nanaism/yomiyasu

yomiyasu（よみやす） – AI‑generated Japanese text polishing skill

What it is – yomiyasu is a skill (plugin) for LLM‑based coding assistants such as Claude Code, Cursor, or Codex. It rewrites Japanese output that sounds “AI‑smelly” into clear, information‑dense prose that a human can read without extra mental work.

Why it exists – The authors observed that large‑language models often produce Japanese with:

• vague pronouns and missing subjects,
• metaphorical verbs like 壊れる (“break”) or 溶かす (“melt”),
• buzz‑words and unnecessary bold/​list formatting, and
• overall low information density. These traits make the text awkward for technical documentation, business specs, or essays.

How it works – The skill encodes seven transformation rules that are applied to the model’s raw output:

1. Restore full SVOCM – make the subject explicit.
2. Clarify actor and action – use “please do …” or “can …” instead of vague commands.
3. Remove non‑human subjects – avoid treating tools or concepts as if they have agency.
4. Replace metaphorical verbs with concrete technical actions.
5. Strip filler introductions and turn double‑negatives into simple affirmative sentences.
6. Preserve meaning while dropping boilerplate – keep numbers, constraints, and facts, but delete AI‑generated disclaimer‑style text.
7. Control sentence length and punctuation – aim for 30‑45 characters per sentence, ≤2 commas, and minimal decorative markup.

The README includes side‑by‑side examples showing the raw AI text and the polished version after the skill is applied, illustrating how the rules improve readability.

Installation – The skill can be added with a single command:

npx skills add nanaism/yomiyasu   # Claude Code, Cursor, etc.

Alternative methods are provided for the openskills CLI, a Claude Code marketplace plugin, or manual download of the yomiyasu.skill file from the releases page.

Usage – In a chat or coding session, paste the draft and ask:

この文章を読みやすくして。
（ここに文章）

Optionally specify a domain (tech, business, essay) to bias the style.

Extra tool – The repo ships scripts/yomiyasu_lint.py, a pure‑Python linter that scores a document for “AI‑smell” (excessive bold, emojis, list density, etc.). It can be run locally or in CI to enforce a clean writing style.

Scope – This is a real, open‑source utility aimed at improving Japanese generated by LLMs. It does not train or fine‑tune models; it provides a prompt‑based transformation layer that can be plugged into existing AI coding environments.

License – MIT.