Artificial Intelligence 5 min read 4h ago

Non-English AI Agent Skills Surge

Analysis of 1.87 million distinct AI agent skill files on GitHub finds 14.3% written in a language other than English, with Chinese the leading non-English language at 6.2%.
The non-English share rose from 13.0% in Q1 2026 to 16.3% in Q2 2026, a pace the analysts describe as far faster than any comparable shift in GitHub documentation over the past
Nearly a third of skills (30.4%) carry a commit trailer identifying an AI agent as author, with Claude named in the overwhelming majority of those trailers, while the platform's

A large-scale analysis of AI agent skill files published to GitHub has found that 14.3 per cent of nearly 1.9 million distinct contents are written in a language other than English — and the proportion is rising markedly faster than comparable shifts in software documentation have ever moved. The finding, drawn from the GitSkills dataset of 3.8 million SKILL.md files spanning 282,200 public repositories, points to an AI development ecosystem already extending well beyond its Californian origins, though the true extent of that shift is, by the analysts' own account, considerably larger than any language count can reveal.

What a skill is, and why language matters

Anthropic published the SKILL.md specification in October 2025. A skill is, in essence, a plain-prose instruction file held in a folder; an AI agent loads it when it judges the task relevant. The format spread rapidly — nine months after publication, 3.8 million skill files existed across public GitHub repositories. Unlike conventional source code, skills carry no runtime constraint on language: a multilingual model can execute instructions written in Mandarin or Portuguese as readily as in English, which means a developer in Shenzhen or São Paulo faces no technical barrier to writing in their own tongue. Whether the agent follows non-English instructions as reliably as English ones is, the analysts note, a harder question that a file crawl cannot settle.

The distribution

Running language identification across 1,870,299 distinct skill contents — after stripping front matter and fenced code — the analysis found English at 85.3 per cent, followed by Chinese at 6.2 per cent, Japanese at 1.7 per cent, German at 1.6 per cent, Korean at 1.2 per cent, Portuguese at 1.1 per cent, Spanish at 0.9 per cent and French at 0.4 per cent. The Chinese-language skills split 104,985 simplified characters against 9,112 traditional. Against GitHub's broader repository documentation, where a 2026 ICSE study found 13.0 per cent non-English with Chinese at 3.3 per cent, skills are broadly comparable in overall non-English share but markedly more Chinese — 6.2 per cent against 3.3 per cent.

Ad Unit · 728×90 / Responsive

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%

Because skills carry commit history, each file has a creation date, and the trend is unambiguous. The non-English share stood at 13.0 per cent in the first quarter of 2026 and rose to 16.3 per cent in the second — three percentage points across 255,068 skills, with confidence intervals that do not overlap. For comparison, GitHub-wide non-English documentation took ten years to travel from 3.7 per cent to 13.0 per cent. The quarterly rise is not carried evenly across languages: European languages (German, French, Spanish, Portuguese, Italian, Russian and Dutch grouped together) more than doubled their share from 2.7 per cent to 5.5 per cent, Chinese climbed from 4.2 per cent to 5.3 per cent, while Japanese fell from 3.2 per cent to 2.5 per cent and Korean remained roughly flat. This acceleration is consistent with the broader surge in AI adoption and development activity observed across multiple independent measures in recent years.

Why the published figures disagree

Several other studies have reported markedly different English shares — ranging from 65.0 per cent in a curated healthcare-skill corpus to 92.6 per cent in a skills marketplace sample and 99.7 per cent in an English-seeded crawl. The analysts argue these are not contradictions but five distinct populations: curated marketplaces skew English, domain slices reflect wherever a domain is active, and a crawl seeded with English queries will return English results. To rule out a tooling artefact, they ran both their own identifier and the one used in the 92.6 per cent study over the same documents; the two tools agreed on 97.6 per cent of documents and produced English-share estimates just 1.2 points apart — far smaller than the seven-point gap between studies. Quality filtering was also tested and found to shift the headline figure by only 0.2 points.

English skills spread; non-English skills get tended

Two behavioural differences separate the two populations. English skills are copied substantially more: among skills copied six or more times, the non-English share falls to 5.8 per cent, against 15.7 per cent among skills never copied. The analysts attribute this partly to search — discovery on GitHub is lexical, so a skill written in Chinese will not surface for a developer searching in English, even though a multilingual model could execute it. Non-English skills, meanwhile, are revised more frequently: at a 90-day window, 33.9 per cent of non-English skills had been revised, against 28.7 per cent of English ones, with the gap opening across the first month and holding. The implication is that the ecosystem is globalising in what gets written well before it globalises in what gets reused. This has parallels to broader questions about how human authors engage with AI-generated outputs — a lag driven, in this case, by discoverability rather than comprehension.

A third of skills carry an agent's fingerprint

The analysts also examined how many skills were written by AI agents rather than humans. The platform's own bot flag catches just 1.0 per cent of files, because agent-written code is typically committed under a human account. The more reliable signal is the commit trailer — the Co-Authored-By line that a coding agent appends to commits it authored. By that measure, 30.4 per cent of skills carry an agent's fingerprint, with Claude accounting for the overwhelming majority of those trailers; Cursor, Copilot and Codex trail well behind. Japanese skills are the most heavily agent-authored at 43.4 per cent; Chinese skills stand at 23.2 per cent. The analysts caution that any skill whose trailer was stripped or squashed counts as human in their data, making 30.4 per cent a floor.

The floor beneath the real number

The 14.3 per cent non-English figure is, by the analysts' explicit framing, a lower bound. India — GitHub's fastest-growing developer population by a wide margin, having added 5.2 million accounts in a single year to reach 21.9 million — writes in English and is invisible to any language count. So are Nigeria and Singapore. GitHub's own Octoverse 2025 data show new sign-ups running at roughly 25 a minute across Asia-Pacific against 12 across Europe, and Stanford's 2026 AI Index places generative-AI adoption at 64 per cent in the United Arab Emirates and 61 per cent in Singapore, against 28.3 per cent in the United States, which ranks twenty-fourth. The analysts note that three independent measurements of where AI development is happening — none of them examining SKILL.md files — all point in the same direction. The language share visible in the corpus, they conclude, understates the geographical diversification considerably, and the gap between what can be measured and what is actually happening is itself a feature of the data worth attending to. The choice of registry or corpus, as they put it, may hold more of the answer about what the AI ecosystem looks like than anything else a researcher reports. This theme — of understanding the limits of what AI-generated artefacts reveal — runs through much of the current debate about agent behaviour and its measurement.

##AgentSkills ##GitHub ##AIEcosystem

Ad Unit · 300×250 / Responsive