Skillselion Research · Edition of 2026-10-23

The Agent Skill Language Census

What language are AI agent skills written in? Almost all use Latin script. On 2026-10-23 we read the stored SKILL.md text of 15,739 skills in the Skillselion catalog: 909 (5.8%) have at least 30 letters, and at least 10% of their letters, outside Latin script, and they hold 15.4% of those skills' installs. Three Lark accounts carry most of that share.

909

skills whose SKILL.md text has at least 30 non-Latin letters, making up at least 10% of all its letters

15.4%

of installs on documented skills go to them

91.3%

of those installs come from three Lark accounts

1.6%

non-Latin install share once those three accounts are set aside

We dumped all 86,544 catalog listings, kept the 63,961 skills, and classified the SKILL.md text the catalog stores for 15,739 of them by Unicode script. Those 15,739 skills hold 95.1% of all 229,193,994 skill installs. Skills labelled Chinese make up 741 of the 909 non-Latin skills, followed by Korean (103) and Japanese (43). Almost all of them appear in our catalog under a Latin-script summary. Every number recomputes from the dataset; the counting rules and their limits are in the methodology.

Finding 01

How many AI agent skills are written in a non-Latin script?

Of the 15,739 skills whose SKILL.md text the catalog stores, 909 (5.8%) have at least 30 letters, and at least 10% of their letters, outside Latin script, and this page calls them non-Latin skills. The bar sits well below a majority: in 349 of the 909, non-Latin letters make up less than half of all letters. They carry 33,558,489 installs, 15.4% of the 217,952,427 installs on documented skills. Their share of installs is larger than their share of listings, and Finding 03 explains most of the gap. A different cut-off moves the result only a little. A looser rule (10 letters, 5%) gives 6.6% of skills and 16.1% of installs, a stricter one (100 letters, 25%) 4.7% and 15.1%.

5.8%

of documented skills have at least 30 letters, and at least 10% of their letters, outside Latin script

Finding 02

Which non-Latin languages do agent skills use?

741 of the 909 non-Latin skills (81.5%) are labelled Chinese, meaning Han is their largest non-Latin script and they have fewer than 10 kana letters. They come from 201 publishing accounts and hold 98.4% of all non-Latin installs. Korean follows with 103 skills from 14 accounts, then Japanese with 43 from 14 accounts and Hebrew with 19 (all from 1 account). 2 skills are labelled Cyrillic and 1 Arabic. The census counts scripts, so Russian and Ukrainian would share one Cyrillic row, and a skill written in Spanish, German or French counts as Latin script (see the methodology).

Chinese (33,006,740 installs)741 skills
Korean (433,807 installs)103 skills
Japanese (106,093 installs)43 skills
Hebrew script (4,126 installs)19 skills
Cyrillic script (1,195 installs)2 skills
Arabic script (6,528 installs)1 skill

Finding 03

How much of the non-Latin share comes from Lark?

Most of it. Three publishing accounts with Lark or Feishu names have stored SKILL.md text in the catalog: larksuite, the GitHub organization whose profile names Lark Technologies Pte. Ltd. and which hosts Lark's cli repository, and open.feishu.cn and open.larksuite.com, which publish skills from the Feishu and Lark open-platform domains (Feishu is the name Lark uses in China). Another account named feishu lists 27 lark-named skills with no stored text, so they are not classified; its GitHub profile does not say who runs it. Together they publish 86 Chinese-labelled skills with 30,632,009 installs: 92.8% of all Chinese-labelled installs and 91.3% of every non-Latin install in the census. open.feishu.cn contributes 25 of those skills and 19,520,090 installs, larksuite 36 skills and 10,942,638 installs, and open.larksuite.com 25 skills and 169,281 installs. The stored copy of lark-doc in larksuite's cli repository, for example, is 46% Han characters by letter count. The same three accounts also publish 11 documented skills labelled Latin script. Install counts here are registry counts: a suite installed as a bundle can record an install on each of its skills, so treat these totals as registry events, not separate users.

Set those three accounts aside and the share falls sharply: 823 of the remaining 15,642 documented skills (5.3%) are non-Latin skills, holding 2,926,480 of 184,772,795 installs, or 1.6%.

91.3%

of non-Latin installs come from three Lark accounts

Finding 04

Who writes non-Latin skills outside the Lark accounts?

Outside those three accounts, 655 Chinese-labelled skills come from 198 accounts and share 2,374,731 installs. The largest of them by installs is dontbesilent2025, with 33 Chinese-labelled skills and 673,831 installs on them. The Korean set leans on one project: NomaDamas publishes 68 of the 103 Korean skills, and they hold 91.5% of Korean installs (396,820). Its k-skill collection includes skills named hwp, korean-law-search and delivery-tracking. In Japanese, InterfaceX-co-jp's 7 skills, all from its genshijin repository, hold 78.8% of Japanese installs (83,561). As in Finding 03, these are registry install counts, and skills from one repository installed together each record an install.

Among documented skills, script and adoption show no clear link at the median: the median non-Latin skill has 1,340 installs and the median Latin-script skill has 1,337.5.

1,340 vs 1,337.5

median installs, non-Latin vs Latin-script documented skills

Finding 05

Does the Skillselion catalog show these skills in their own script?

Mostly no, and that is a gap in our own catalog. 891 of the 909 non-Latin skills (98.0%) appear on Skillselion under a Latin-script summary, and across all 63,961 skill listings only 226 catalog descriptions (0.4%) reach the description non-Latin bar set out in the methodology. So for 891 of those 909 skills, a reader who searches our catalog sees Latin-script listing text, although the stored SKILL.md text has at least 30 non-Latin letters making up at least 10% of its letters. In two categories non-Latin skills hold most of the documented installs: 74.7% in Office & Documents and 57.7% in Automation & Workflows.

98.0%

of non-Latin skills carry a Latin-script summary in the Skillselion catalog

Method

Methodology

The census is a complete dump of the Skillselion production catalog taken on 2026-10-23: 86,544 listings in total, of which the 63,961 skill listings form this census, with 229,193,994 cumulative installs aggregated from public registry data, chiefly the skills.sh registry. Installs count registry install events, not active users. The text we classify is the SKILL.md excerpt the catalog stores for each skill. It exists for 15,739 skills (24.6%), and those skills hold 95.1% of all installs. The 48,222 skills without stored text sit mostly in the long tail, with a median of 43 installs against 1,338 for documented skills. Every share on this page is a share of documented skills unless it says otherwise.

Before counting, we normalize line endings and remove fenced code blocks (a fence left open by the end of a stored excerpt is removed through to the end), inline code, URLs and HTML tags (a tag may span lines), because code is written in Latin script whatever the language of the prose. We then count letters (Unicode general category L) by the script prefix of each character's Unicode name: Latin, Han (CJK ideographs), kana (Hiragana and Katakana), Hangul, Cyrillic, Arabic, Hebrew, Thai, Devanagari and Greek, with any other letter counted in a non-Latin "other" group. The groups follow the script names in Unicode Standard Annex #24. A skill counts as non-Latin when at least 30 of its letters, and at least 10% of all its letters, are outside Latin script. Any such skill with 10 or more kana letters is labelled Japanese; otherwise the label follows the largest non-Latin script, with Han read as Chinese and Hangul as Korean, so Japanese text written only in kanji would count as Chinese. The looser and stricter rules in Finding 01 show how far the cut-off moves the result. A catalog description counts as non-Latin when at least 5 of its letters, and at least 10% of all its letters, are outside Latin script. Publishing accounts are counted case-insensitively, because GitHub owner names are case-insensitive.

Script is not language. Latin script covers English, Spanish, German and many others, and this census does not split them. The language tag the catalog stores covers 9,500 skills (14.9%): of its 8,818 Latin-script language tags, 8,796 (99.8%) are English. On the 2,453 skills that have both a tag and stored text, the tag and our script detection agree on 2,355 (96.0%). Listings are counted as the catalog records them, so copies of the same skill each count once; duplication is measured in The Agent Skill Clone Census. Every number on this page recomputes from the dataset below with the verification script described in How Skillselion Research is done.

Skillselion is an independent project and is not affiliated with Anthropic, OpenAI, Cursor, Claude, Claude Code, Codex, Lark, Feishu, or any other publisher named on this page.

Data

Download the dataset

The dataset has 63,961 rows, one per skill listing, with the publishing account, category, installs, whether SKILL.md text is stored, letter counts per Unicode script group, the language label, the catalog's language tag and the script of the catalog description. It is free to use with attribution under CC BY 4.0: cite "Skillselion Agent Skill Language Census, October 2026" and link this page.

The file ships with the next Skillselion site release, and this section will carry the download link once the file is served.

FAQ

Questions journalists ask

What language are AI agent skills written in?

Mostly Latin script. Of the 15,739 skills whose SKILL.md text the Skillselion catalog stored on 2026-10-23, 909 (5.8%) have at least 30 letters, and at least 10% of their letters, outside Latin script, and the other 14,830 do not. The 909 non-Latin skills hold 15.4% of installs on documented skills, and 741 of them are labelled Chinese.

How many AI agent skills are written in Chinese?

741 of the 15,739 skills whose SKILL.md text the Skillselion catalog stored on 2026-10-23 are labelled Chinese: at least 30 of their letters, and at least 10% of all their letters, are outside Latin script, Han is their largest non-Latin script, and they have fewer than 10 kana letters. They come from 201 publishing accounts and have 33,006,740 installs. Three accounts with Lark or Feishu names, open.feishu.cn, larksuite and open.larksuite.com, publish 86 of them and hold 92.8% of those installs. The other 655 Chinese-labelled skills come from 198 accounts. The 48,222 skills without stored text were not classified.

Why does the census report scripts instead of languages?

Script can be counted from the text alone, and anyone can re-check the count. Telling English from Spanish needs a language classifier, which this census does not use. The language tag the catalog stores covers 14.9% of skills and agrees with our script detection on 96.0% of the skills where both exist.

Can I reuse these numbers or the dataset?

Yes. Cite "Skillselion Agent Skill Language Census, October 2026" and link this page. The raw dataset (63,961 rows, one per skill listing, with per-script letter counts and the language label) will be released under CC BY 4.0 with the next site release, and every figure recomputes from it with the verification script described in our methodology.