🔶 Agent 技能使用哪些语言编写？

2026 年第一季度，新编写的 Agent 技能中有 13.0% 使用英语以外的语言；一个季度后，这一比例达到 16.3%。在 255,068 个技能的样本中，三个月增长了 3 个百分点，且置信区间完全没有接近重叠。作为对比，GitHub 整体的非英语文档比例花了十年才从 3.7% 升至 13.0%。因此，这里发生的事情速度完全不同，最合理的解释是：AI 开发已在旧金山以外的地方兴起。

审视数据后发现，这一结论比其直观版本更强，因为英语并不等同于美国。GitHub 增长最快的开发者群体，且增长幅度显著领先其他群体，是印度；印度使用英语，尼日利亚和新加坡也是如此，因此语言统计无法识别这些开发者。非英语占比因而不能衡量这个生态中有多少位于美国以外；它只是这一比例的下限，下面所有内容都应据此理解。

作为背景，技能是一个文件夹中的 SKILL.md 文件，使用自然语言保存针对 AI Agent 的指令；当 Agent 判断任务相关时，便会加载它。Anthropic 于 2025 年 10 月发布这一规范，它像食谱一样传播：有人复制它。九个月后，在 282,200 个公开仓库中已有 380 万个此类文件，这正是 GitSkills 数据集所收集的内容。若按传统软件的定义来看，技能是一种奇特的软件，因为它正是 AI 改变的软件形态之一。它们以人类语言写成，运行时则是多语言模型，因此从技术上没有理由必须用英语编写：深圳或圣保罗的开发者可以用自己的语言更精确地描述流程，而 Agent 会遵循它。至于它是否能遵循得同样好，则是一个更好的问题，也远比文件爬取能够解决的问题困难。

分布

我们在去除 front matter 和围栏代码后，对每一份独特技能的正文进行了语言识别。

[横向条形图：语言分布。英语 85.3%，中文 6.2%，日语 1.7%，德语 1.6%，韩语 1.2%，葡萄牙语 1.1%，西班牙语 0.9%，法语 0.4%。]

共有 1,870,299 份独特技能内容。非英语内容占 14.3%，其中中文占比最高。

语言	独特技能占比
英语	85.3%
中文	6.2%
日语	1.7%
德语	1.6%
韩语	1.2%
葡萄牙语	1.1%
西班牙语	0.9%
法语	0.4%

因此，14.3% 的技能不使用英语。按文字系统划分，中文技能中有 104,985 份使用简体中文，9,112 份使用繁体中文。上表的合计与该比例并不完全一致，因为有 6,810 份技能的识别结果低于我们的置信度下限，因而未被归入任何语言。值得比较的对象是 GitHub 自身的文档，而不是 issue 或 pull request；一项 2026 年 ICSE 研究指出，仓库文档中非英语内容占 13.0%，且中文出现在 3.3% 的仓库中。整体来看，技能并不特别：14.3% 对 13.0% 基本持平。但其中文比例显著更高，为 6.2%，而非 3.3%。

为什么所有已发布数字都不一致

我们的数字不是唯一公开的数字，而这些已发布数字彼此并不一致。

报告的英语占比	语料库	方法
65.0%	557 个医疗健康技能，ClawHub（2605.02709）	未说明
81.8%	26,502 个技能，ClawHub（2604.13064）	未说明
85.3%	1,870,299 份独特内容，GitHub（我们的数据）	py3langid，conf >= 0.80
92.6%	133,149 个技能，skills.sh（2607.01456）	fast-langdetect
99.7%	以英语为种子的爬取（2606.03565）	基于种子

这并非互相矛盾，而是五个不同总体：精选市场倾向英语，特定领域的切片会偏向该领域活跃的地区，而使用英语查询作为种子的爬取会找到英语内容。第一个需要排除的候选解释是我们自己：如果我们的识别器只是比其他工具识别出的英语更少，那么整个比较就会是工具造成的伪象。因此，我们在相同文档上运行了两种工具：我们使用的 py3langid，以及 92.6% 研究所用的 fast-langdetect。两者在 97.6% 的文档上结论一致，英语占比仅相差 +1.2 个百分点，而观测到的差距约为 7 个百分点。质量筛选似乎是下一个合理候选解释，但同样无法解释：如果筛选有效 front matter 的语料库在不经意间丢弃了非英语技能，或许能解释部分差异；但非英语技能的 front matter 有效率略高，为 88.1%，英语技能为 86.6%，而筛选只会让英语占比从 85.6% 变为 85.4%。剩下的解释就是你在哪里看。这一结论远不止适用于这个数据集：当有人告诉你“AI 生态系统”的样子时，他们爬取的是哪个注册表，可能比他们说的其他任何内容都更能说明答案。

技能正变得更少使用英语

技能具有提交历史，因此每个技能都有创建日期，这使静态饼图变成了一条趋势线。

[图表：按月统计的非英语占比及置信区间，纵轴为 0%、8%、16%、24%，横轴为 2025 年 10 月至 2026 年 7 月。置信带为 95% Wilson 区间。2026 年 7 月以阴影标示：采集在月中进行，因此该队列不完整，不纳入比较。]

季度	非英语占比
2026 年 Q1	13.0% [12.8, 13.1]
2026 年 Q2	16.3% [16.1, 16.4]

按月看，这一上升并不平滑：2 月曾降至 10.9%，随后 3 月恢复至 14.2%；但在观察窗口内的方向没有疑问，1 月为 13.1%，6 月为 17.6%。对于一种仅有 18 个月历史的格式，这大致符合预期：当没有既有惯性需要克服时，新型工件会比成熟工件更快形成其人口结构。但“非英语”并不是单一类别；拆分后可以发现，增长由四个群体中的两个推动，而不是所有群体共同推动。

[小倍数图：按语言分组的季度占比及置信区间。]

[图表：中文，2025 年 Q4 至 2026 年 Q2，+2.1 个百分点。]

[图表：日语，2025 年 Q4 至 2026 年 Q2，-1.0 个百分点。]

[图表：韩语，2025 年 Q4 至 2026 年 Q2，+0.3 个百分点。]

[图表：欧洲语言，2025 年 Q4 至 2026 年 Q2，+4.2 个百分点。]

阴影带为 95% Wilson 区间；空心的最后一个数据点是被审查的 7 月队列，会绘制但绝不参与比较。欧洲语言组包括德语、法语、西班牙语、葡萄牙语、意大利语、俄语和荷兰语。

语言组	2026 年 Q1	2026 年 Q2	变化
中文	4.2% [4.1, 4.4]	5.3% [5.2, 5.4]	+1.1
欧洲语言	2.7% [2.6, 2.8]	5.5% [5.4, 5.6]	+2.8
韩语	2.2% [2.1, 2.3]	2.0% [1.9, 2.0]	-0.2
日语	3.2% [3.1, 3.3]	2.5% [2.4, 2.5]	-0.7

欧洲语言——即归为一组的德语、法语、西班牙语、葡萄牙语、意大利语、俄语和荷兰语——在观察窗口内增长逾一倍；中文则稳定上升。日语和韩语则都不是如此：日语在 2025 年末是最常见的非英语语言，但随着其他语言的加入，它在 2026 年上半年占比下滑；韩语则全程基本持平。被审查的 7 月队列暗示日语正在回升，但我们没有将其计入。变化以两个完整季度——2026 年 Q1 和 Q2——之间的差异计算，因为最后一列对应采集月份 7 月，数据不完整，所以只会出现在图表中，而不会用于比较。

我们为何相信这一趋势

这样的趋势恰恰很可能被证明是伪象，因此我们花在试图推翻它上的时间，比发现它更多。只有 24% 的技能具有提交历史，而这个子样本偏向被大量复制的技能；这很重要，因为复制与语言高度相关。换言之，担忧在于：我们看到的可能是选择效应，而不是人们写作内容的变化。在将复制次数固定为 1 的条件下，增长幅度比标题数字更大，从 14.7% 升至 18.1%；若每个仓库只计一次，以避免批量上传者左右结果，这一增长仍然存在，从 14.5% 升至 16.8%。

时钟

提交时间戳以 UTC 存储，因此在我们看到文件之前，作者的本地时区就已经消失了。但人们大多在清醒时提交；如果一组技能由世界某个地区的人编写，其提交应在该地区的夜晚消失。

[24 小时钟盘图：每种语言一个钟盘。]

46.5%

西班牙语/葡萄牙语 n=9,936

中心数字是该时间窗口内首次提交的占比。非英语群体样本较小，因此应关注对比，而不是小数点后的精确数字。

语言	东亚夜间的提交占比	n
英语	35.7%	384,979
中文	15.3%	21,939
日语	15.9%	12,908
韩语	18.7%	9,388
西班牙语/葡萄牙语	46.5%	9,936

在这个时间窗口中，中文技能的比例降至英语比例的一半以下；英语自身则在全部 24 小时内保持平坦，这是一种全球分布人群的特征，即不存在单一的夜间。西班牙语和葡萄牙语则呈现相反走势，并在 19:00 UTC 达到峰值——这对应巴西的午后和伊比利亚半岛的深夜——将这些作者定位在美洲。语言识别器不包含任何关于文件提交时间的信息，因此这两个信号是独立的，并且它们相互一致。

诚实的限制。 这是总体层面的时相估计，最多只能精确到数小时；它无法区分 UTC+8 和 UTC+9，语言并不等于国家，而且对任何个体作者完全没有说明力。我们没有找到在这一粒度上验证按小时推断方法的已发表研究，因此应将其视为佐证，而不是地理定位。原始 git 提交实际上记录了作者的 UTC 偏移量，而该数据集将其标准化掉了；这正是任何希望在此基础上继续构建的人应修复的问题。

来自外部的同一故事

我们正在读取一个平台上的一种工件类型，因此重要的问题是：是否有其他人通过测量别的对象看到了同样的变化？答案是肯定的，而且规模比我们能够观测到的更大。

GitHub 自己的 Octoverse 2025报告称，印度在一年内新增 520 万开发者，约占全球新增 3,600 万账户的 14%，使其总数达到 2,190 万并跃居全球第二。这是其 2020 年规模的 4.9 倍。同期，巴西增长 4.1 倍，印度尼西亚增长 4.8 倍，日本增长至三倍；目前 APAC 地区的新注册速度约为每分钟 25 个，欧洲则为 12 个。印度、巴西和印度尼西亚合计约占所有新账户的一半。斯坦福大学的 2026 AI Index显示，阿联酋的生成式 AI 采用率为 64%，新加坡为 61%，而美国为 28.3%，排名第 24。其政策章节记录称，来自“世界其他地区”的开源贡献如今在 GitHub 上已超过欧洲并接近美国；其教育章节发现，AI 工程技能在阿联酋、智利和南非增长最快。

这些内容没有任何一项与 Agent 技能直接相关，而这正是它们有价值的原因：三项独立测量都在观察 AI 开发发生在哪里，没有一项在查看 SKILL.md 文件，却都指向同一个方向。我们的数字是同一现象在一个只有 18 个月历史的语料库中的显现。我们的数字同样看不到其中的大部分。印度使用英语，因此 GitHub 增长最大的单一引擎对语言统计不可见；尼日利亚和新加坡也是如此。这就是为什么 14.3% 应被理解为当今这个生态中位于美国以外的比例下限。

被复制，还是被精心维护？

[图表：按复制次数统计的非英语占比。1 次复制：15.7%，n=1,483,575；2 次复制：10.8%，n=179,102；3–5 次复制：9.1%，n=135,271；6 次及以上复制：5.8%，n=72,351。覆盖全部 1,870,299 份独特内容。剔除 10 个聚合仓库，或按独特所有者计数后，差距会扩大。]

英语技能获得更多复制。 技能被复用得越多，非英语占比下降得越稳定：在从未被复制的技能中为 15.7%，在被复制 6 次或更多次的技能中则降至 5.8%。这一点需要辩护，因为 GitHub 上许多看似复制的行为实际上是归档。十个仓库持有这里所有技能文件的 14.5%，它们是注册表和镜像而非作者；因此按集中度计算，282,200 个仓库的行为更接近约 250 个仓库。这些聚合器实际上更倾向于非英语内容，占比 20.5%，而其他仓库为 13.7%，因此无论它们如何影响数据，都会与这一模式相反，而非制造该模式。排除它们后，差距仍然存在，从 15.1% 降至 5.3%；按独特所有者计数后——即某个参与者将一个技能供应到自己的十个仓库中时只算一次——差距略微扩大至 15.1% 对 4.6%。

但非英语技能获得更多修订。 这需要进行年龄校正，因为非英语技能平均更年轻，可被修改的时间更少。因此，下方比较固定了年龄，并询问至少已有 N 天历史的技能中，有多大比例会在其最初 N 天内被修订。

[图表：7 天、30 天和 90 天窗口内的修订率。7 天内：英语 14.4%，非英语 15.7%；30 天内：英语 21.7%，非英语 26.9%；90 天内：英语 28.7%，非英语 33.9%。]

在相同年龄下比较：在至少已有 N 天历史的技能中，于最初 N 天内被修订的占比。

窗口	英语	非英语
7 天	14.4%	15.7%
30 天	21.7%	26.9%
90 天	28.7%	33.9%

差距在第一个月内拉开后保持不变：一周时为 +1.3 个百分点，一个月时为 +5.2 个百分点，三个月时同样为 +5.2 个百分点。这留下了两种行为截然不同的群体：英语技能会传播——写一次、被广泛复制、之后很少再被触碰；而非英语技能则被精心维护——复制更少、修订更多。复制差异最可能的原因是搜索。发现依赖词汇，而开发者的搜

In the first quarter of 2026, 13.0% of newly written agent skills were in a language other than English, and one quarter later it was 16.3%. That is three points in three months across 255,068 skills, with confidence intervals nowhere near touching. For comparison, GitHub-wide non-English documentation took ten years to travel from 3.7% to 13.0%, so whatever is happening here is happening at a different speed entirely, and the most plausible explanation is that AI development has arrived somewhere other than San Francisco.

Reviewing the data, it turns out that the claim is stronger than the obvious version of it, because English is not a proxy for American. GitHub's fastest-growing developer population by a wide margin is India, which writes in English, as do Nigeria and Singapore, so a language count cannot see any of them. The non-English share is therefore not a measure of how much of this ecosystem sits outside the United States. It is a floor beneath it, and everything below should be read that way.

For context, a skill is a SKILL.md file in a folder, holding instructions for an AI agent in plain prose, loaded when the agent judges the task relevant. Anthropic published the specification in October 2025, and it spreads the way a recipe spreads: somebody copies it. Nine months later there were 3.8 million of them across 282,200 public repositories, which is what the GitSkills dataset collects. Skills are strange as software, by which we mean the traditional kind, because this is one of the things AI has upended. They are written in human language and the runtime is a multilingual model, so there is no technical reason to write one in English: a developer in Shenzhen or São Paulo can state a procedure more precisely in their own language, and the agent will follow it. Whether it follows it as well is a better question, and much harder to answer than anything a file crawl can settle.

the distribution

We ran language identification over the prose body of every distinct skill, after stripping front matter and fenced code.

horizontal bar chart, language distribution English 85.3% Chinese 中文 6.2% Japanese 日本語 1.7% German Deutsch 1.6% Korean 한국어 1.2% Portuguese Português 1.1% Spanish Español 0.9% French Français 0.4% 1,870,299 distinct skill contents. The 14.3% that are not English are led by Chinese. Language	Share of distinct skills
English	85.3%
Chinese	6.2%
Japanese	1.7%
German	1.6%
Korean	1.2%
Portuguese	1.1%
Spanish	0.9%
French	0.4%

So 14.3% of skills are not in English, and split by script the Chinese ones run 104,985 simplified against 9,112 traditional. The rows above do not quite sum to that, because 6,810 skills came back below our confidence floor and are counted as neither. The comparison worth making is against GitHub's own documentation instead of its issues or pull requests, and a 2026 ICSE study put repository documentation at 13.0% non-English, with Chinese at 3.3% of repositories. In aggregate that makes skills unremarkable, 14.3% against 13.0% being a dead heat. They are markedly more Chinese, though, 6.2% against 3.3%.

why every published number disagrees

Ours is not the only published figure, and the published figures do not agree with each other.

Reported English share	Corpus	Method
65.0%	557 healthcare skills, ClawHub (2605.02709)	not stated
81.8%	26,502 skills, ClawHub (2604.13064)	not stated
85.3%	1,870,299 distinct, GitHub (ours)	py3langid, conf >= 0.80
92.6%	133,149 skills, skills.sh (2607.01456)	fast-langdetect
99.7%	English-seeded crawl (2606.03565)	seeded

These are not contradictions, they are five different populations: curated marketplaces skew English, domain slices skew toward wherever that domain happens to be active, and a crawl seeded with English queries will find English. The first candidate to rule out is us, because if our identifier simply saw less English than everyone else's then the whole comparison would be an artifact of tooling. So we ran both over the same documents, py3langid which we use and fast-langdetect which the 92.6% study used. They agree on 97.6% of documents, and their English shares sit +1.2 points apart against a gap of around seven. Quality screening looks like the next good candidate and leads nowhere either: if corpora that filter for valid front matter were quietly discarding non-English skills that would explain some of the spread, but non-English skills have slightly better front-matter validity, 88.1% against 86.6%, and filtering moves the English share only from 85.6% to 85.4%. What is left is where you looked. That generalises well past this dataset, so when someone tells you what "the AI ecosystem" looks like, the registry they scraped may hold more of the answer than anything else they say.

skills are getting less English

Skills carry commit history, so each one has a creation date, and that turns a static pie chart into a trend.

non-English share by month, with confidence band 0%8%16%24% 25·1025·1125·1226·0126·0226·0326·0426·0526·0626·07 Band is the 95% Wilson interval. July 2026 is shaded: collection ran mid-month, so that cohort is censored and excluded from comparisons. Quarter	Non-English share
2026 Q1	13.0% [12.8, 13.1]
2026 Q2	16.3% [16.1, 16.4]

Month by month the climb is not smooth, since February dips to 10.9% before March resumes at 14.2%, but the direction across the window is not in doubt: 13.1% in January against 17.6% in June. That is roughly what you would expect of a format eighteen months old, since new artifact types acquire their demographics much faster than mature ones when there is no incumbency to overcome. But "non-English" is not one thing, and broken out, the rise turns out to be carried by two of the four groups rather than by all of them.

small multiples, share by quarter per language, with confidence bands

0%4%8% 25-Q326-Q3

Chinese +2.1 pts 25-Q4 → 26-Q2

0%4%8% 25-Q326-Q3

Japanese -1.0 pts 25-Q4 → 26-Q2

0%4%8% 25-Q326-Q3

Korean +0.3 pts 25-Q4 → 26-Q2

0%4%8% 25-Q326-Q3

European +4.2 pts 25-Q4 → 26-Q2

Shaded band is the 95% Wilson interval; the hollow final point is the censored July cohort, plotted but never compared. European groups German, French, Spanish, Portuguese, Italian, Russian and Dutch.	2026 Q1	2026 Q2	Change
Chinese	4.2% [4.1, 4.4]	5.3% [5.2, 5.4]	+1.1
European	2.7% [2.6, 2.8]	5.5% [5.4, 5.6]	+2.8
Korean	2.2% [2.1, 2.3]	2.0% [1.9, 2.0]	-0.2
Japanese	3.2% [3.1, 3.3]	2.5% [2.4, 2.5]	-0.7

European languages, by which we mean German, French, Spanish, Portuguese, Italian, Russian and Dutch grouped together, more than double across the window while Chinese climbs steadily, and Japanese and Korean do neither: Japanese was the most common non-English language at the end of 2025 and slipped through the first half of 2026 as everyone else arrived, while Korean stays flat throughout. The censored July cohort hints that Japanese is recovering, and we are not counting it. Changes are measured between the two complete quarters, 2026 Q1 and Q2, since the final column is the July collection month and is censored, so it appears in the chart but never in a comparison.

why we believe it

A trend like this is exactly the kind of thing that turns out to be an artifact, so we spent longer trying to break it than we did finding it. Commit history exists for only 24% of skills, and that subsample leans toward heavily copied ones, which matters because copying turns out to be strongly related to language. The worry, in other words, is that we are watching a selection effect and not a change in what people write. Holding copies fixed at one, the rise is larger than the headline, 14.7% to 18.1%; counting each repository only once, so that no bulk uploader can swing it, the rise survives at 14.5% to 16.8%.

the clock

Commit timestamps are stored in UTC, so an author's local timezone is gone before we ever see the file. But people mostly commit while they are awake, and if a group of skills is written by people in one part of the world, their commits should vanish during that region's night.

24-hour dials, one per language

46.5%

Spanish/Portuguese n=9,936

Centre figure is the share of first commits in that window. The non-English groups are small, so read the contrast, not the decimals. Language	Commits during East Asian night	n
English	35.7%	384,979
Chinese	15.3%	21,939
Japanese	15.9%	12,908
Korean	18.7%	9,388
Spanish/Portuguese	46.5%	9,936

Chinese-language skills fall to less than half the English rate in that window, while English itself stays flat across all twenty-four hours, which is the signature of a globally distributed population with no single night. Spanish and Portuguese run the opposite way and peak at 19:00 UTC, mid-afternoon in Brazil and late evening in Iberia, which places those authors in the Americas. Nothing in the language identifier knows what time a file was committed, so the two signals are independent, and they agree.

Honest limits. This is a population-level phase estimate, good to a couple of hours at best; it cannot separate UTC+8 from UTC+9, a language is not a country, and it says nothing whatsoever about any individual author. We found no published validation of hour-of-day inference at this granularity, so treat it as corroboration and not as geolocation. A raw git commit does record the author's UTC offset, and this dataset normalised it away, which is the fix for anyone building on this.

the same story from outside

We are reading one artifact type on one platform, so the question that matters is whether anyone measuring something else sees the same movement, and they do, at a larger scale than we can.

GitHub's own Octoverse 2025 reports that India added 5.2 million developers in a single year, about 14% of the 36 million accounts opened worldwide, which takes it to 21.9 million and second place globally. That is 4.9 times its 2020 population. Brazil grew 4.1 times over the same period, Indonesia 4.8 and Japan tripled, and new signups now run at roughly 25 a minute across APAC against 12 across Europe. India, Brazil and Indonesia together account for about half of all new accounts. Stanford's 2026 AI Index puts generative-AI adoption at 64% in the United Arab Emirates and 61% in Singapore, against 28.3% in the United States, which ranks twenty-fourth. Its policy chapter records open-source contributions "from the rest of the world now outpacing Europe and approaching the United States on GitHub", and its education chapter finds AI engineering skills accelerating fastest in the UAE, Chile and South Africa.

None of that is about agent skills, which is exactly what makes it useful: three independent measurements of where AI development is happening, none of them looking at SKILL.md files, all of them pointing the same way. Our number is the same phenomenon surfacing in a corpus eighteen months old. Our number also cannot see most of it. India writes in English, so the largest single engine of GitHub's growth is invisible to a language count, and so are Nigeria and Singapore, which is why 14.3% should be read as the floor beneath whatever share of this ecosystem now sits outside the United States.

copied, or tended?

non-English share by copy count 15.7% 1 copy n=1,483,575 10.8% 2 copies n=179,102 9.1% 3-5 copies n=135,271 5.8% 6+ copies n=72,351 Across all 1,870,299 distinct contents. Removing the ten aggregator repositories, or counting distinct owners, widens the gap.

English skills get copied more. The non-English share falls steadily the more a skill is reused, from 15.7% among skills nobody has ever copied down to 5.8% among those copied six or more times. That one needed defending, because a great deal of what looks like copying on GitHub is really archiving. Ten repositories hold 14.5% of every skill file here, and they are registries and mirrors rather than authors, so 282,200 repositories behave, by concentration, like about 250. Those aggregators turn out to lean non-English, 20.5% against 13.7% elsewhere, so whatever they are doing to the numbers works against this pattern instead of producing it. Excluding them leaves the gap where it was, 15.1% down to 5.3%, and counting distinct owners, so that one actor vendoring a skill into ten of their own repositories counts once, widens it slightly to 15.1% against 4.6%.

But non-English skills get revised more. That needs an age correction, since non-English skills are younger on average and have had less time to be touched, so the comparison below holds age fixed and asks what share of skills at least N days old were revised within their first N days.

revision rate at 7, 30 and 90 day windows 14.4% 15.7%within 7 days 21.7% 26.9%within 30 days 28.7% 33.9%within 90 days Compared at equal age: among skills at least N days old, the share revised within their first N days. Window	English	Non-English
7 days	14.4%	15.7%
30 days	21.7%	26.9%
90 days	28.7%	33.9%

The gap opens over the first month and then holds, +1.3 points at a week, +5.2 at a month and +5.2 at three, which leaves two populations behaving quite differently: English skills propagate, written once and copied widely and rarely touched again, while non-English skills are tended, copied less and revised more. The likely reason for the copying half is search. Discovery is lexical, and a developer searching in English will not surface a skill written in Chinese even where a multilingual model could execute it perfectly, so what stands between a Shenzhen developer's skill and the person who needs it is a text match. If that is right, the ecosystem globalises in what gets written well before it globalises in what gets reused, and the lag between the two is a tooling problem somebody could fix.

who is actually writing these?

The GitSkills authors asked one more thing worth answering: how many skills do agents write themselves? The obvious way to check fails immediately, because GitHub's own bot flag catches almost nothing: agent-written code is committed under the human's account. The signal that does work is the trailer, the Co-Authored-By line a coding agent appends to commits it authored, which is the tool claiming authorship instead of us inferring it from prose.

Measure	Value
Name an AI agent in a commit trailer	30.4% [30.3, 30.6]
Flagged as a bot by the platform	1.0%
Japanese skills, agent-authored	43.4%
Chinese skills, agent-authored	23.2%

Nearly a third of skills carry an agent's fingerprint and the platform sees almost none of it, with Claude accounting for the overwhelming majority of those trailers while Cursor, Copilot and Codex trail far behind, and the trailers even carry model versions. Read it as a floor, since a skill whose author stripped the trailer, or squashed it away, or used a tool that never emits one, counts as human here.

what this does not show

Language identification keys on script and function words, so a German skill thick with English technical vocabulary gets pulled toward English, which is another reason the non-English share is a lower bound. Dates cover only part of the corpus and only deduplication representatives, so "created" means the first commit touching that copy and not the first appearance of that content anywhere. And a crawl sees only survivors, so any skill created and deleted before July 2026 is invisible to us, which inflates every maintenance figure here by an amount we cannot estimate.

what we would want to know next

The number we cannot get from this data is whether any of it costs anything. A skill written in Chinese and never copied might be worse, or might be identical work that nobody found, and those two worlds look the same from a file crawl while implying opposite things. Separating them needs execution traces, and if somebody has those we would like to see them. The larger question is what the floor actually rests on. 14.3% of skills are not in English and the share is climbing three points a quarter, while the fastest-growing developer population on the platform writes in English and never appears in that count at all. Whatever the real number is, everything we can measure says it is moving in one direction, and faster than anything comparable has moved before.

Article 02 stays on this corpus and asks which programming languages skills talk about. Our first pass had Shell/Bash leading every language at 37.5%, which turned out to be an artifact of counting pasted commands, and measured by what a skill actually ships Python leads at 7.6% while Shell drops to 3.1%.

credit

None of this exists without the dataset, which was built and released by someone else, and all credit for collecting, deduplicating and documenting 3.8 million skill files belongs to its authors:

Giuseppe Destefanis, Daniel Graziotin, Matteo Vaccargiu, and Marco Ortu. 2027. GitSkills: A Dataset of Agent Skills on GitHub. In Proceedings of the 24th International Conference on Mining Software Repositories (MSR '27).

Preprint arXiv:2608.10906, archive 10.5281/zenodo.21875637, Parquet mirror mvaccargiu/gitskills, sample giuseppedestefanis/gitskills-sample, licence CC BY 4.0.

GitSkills is the dataset for the MSR '27 Mining Challenge, and we are not affiliated with its authors, with MSR, or with the challenge, so nothing here should be read as endorsed by them: the dataset is theirs, and the analysis and any error in it is ours. Several of the research questions we take up, including which natural languages skills are written in, are ones the GitSkills authors posed and left open.

Analysis code lives at github.com/plicara/articles under gitskills-analysis/, where every figure is generated from a single machine-readable export and never typed by hand, so the whole thing can be regenerated and checked.

Found something wrong? We would genuinely like to know.