The AI Smell Is Not an English Problem

Ask an LLM to write Japanese and you get a specific kind of wrong: prose that is grammatically flawless and reads like a contract drafted by a vending machine. yomiyasu, at 9.2, exists to fix that — and it's the only skill in our directory aimed squarely at de-AI-flavoring a non-English language.

The phrase "AI-sounding text" is usually treated as an English-language complaint — the em-dashes, the "delve into," the paragraph that ends by summarizing itself. Reading yomiyasu — 9.2, a Japanese 推敲 (close-polishing) skill — convinced us the phenomenon is deeper than any one language. AI output has a register, and the register translates.

What AI flavor looks like in Japanese

In Japanese, the model's tells are different from English but equally mechanical: uniformly formal です/ます register in contexts that call for plain form, creating stiff distance in text that should feel close. Formulaic politeness — honorific scaffolding deployed in perfectly even layers regardless of who is speaking to whom. Template rhythm: sentence after sentence of identical length and construction, because the underlying token distribution smooths exactly where a human writer would vary. And connective tissue that sounds cohesive — しかし、さらに、その結果 — deployed by frequency rather than logic.

The skill's insight, and the reason it scores 9.2, is that these are diagnosable and correctable one at a time. yomiyasu doesn't say "make it natural." It decomposes the AI smell into concrete axes — register appropriateness, politeness calibration, rhythmic variation, logical connectives — and polishes each pass separately. That decomposition is what a 9.2 looks like in our scoring: the problem, named precisely enough to be solved.

Why non-English writing skills are structurally rare

Here's the uncomfortable pattern in skill collections generally: the long tail of writing skills is overwhelmingly English-first — English copywriting, English editing, English tone guides — with other languages handled implicitly, on the theory that the model's multilingual baseline is "good enough."

yomiyasu at 9.2 is evidence the baseline is not good enough, and in a predictable direction. Models are trained on less high-quality Japanese prose than English prose, so the smoothing — the averaging-away of idiosyncratic, human-feeling choices — is worse in Japanese, not better. The weaker the language's representation relative to English, the stronger the template smell. Non-English polishing skills aren't a nice-to-have layer on top of English skills; they're higher-leverage, because the disease they treat is more advanced in their target language.

The rare counterexamples prove the demand. The baoyu family's baoyu-translate — 9.4 in our directory, one of the highest scores anywhere — encodes a house translation style rather than trusting the multilingual baseline, precisely because the baseline wasn't good enough for serious bilingual publishing. And baoyu-article-illustrator at 9.1 takes the same stance for visuals: an article's images should hold to one style the way its prose holds to one voice. yomiyasu at 9.2 is the same decision, made for Japanese.

That argument generalizes beyond Japanese, and it's the argument we'd make to any skill author: the languages with the worst AI-flavor problem are exactly the languages where a serious polishing skill has the most headroom. This is one of the few areas in the directory where an underserved niche is underserved for a real reason, not an accident.

The general lesson, cross-lingual

Strip away the Japanese specifics and yomiyasu's method works everywhere. The AI smell, in any language, comes from the same three sources: uniform register where humans vary, filler cohesion where humans connect by meaning, and rhythmic flatness where humans break pattern deliberately. A polishing pass that names those three axes in its target language — with that language's own examples, not translated ones — will outperform any amount of vague "sound natural" instruction.

推敲 — the traditional Japanese craft of repeated close revision toward the one right word — is a felicitous frame for this. It treats polish not as cleanup but as the actual work of writing. The model's draft is the starting point, and the skill exists to keep the last mile from being skipped.

yomiyasu scores 9.2 for doing that in one language, precisely. Our takeaway for the ecosystem is bigger: every language deserves its own version of this skill, and almost none of them have one. The AI smell isn't an English problem — it's everywhere the models write. The medicine, so far, is mostly only in Japanese.