The fix for AI-flavored Japanese is syntax, not a banned-word list
yomiyasu is an agent skill that deodorizes LLM-written Japanese, repairing the sentence structure underneath instead of playing buzzword whack-a-mole.

What it does yomiyasu is an agent skill — a structured instruction set, not a program — that you load into coding agents such as Claude Code, Codex, or Cursor. When the agent writes practical Japanese (tech articles, design and spec documents, PR descriptions, internal reports), the skill pushes the output through seven transformation principles aimed at natural, information-dense prose. The before/after samples in the README are almost uncomfortably recognizable: missing subjects, concepts that quietly break things, time that gets melted, and the ceremonial いかがでしたでしょうか sign-off.
The interesting bit
The interesting part is the diagnosis: earlier fixes banned surface words like 手触り or 解像度, which just swaps one vague term for another while the broken sentence structure survives. yomiyasu targets the skeleton instead — restore the full SVOCM structure, stop letting inanimate concepts act with intent, and convert metaphorical verbs like 壊れる, 倒す, 溶かす into concrete operations and state changes. The authors also found that piling on rhetorical rules backfires — models over-interpret and invent coinages — so the skill stays deliberately narrow.
Key highlights
- Seven principles, from restoring vague demonstratives (
これ,両者) to concrete nouns, to rewriting double negatives and rhetorical contrast as plain affirmatives. - Quantified style guardrails: average sentence length of 30–45 characters, 0–2 commas per sentence, no emoji, no sentence-ending colons, restrained bold and bullets.
- A meaning-preservation rule: technical constraints and figures survive untouched, AI-typical disclaimers and filler get cut, and the model may not add subjects or specs absent from the source.
- Before/after comparisons built from open-license public documents and OSS design notes; a linked Zenn article digs into the linguistic pathology analysis and multi-corpus validation.
- Ships as a standard skill for the major agent environments — Claude Code, Codex, and Cursor are the ones the README names.
Caveats
- Japanese only; if you never write or review Japanese, there is nothing here for you.
- It is an instruction set, not an algorithm — the ceiling is the host model’s obedience to its own instructions.
- The README warns that other Japanese proofreading or style-adjustment skills enabled in the same environment can conflict with it.
Verdict If an agent drafts or reviews Japanese for you — docs, specs, PR text — this is a genuinely thoughtful fix, and the seven principles are worth reading even if you never install the skill. Everyone else can admire it from a distance.
Frequently asked
- What is nanaism/yomiyasu?
- yomiyasu is an agent skill that deodorizes LLM-written Japanese, repairing the sentence structure underneath instead of playing buzzword whack-a-mole.
- Is yomiyasu open source?
- Yes — nanaism/yomiyasu is open source, released under the MIT license.
- What language is yomiyasu written in?
- nanaism/yomiyasu is primarily written in Python.
- How popular is yomiyasu?
- nanaism/yomiyasu has 3.4k stars on GitHub and is currently accelerating.
- Where can I find yomiyasu?
- nanaism/yomiyasu is on GitHub at https://github.com/nanaism/yomiyasu.