates skills.sh has any presence on ClawHub or Tencent SkillHub, or whether a name is being discussed on X and Bluesky while showing modest install numbers. The Social Buzz table is the part with no equivalent in the source registries, and the raw per-platform columns are what make the comparison auditable.
Do not use it as a quality signal. The WIS is built from install counts and social mentions, both of which measure attention rather than whether a skill works. Do not use it where a missing value must be interpreted as zero, and do not redistribute it until the CC BY 4.0 badge and the NOASSERTION metadata are reconciled by reading the LICENSE file.
Three things to verify first: open data/2026-09-10/rankings/best-100.csv and confirm the column set matches the README preview; read docs/methodology.md to find out how the within-platform percentile composite is computed and what the Cov letter means; and confirm that data/latest resolves to the newest dated directory, since that alias is the only path in the material that does not require you to track the date yourself.
Editorial conclusion
Adopt best-skills if you need to see a skill's install counts across skills.sh, ClawHub and Tencent SkillHub side by side, or if you want the raw CSV to re-rank yourself. Do not adopt it if you need a guaranteed schema, a stable historical series, or a licence you can hand to a legal team without a conversation. Before depending on it, download data/2026-09-10/rankings/best-100.csv, read docs/methodology.md to confirm how the within-platform percentile composite is computed, and check whether the LICENSE file contains the CC BY 4.0 text the README badge claims.
Community notes