One Skill, Two Languages — The Hidden Tax on Bilingual SKILL.md Files
Our directory is an English site evaluating a multilingual ecosystem. That asymmetry turned out to be a lens: when a skill's audience is Chinese, its language is content, not packaging. When a skill tries to be both, the score pays for it.
Roughly one in ten skills in our directory carries substantial Chinese content — either written natively for Chinese platforms or as a bilingual fork of an English skill. We score them on the same six-dimension rubric as everything else, and the results split cleanly into two groups that look identical from the outside.
Group 1: Chinese-native, Chinese-targeted — scores well
The Chinese social publishing roundup found eight skills at 9.1–9.4. What they share isn't language — it's that the language is load-bearing:
baoyu-post-to-wechat— 9.3: publishing to WeChat Official Accounts. The skill must understand 公众号 formatting conventions, title norms, and platform quirks. Translating this skill to English would make it worse, because its content is the platform.baoyu-xhs-images— 9.4: Xiaohongshu card series. XHS has a visual grammar — dense text, sticker conventions, emoji patterns — that exists nowhere else. The skill encodes that grammar in the language it's used in.baoyu-wechat-summary— 9.3: WeChat group chat summarization. The input is Chinese chat text; the skill's conventions for what a "structured highlight" means are Chinese-community conventions.
For this group, "would translating it help?" is a category error. The skill is the platform, and the platform is Chinese.
Group 2: Bilingual-by-design — pays a tax
Then there are skills written to serve both audiences in one file: English trigger, Chinese examples; or worse, parallel sections that say the same thing twice. They score consistently lower on two dimensions:
Trigger quality drops. A trigger clause that must match requests in two languages is written twice as long and half as sharp. "Use when the user asks to 翻译 or translate" is a weaker discriminator than either single-language version, because it's carrying both vocabularies without committing to either.
Content quality drops. The SKILL.md budget is finite. Spending it on translated duplication means spending less on the thing that differentiates a 9 from an 8: the platform conventions, the failure cases, the checkable contract. Bilingual skills routinely have half the concrete guidance per language.
The pattern that works — and we see it in the top bilingual families — is one skill per language, linked, not merged. baoyu-translate (9.4) doesn't try to be a translation skill in two languages; it's a translation skill with a house style, and the output language is a parameter, not a section of the manual. Compare the average forked-bilingual skill at 8.2–8.6.
What this means for skill authors
Three rules falling out of the data:
- If the platform is Chinese, write Chinese. Don't dilute a WeChat skill with English sections to chase international reach — the international user installing a WeChat skill needs it to be more Chinese-platform-precise, not less.
- If the skill is language-agnostic, write English and make language a parameter. Translation, diagrams, image generation — the operation doesn't change with locale, so the skill shouldn't either.
- Never merge. Two linked skills outscore one bilingual file on every dimension we measure. The fork costs maintenance; the merge costs quality.
What it means for readers
On an English directory, Chinese-language skills are a feature, not a bug — they're the ones that actually work on Chinese platforms. Our advice for users: match the skill's language to the platform you're publishing to, not to yourself. If you're an English speaker posting to Xiaohongshu, you want the XHS skill as native as it comes, and you translate your own brief to the agent.
The directory's writing and marketing categories hold both groups; the scorecard's content-quality dimension is where the tax shows up if you want to verify the pattern yourself.