---
title: "Four fixers for AI-sounding Japanese vs. a 137-item list"
description: "A deck that condenses 137 checklist items into 16 principles drew attention. I use it as a yardstick for yomiyasu, natural-japanese, meiseki, and textlint's AI rules."
url: https://udon-dev.com/en/notes/japanese-writing-skills-compared/
lang: en
published: 2026-11-03
topic: [ai]
tags: ["agent-skills", "claude-code", "natural-japanese", "textlint", "writing"]
---

# Four fixers for AI-sounding Japanese vs. a 137-item list

- Keisuke Tameyasu's deck rates 137 items, including the feedback he gave AI on its slides, on a five-level scale, and groups them into 16 principles and four pillars (substance, form, reasoning and evidence, and how to treat IT writing)
- For IT documents, the deck marks katakana loanwords, stiff Sino-Japanese vocabulary, hype words, and preambles as "do not fix". That is the opposite of a textlint rule that flags hype words
- yomiyasu mainly handles metaphors and subject-predicate links, natural-japanese not adding what the source lacks, meiseki syntax you can grasp in one read, and the textlint preset the shape of bullet lists and hype words
- natural-japanese and yomiyasu let the machine find candidates only; whether to fix them is decided from context. The deck's checks split the work the same way

When you tell an AI that its writing "sounds like AI," the same habits can come back in the next output. Keisuke Tameyasu's deck "[Where to judge AI versus human writing](https://speakerdeck.com/tame/tame-skill-v2-dot-0)," published on November 2, 2026, tries to stop that loop: it turns the feedback into a 137-item checklist and rules on each item one by one.

（図: The deck's cover. The title "137 checks to fix AI-like slides," the subtitle "16 principles for human-like documents," and the credit "November 2026, Keisuke Tameyasu," all in Japanese）
Source: [Where to judge AI versus human writing (Keisuke Tameyasu, Speaker Deck)](https://speakerdeck.com/tame/tame-skill-v2-dot-0)

Several other skills and checkers for fixing Japanese prose are public. Using the deck's principles as a yardstick shows that each tool targets different things, and that some tools treat the same pattern in opposite ways.

## How the deck was built, and its four pillars

The checklist items come from four sources. The core is the feedback Tameyasu gave AI while it made slides. The rest comes from Kimitaka Yokoi's talk "What is AI-smelling writing?", a report made with Deep Research titled "What is AI-like writing?", and the yomiyasu 2.4 skill.

Tameyasu rated the 137 items on a five-level scale: ◎ always follow, ◯ follow as a rule, △ suggest when noticed, × use in the opposite direction from the original item, and -. Items that give the writer the same instruction were merged into 16 principles, which are further split into four pillars. The deck's conclusion: "the substance is concrete, the form varies, and the reasoning and evidence belong to the writer."

| Pillar | Example principles | Before and after |
|---|---|---|
| Concrete substance | Say who did what and how. Use words that work in the reader's field | Replace "lean toward not keeping it" with "As a rule, we exclude items we cannot decide on" |
| Varied form | Do not repeat the same form three times. Make headings item names that show the content | Change headings shaped like "What is X: how it works and how to use it" to "How X works and how to use it" |
| Reasoning and evidence belong to the writer | Write only what the writer said and verified. Do not mention the document itself or how it was made | Delete "I keenly felt my failure," which the original remarks never said |
| Keep IT writing conventions | Leave IT writing conventions as they are. Some items are used in the opposite direction from the original | Keep "provide a solution" (the katakana loanword ソリューション) as is |

The form pillar also includes slide-specific items, such as font size (citations from 8.5pt up to 12pt) and the color of supplementary text (from light gray to a shade close to the body text). Tools that only fix prose do not cover this part.

## Items marked "do not fix" for IT documents

What sets the deck apart from other writing standards is its fourth pillar. Among patterns that look AI-like, it marks those that are common in IT documents as "do not fix."

| Pattern left alone | Example from the deck |
|---|---|
| Katakana loanwords for IT terms | "provide a solution" (ソリューションを提供する) |
| Stiff Sino-Japanese vocabulary and nominalizations common in IT | "it is possible to carry out ~" (〜を実施することが可能です) |
| Hype words and stock phrases (the deck does not say why it removed them from the fix criteria) | "overwhelming productivity" (圧倒的な生産性) |
| Preambles and descriptions of what comes next | "This article introduces AI writing" (本記事ではAIの文章について紹介します) |
| Several topics on one slide | Cause, countermeasure, and case study on one slide |
| Bullets that do not fit on one line | Bullets need not fit on one line |

One item is used in the opposite direction from the original. The original item was "explicitly separate facts, guesses, and value judgments." The deck says it is more common to write a guess as part of the sentence's content than to add a disclaimer like "this is a guess."

Forms that stack "perform" (行う) on a nominalization, such as "perform the carrying out of a configuration change" (設定変更の実施を行う), get rewritten as "change the configuration" (設定を変更する). Stiff phrasing established in IT, such as "it is possible to carry out ~," is kept. This boundary is hard to judge mechanically and uniformly, and the deck's own checks say to "treat the machine's flags as candidates only, and judge by eye whether to change the meaning."

## What the four tools cover

I compare four public tools for fixing Japanese prose whose contents I could check. Three are agent skills that you load into Claude Code or Codex, and one is a set of textlint rules. The versions I checked in the repositories on November 3, 2026, were yomiyasu v2.5.0, natural-japanese v1.6.4, meiseki v0.10.0, and the textlint preset 1.7.0.

| Tool | Main focus | Machine checks |
|---|---|---|
| [yomiyasu](https://github.com/nanaism/yomiyasu) | Metaphorical verbs, personification, subject-predicate links, commas | Bundled lint, diff, and check (Python) |
| [natural-japanese](https://github.com/coji/natural-japanese) | Not adding facts or certainty absent from the source, bold text, phrases whose referent is unclear | lint.py with morphological analysis (sudachipy) |
| [meiseki](https://github.com/bamboo-nova/meiseki) | High-load syntax such as double negatives, long prenominal modifiers, and chains of "no" (の) | A prh dictionary and the textlint preset (Node.js) |
| [textlint AI writing preset](https://github.com/textlint-ja/textlint-rule-preset-ai-writing) | Bullets with bold text and colons, bullets with emoji, hype words | Five textlint rules |

### yomiyasu

（図: The yomiyasu logo: "よみやす" in hiragana and "yomiyasu" in Latin letters）
Source: [nanaism/yomiyasu (GitHub)](https://github.com/nanaism/yomiyasu)

A skill aimed at working documents such as technical articles, design documents, and PR descriptions. The first of its seven transformation principles is to pair subjects with predicates, modifiers with what they modify, and demonstratives with what they point to. It rewrites metaphorical verbs such as "break," "tip over," "take effect," and "melt," and fixes personification that gives tools and concepts a will. The deck's "replace metaphorical verbs with the actual action" and "say who did what" nearly match this principle. That overlap is natural, since yomiyasu 2.4 is one of the deck's sources.

It sets targets for sentence length and commas: an average of 30 to 45 characters per sentence, and 0 to 2 commas per sentence. Version 2.5.0 added a check on where commas go. A measurement comparing human writing (the livedoor news corpus) with AI-generated text found that a comma right after a short object of six characters or fewer (「を、」) occurred 13.2 times per 10,000 sentences in human text and 44.6 times in AI text. The [CHANGELOG](https://github.com/nanaism/yomiyasu/blob/main/CHANGELOG.md) says the rule flags position rather than count, because some genres of human writing also use many commas. That points the same way as the deck's "put commas only at breaks in meaning."

For environments that cannot install the full skill, there are shorter versions, yomiyasu-mini and yomiyasu-pocket.

### natural-japanese

A skill for fixing meeting notes, reports, technical articles, and PR descriptions. Its rules for editing include not adding facts, numbers, opinions, or policies absent from the original, and not changing the degree of certainty. The README's example is a research report whose original notes held two unverified hypotheses. The output without the skill said they were "strongly suspected," while the output with the skill said "neither has been verified." This corresponds to the deck's "write with the certainty of the original source" and "do not claim you practiced what you did not."

In a comparison that had five models write five kinds of documents, bold text appeared 551 times across 25 outputs without the skill and 13 times across 25 outputs with it.

The README also has a section on weaknesses. It says the skill misses about half of idioms translated word by word from English and phrasings that bundle actions into nouns. Forms the deck marks as "fix," such as "perform the carrying out of a configuration change," fall within this weak area.

### meiseki

A skill narrowed to "clarity," meaning that a reader who does not know the content understands it correctly in one read. It treats syntax, not difficult vocabulary, as the cause of hard reading, and fixes double negatives, nested negation and conditions, long prenominal modifiers, distance between subject and predicate, chains of "no" (の), and excessive nominalization.

The README says it goes in the opposite direction from skills that rewrite general statements into "I think" opinions, and that it adds no voice or stance to the text. The deck has an item "include the writer's own observations, experience, and judgment" (◯ follow as a rule), which means including the writer's material when it exists. But the deck also says not to write emotions the writer did not express or things the writer did not practice. On not adding what the writer's material lacks, the deck and meiseki can be read as pointing the same way.

The same repository also holds other skills, such as yorisoi, which rewrites text into Easy Japanese (yasashii nihongo, a simplified register for non-native readers), and rikai, which makes comprehension quizzes from documents.

### textlint AI writing preset

A set of textlint rules, not an agent skill. Its principle is "constrain structure, not expression." It contains these five rules:

- `no-ai-list-formatting`: bullets with bold text and a colon, like "**Important**: ~", and bullets that start with emoji
- `no-ai-hype-expressions`: hype words such as "revolutionary," "game changer," and "paradigm shift"
- `no-ai-emphasis-patterns`: bold text in the middle of a sentence, like "This is a **very** important item," and bold headings
- `no-ai-colon-continuation`: code or bullets that follow a colon after a sentence ending in a predicate, like "Run:"
- `ai-tech-writing-guideline`: suggestions for redundant phrasing (「〜することができます」 "is able to do" to 「〜します」 "does"), passive voice, abstract wording, and more

It works with textlint 15.1.0 and later. Start it as an MCP server with the `npx textlint --mcp` command, and AI tools such as Claude Code can call it. The preset assumes a loop where AI writes, textlint checks, and AI fixes based on the results. meiseki uses this preset to detect formatting and hype words.

## Patterns the tools treat differently

Putting the deck's rulings next to the tools, three patterns split.

| Pattern | Tameyasu's checklist | Tools |
|---|---|---|
| Hype words ("overwhelming," "revolutionary") | Removed from the fix criteria | textlint's `no-ai-hype-expressions` and meiseki detect them |
| Preambles ("This article introduces ~") | Allowed | yomiyasu deletes them only when removing them changes neither the claim nor its emphasis |
| Stiff Sino-Japanese vocabulary and nominalizations ("it is possible to carry out ~") | Keep the ones common in IT | meiseki targets excessive nominalization, and textlint's `ai-tech-writing-guideline` targets 「〜することができます」 ("is able to do") |

Hype words show the largest gap between the deck and the tools. The textlint rule keeps a list of hype words and flags them uniformly, while Tameyasu's checklist ruled that they need no fix in IT documents. When you add textlint to product introductions or external-facing documents, deciding which words to exempt with the rule's `allows` setting should help you avoid flags on words you decided to keep.

Conversely, some patterns are marked "fix" by the checklist and also have fix rules in the tools. yomiyasu handles metaphorical verbs and sentences without a subject, the textlint preset and meiseki handle bullets with bold text and colons, and natural-japanese handles adding facts absent from the source.

## What to leave to the machine and what to read by eye

The deck's checks split into what the machine counts and what you read by eye.

The machine counts commas and sentence length, repeated sentence endings, "perform ~" (〜を行う) and inconsistent spelling, item numbers in the main body and references to the document itself, and the number of font sizes and colors. Reading by eye covers four things: whether title and structure patterns repeat, whether headings and conclusions connect logically, whether anything is written that was never said, and whether a first-time reader can state the key points and conclusion.

natural-japanese and yomiyasu split the work the same way. natural-japanese surfaces suspicious spots with lint.py, and the agent decides from context whether to fix them. yomiyasu's lint also offers no replacement words for its rules on phrases common in AI writing. yomiyasu's README names the limits of the earlier approach: replacing banned words only swapped them for other vague words and did not fix the sentence structure.

Choosing from the targets each README lists: to fix metaphors and subject-predicate links, use yomiyasu. To keep facts and certainty unchanged in meeting notes and reports, use natural-japanese. For documents like procedures that must come across correctly in one read, use meiseki. To stop bullet formatting and hype words in CI or the editor, use the textlint preset. If, like Tameyasu, you have decided which conventions to keep in your field, copying that list into a tool's exclusion settings first should cut down on sorting through flags.

I also wrote [Why images look AI-made at a glance](/en/notes/ai-likeness-cues/), which organizes AI-likeness cues from the research side. When skills get loaded is covered in [Everything Claude Code's 311 skills: what loads, how to pick](/en/notes/ecc-skill-selection/).