yomiyasu: an Agent Skill that de-stinks AI-generated Japanese
What it does
yomiyasu (よみやす, roughly "easy to read") is an Agent Skill built to de-stink AI-generated Japanese. Let's be honest: when you ask Claude, ChatGPT, or Gemini to write in Japanese, the result almost always carries a distinct "AI smell" — odd word choices, vague subjects, an obsession with bold text and bullet points, and, ironically, a lower information density than a human would produce. yomiyasu takes that unnatural Japanese and revises it into prose a human actually enjoys reading while keeping the information density high. It is designed for practical documents such as technical articles, design specs, PR descriptions, and internal reports, and you load it straight into AI coding environments like Codex, Claude Code, and Cursor without switching tools. Crucially, it is not a separate app you paste text into; it lives inside the agent loop, so the same assistant that drafted your text can immediately refine it.
Why it is trending
The project went open source at the end of September 2026 and reached more than 1,200 stars in about three days, which is a fast climb. The reason is refreshingly simple: almost everyone who writes Japanese with AI has hit the "this reads weird" wall, yet most people could only fight it with lists of banned words that treat the symptom and not the disease. The author did not stop at complaining. They published a long write-up that dissects the linguistics of "AI-smelly Japanese" into four pathologies — banned words only swap surface vocabulary, over-applied editing rules invent new weird coinages, vague subject-object relations paired with inanimate subjects, and formatting-heavy output that lowers information density — and even validated the claims across several corpora. That "serious to the point of being pedantic" attitude plays very well on Hacker News and X, so the project spread quickly.
Technical highlights
What makes yomiyasu special is that instead of banning words, it treats the sentence skeleton — the syntactic structure — as seven transformation principles and applies them systematically across major LLM environments. For example, the first principle is "full restoration of SVOCM": it forces the model to name the actor (developer, operator, system) explicitly and replaces vague pronouns like "it", "this", or "the mechanism" so you never get a sentence where a concept is dragged around by a metaphorical verb. It also fixes the "over-applied editing rules" problem that produces strange coinages, and it re-compresses the core information that bold text and bullet points diluted. In other words, it does not swap surface words; it operates at the syntactic level, so the result genuinely reads like a human wrote it rather than like a fresh batch of AI vocabulary. Beyond SVOCM, the other principles cover clarifying modifier-head relations, recovering information density that formatting ate, and de-formatting prose back into plain narrative — together they form a repeatable recipe rather than a vibe. It supports Codex, Claude Code, Cursor, and Gemini, and it ships under the MIT license, so you can fork and tweak it without guilt.
Who it is for
If you write technical blogs, design docs, or PR descriptions in Japanese on a daily basis, or you are an engineer at a Japanese company who needs to polish an AI draft into a report you can ship, this Skill is close to a must-have. Picture a PR description: the model drafts it, you mount yomiyasu, and the final text names who changed what and why instead of leaning on "this improves the behavior". It is especially smooth for people who already live in Claude Code, Cursor, or Codex — you just mount it and call it, polish the Japanese right after writing the code, and never open a separate translation or polishing website again.
Quick try
Load it as an Agent Skill into your coding environment. For Claude Code, run: git clone https://github.com/nanaism/yomiyasu ~/.claude/skills/yomiyasu , then invoke yomiyasu in a conversation and paste the Japanese you want polished. The same Skill folder works in Cursor and Codex with their respective skill loaders.
Versus alternatives
Existing "Japanese naturalization" approaches fall into two buckets: prompts that ban a list of words, and generic polishing Skills. yomiyasu differs by being linguistics-driven and corpus-validated. The first bucket only swaps surface words; the second leans on the model's free interpretation and often overcorrects. yomiyasu constrains the rewrite with seven syntactic principles, so it neither misses weird sentences nor invents new ones. Compared with a generic grammar checker like textlint, it reasons about meaning and actor clarity, not just surface rules; and compared with simply asking the model "make this natural", it applies a fixed, auditable recipe so the output is consistent across runs. For serious work documents, that predictable, reproducible quality beats luck-based polishing, and it also spares you from hand-writing a long style constraint every single time.