← All posts

21.8% of the Corpus Is Chinese — and Most of It Is One Author

2026-10-23

The corpus looks bilingual until you group it by author.


Scanning all 765 SKILL.md bodies for Chinese characters:

MetricValue
Skills containing Chinese167 (21.8%)
Average score, Chinese8.69
Average score, English-only8.39

A 0.30-point gap with Chinese text scoring higher. Read alone, that invites a conclusion about bilingual documentation quality. Group by author and the conclusion collapses.

Who writes the Chinese skills

AuthorChinese skills
aaron-he-zhu119
JimLiu16
yan-labs9
zenstory-ai6
wenziai6
Everyone else11

119 of 167 — 71% — come from one catalogue. That catalogue also happens to score well across every dimension, so the "Chinese scores higher" effect is mostly "one author scores higher".

This is the same trap as the category averages: a group statistic that looks like a property of a group is usually a property of who is in it.

What is actually bilingual here

Strip out the large catalogue and 48 skills contain Chinese — about 6% of the directory. Those are small, mostly single-author skills, and several are explicitly bilingual: the baoyu publishing family, translation utilities, and the China/global publishing stacks.

That is a real and coherent pocket of the ecosystem. It is just much smaller than 21.8%.

Why this matters for trigger matching

A bilingual description is not a neutral feature. The router matches on the user's phrasing, and a description that contains both languages can fire on either — which is a genuine advantage for users who code-switch.

But it also doubles the surface area of the description. A skill that tries to cover Chinese and English trigger phrases in one field runs out of room for the thing that matters most: the exclusions.

The practical takeaway

If you maintain a bilingual skill, check three things:

  1. Are both languages' trigger phrases in the description? Not just in the body.
  2. Is the exclusion stated in both? A boundary only in Chinese does not stop an English request.
  3. Did the second language crowd out the first? If adding translations pushed the "when not to use" clause out, the translation cost more than it bought.
  4. The corpus says bilingual skills are a small, deliberate subset — not an accident of the ecosystem, and not a quality signal either way.