yomiyasu

yomiyasu

Popular

Agent Skill for rewriting unnatural AI-generated Japanese into human-readable, high-density natural prose. Use for requests like "make this text readable," "remove the AI feel," "eliminate AI-isms," "make it natural Japanese," or "deodorize the text," as well as when polishing technical articles, business specifications, PR descriptions, essays, or note posts. It dismantles non-living subjects, concretizes metaphorical verbs, completely eliminates emojis and trailing colons, removes unnecessary explanatory parentheses, removes unnatural half-width spaces around English words, and flattens excessive bolding, bullet points, and negative contrasts to reconstruct sentences that convey who did what to whom within each sentence.

962stars
19forks
Updated 10/1/2026
SKILL.md
read-only
name
yomiyasu
description

Agent Skill for rewriting unnatural AI-generated Japanese into human-readable, high-density natural prose. Use for requests like "make this text readable," "remove the AI feel," "eliminate AI-isms," "make it natural Japanese," or "deodorize the text," as well as when polishing technical articles, business specifications, PR descriptions, essays, or note posts. It dismantles non-living subjects, concretizes metaphorical verbs, completely eliminates emojis and trailing colons, removes unnecessary explanatory parentheses, removes unnatural half-width spaces around English words, and flattens excessive bolding, bullet points, and negative contrasts to reconstruct sentences that convey who did what to whom within each sentence.

yomiyasu

A skill that dismantles AI-specific hollow prose, unnatural metaphors, and broken SVOCM (subject-verb-object-complement-modifier relationships) to refine text into natural Japanese that humans can read without friction.

Note: If other agent skills intended for style adjustment or Japanese proofreading are enabled in the same environment, conflicting instructions may garble the output. Consider disabling similar skills when using this skill.


1. Core Design Philosophy

Based on the latest academic research and a quantitative study of 70,000 Qiita articles, this skill reconstructs text according to the following principles.
The highest-priority rules of this skill are "Complete Preservation of Meaning (do not change claims, weight, strength of assertions, or sentence functions)" and "No Unnecessary Addition of Information (do not add to the text)."
It leaves what is being said unchanged and fixes only the AI-ness (unnatural metaphors, empty decorations, distorted syntax, and awkward sentence transitions).

Top-Priority Rule: Preservation of Meaning

Before rewriting, verify the following 4 aspects of the original sentence and ensure they remain identical after rewriting. All other principles apply only within the scope of keeping these 4 aspects unchanged.

  1. Claim (What is being said): Maintain the main point and logical relationships you wish to convey. Do not substitute points of discussion like "essence" or "core" with different concepts such as "purpose," nor add verbs arbitrarily to change the point.
  2. Weight (What is treated as most important and what is treated lightly): Do not change things that were negated into parallel listings (e.g., "in addition to A, B"), nor arbitrarily alter rankings.
  3. Strength of Assertions (Definitive statement, conjecture, possibility): Keep definitive statements as definitive statements, and write conjectures or possibilities with their original degree of certainty (do not weaken to "there is a risk of" or raise "it can be said that" to a definitive statement).
  4. Sentence Function (Evaluation, explanation, request, plan, impression): Keep evaluation sentences as evaluations, and explanations as explanations (do not turn sentences of evaluation/property such as "... is important" or "the essence of ... is" into target/plan sentences like "aims to ..." or "enhances ..."). Preserve the function of sentence endings ("let's ...", "please ...", "will ...", "is ...") in line with the original sentence type.

Logical Structure of Paragraphs and Sentences (Connections)

AI-generated text often has ambiguous sentence connections, contains empty preview sentences, or mixes multiple topics within a single paragraph. Structure the logical flow using the following guidelines:

  1. Verifying Connections Between Sentences: For transitional words (however, but, also, and, in short, therefore, etc.), the thematic particle "mo", and demonstratives at the head of a sentence (this, that, these, etc.), always verify what is being connected across adjacent sentences. If the grounds for connection are clear from context, adjust to an appropriate transitional expression; if unclear, do not force a change and clarify with the writer. Do not rely solely on "does it read naturally," but identify logical targets.
  2. Resolving Empty Preview Sentences: For preview sentences lacking content (e.g., "There are points to note," "The following points can be listed"), combine them into a single sentence with the following content sentence while retaining the weight carried by the preview (such as "points to note") in the predicate. If combining into one sentence alters the meaning, leave them uncombined.
  3. Enforcing One Topic per Paragraph: A paragraph consists of a single topic. Do not casually combine paragraphs with different topics. Conversely, if multiple topics are mixed in a single paragraph, split the paragraph appropriately without disrupting sentence order.
  4. Subject-Predicate Alignment: For sentences where subject-predicate alignment is broken, align them naturally to the extent that the original sentence's intent can be read.

Complete Restoration of SVOCM

Go beyond simple word substitutions to clarify "who did what to whom." However, supplement omitted objects or targets only when those words are explicitly stated in preceding or following sentences. Do not invent and add words from context that are unknown. Restoring demonstrative pronouns like "this" or "one side" to concrete nouns is also limited to cases where the referent is clear from context. If words were supplemented, state them explicitly in the output.

Clarifying Agents and Actions

As a rule, maintain sentence-ending functions as in the original text. Do not arbitrarily change types among "let's ...", "please ...", "will ...", or "is ...". The rule to organize requests to the reader as "please ..." and functional explanations as "can ..." applies only when the original sentence is ambiguous between a request and an explanation. Do not arbitrarily convert definitive explanatory sentences into capability expressions ("can ...").

Dismantling Non-Living Subjects (Personification)

Fixing non-living subjects is limited strictly to anthropomorphic expressions where tools or concepts are given feelings or willpower (e.g., "the code speaks," "the system hopes"). Sentences that objectively state the function of tools or mechanisms (e.g., "this tool collects logs") should remain as they are. When fixing, do not add conditions not present in the original ("if you use ...") or capability ("can ...").

Concretizing Metaphorical Verbs and Changing to Everyday Words

Rewrite AI-specific metaphorical expressions like "defeat," "work on/take effect," "melt," or "crush" into everyday words or terms that fit the context. Do not overuse overly stiff two-kanji compound nouns like "occurrence" or "guarantee" in unnecessary situations.
Always preserve the nuances (feelings or evaluative direction) held by metaphors or idioms.

  • Examples of nuance: carelessness in "unintentionally," the feeling of being tired of waiting in "finally," regret in "ended up ...ing," or the quiet yet undeniable presence in "subtly."
  • Metaphorical words may be rephrased. However, the nuance or reason conveyed by that word must be supplemented using other everyday words.
  • Handle adverbs and introductory phrasing in the same manner as metaphorical verbs.
    Change "break/ruin" used against data or systems to everyday words of equivalent breadth, such as "malfunction" or "become unusable." Limit writing "data integrity is lost" or "inconsistency occurs" strictly to cases where it is clear from the original text or context that integrity is being discussed. Leave literal instances of "break" for physical objects like clocks or machines, or bodily ailments (upset stomach, etc.), unchanged. Furthermore, soften vagueness like "quietly" or "without saying anything" to the same scope of unnoticeability, such as "before realizing it" or "without knowing." Limit writing from a mechanism perspective like "without issuing an error" or "without notification" strictly to cases where it is clear from the original text or context that errors or notifications are not issued.
    In addition, idiomatic expressions established in everyday Japanese like "break a bone (exert great effort)," "burn hands (have trouble with)," "scales fall from eyes," or "make neck long while waiting" should not be treated as AI-like metaphors and should be preserved. Even in business specifications, leave them if there is no risk of misinterpreting the meaning. When in doubt, the guideline is whether that phrasing was commonly seen in human writing before AI spread; if it was seen, keep it.

Organizing Self-Labeling and Negative Contrast

When phrases like "what is important is" or "what matters is" carry the evaluation itself, do not erase the evaluation; move it to the predicate and retain it (e.g., "... is important," "... matters"). Trimming introductory text is permissible only when doing so does not change the core claim.
Convert negative contrasts (B, not A) into affirmative statements only when removing the negative does not alter the claim. For negatives carrying meaning or weight—such as rephrasing what tends to be thought of as A into B—do not force them into affirmative statements; keep them negative and refine only the phrasing. Do not list negated A alongside B as "in addition to A, B" or rank them as "B rather than A." When keeping a negative, do not add explanatory sentences for reasons not in the original.

Meaning Preservation and Non-Addition of Information (Do Not Add)

Do not add agents, nouns, actions, causes, conditions, numerical values, examples, or technical jargon that are not evident from the original sentence or context. Do not arbitrarily replace vague words with narrower, specific facts. When information is insufficient to write specifically, do not supplement it on your own in the main body; present it at the end of the output under "Points to confirm with the writer."

Natural Sentence Length and Punctuation Control

Maintain an average sentence length of 30–45 characters, and suppress commas (touten) to 0–2 per sentence to create a smooth reading experience.

Complete Elimination of Formatting Symbols and Unnecessary Spaces

Completely eliminate emojis, trailing colons, em dashes, non-informative rephrasing parentheses, and unnatural half-width spaces between Japanese and Latin characters.


2. Execution Steps

Step 1: Understanding Context and Paragraphs

Check the overall flow of the text and paragraphs before rewriting.

  1. Determining the Domain: Determine the domain from the input text or instructions.
    • tech (Technical Articles): Technical blogs or architecture documents. Keep bullet points to 15% or less, organizing procedures within the scope of information present in the original text or materials.
    • business (Business/Specifications): PR descriptions or internal reports. Retain the reasons for metaphors while changing to objective expressions, explicitly stating boundary conditions and responsible entities within the scope known from the original text.
    • essay (Essays/Personal Posts): Note posts or personal memos. Avoid exaggerated moralizing, retaining simple emotions and concrete experiences.
  2. Understanding Paragraph Topics and Sentence Relationships:
    • Grasp "what is being discussed" for each paragraph in one line. Do not combine paragraphs with different topics, and consider splitting paragraphs where topics are mixed.
    • For each sentence, verify its relationship to the preceding sentence (reason, example, proviso, rephrasing, supplement, summary, etc.) (this note is not included in the output).
  3. Verifying the 4 Aspects of Meaning: Confirm the original sentence's "Claim," "Weight," "Strength of Assertions," and "Sentence Function."

Step 2: Executing Syntactic Transformation

Transform according to references/gemini-syntax.md and domain-specific guidelines using the following steps:

  1. Clarify agents (subjects) and targets within the scope known from the original text. Limit supplementing omitted objects or targets strictly to cases where those words appear in surrounding sentences (if supplemented, list them in "Changes made").
  2. Maintain sentence-ending functions as in the original text. Preserve types among "let's ...", "please ...", "will ...", and "is ...". Restructure into "please ..." or "can ..." only when ambiguous between a request and an explanation, and do not arbitrarily change definitive statements into "can ...".
  3. Restore demonstrative pronouns (this, one side, etc.) to concrete nouns only when the referent is clear from context.
  4. Replace metaphorical verbs and idioms with everyday words or terms fitting the context. Change "break" for data or systems to words of equivalent breadth like "malfunction" or "become unusable," using "integrity is lost" only when the context of integrity is clear. Maintain established idioms and nuances. Preserve established idioms and nuances.
  5. Trim redundant explanatory parentheses and remove extra half-width spaces around English words.
  6. When "what is important is" carries evaluation, retain it in the predicate; if negative contrast (B, not A) carries meaning or weight, adjust the expression while keeping the negative.
  7. Eliminate emojis, trailing colons, and em dashes.
  8. Rhythm adjustments at sentence endings should be performed only when 3 consecutive sentences end identically while fixing AI-ness. Do not change sentence endings for natural text that lacks AI-ness. If there are few places to fix, do not force rewrites.
  9. When converting bullet points into continuous body text, use the sentence-ending function of each item as-is. Do not add evaluative or obligatory phrasing like "... is vital" or "it is necessary to ...". If items are merely listed, leaving them as bullet points is fine.
  10. For transitional words (however, but, also, and, in short, etc.), the thematic particle "mo", and sentence-starting demonstratives, verify the connections before and after to ensure proper alignment (if corrected, list in "Changes made"; if indeterminate, do not force a fix and list in "Points to confirm with the writer").
  11. For empty preview sentences, consolidate them into a single sentence with the content sentence while retaining the weight of the preview in the predicate (if consolidated, list in "Changes made"; if consolidating changes the meaning, leave them and list in "AI-like parts retained").
  12. For sentences with misaligned subject and predicate, align them within the scope of understanding the original intent (if fixed, list in "Changes made").

Step 3: Static Inspection (Linter)

When full mode is specified or when saving files, run the bundled linter to perform static inspection.

# Resolve the absolute path of the script based on the skill installation directory (e.g., ${CLAUDE_SKILL_DIR}) and execute
python3 <skill_dir>/scripts/yomiyasu_lint.py <target_file>

Detection results serve as mechanical review candidates. If technical terms or factual descriptions are contextually justified, retain them without forcing rephrasing. In addition, do not forcibly erase words carrying evaluation like "is important" or negatives necessary for the claim (B, not A) even if flagged by the linter. Limit modification attempts to a maximum of 2 iterations to prevent excessive rephrasing loops meant solely to clear warnings.

Step 4: Inspecting Additions and Deletions (Diff Inspection)

Once the rewritten text is complete, compare the original text with the rewritten text to check for unintended information changes or broken transitions.

# Save original and rewritten text to files and run the diff script
python3 <skill_dir>/scripts/yomiyasu_diff.py original.txt rewritten.txt

Check the candidates output by the script one by one and evaluate the following:

  • Increase/decrease in phrasing types (request, obligation, evaluation, capability, etc.): If it strays from the original sentence's function, revert to the original expression; if it is a valid rephrasing, keep it and list it under "Changes made."
  • Words not in original or deleted words: Check whether information was arbitrarily added or necessary premises were dropped.
  • Changes in bullet points or paragraphs: When converting bullet points to body text, check whether unnecessary evaluations or obligations were added, or if topics were mixed due to paragraph merging.
  • Verifying connection points: Confirm whether transitional words and demonstratives logically connect with preceding and following sentences.

Perform corrections only once without repeating script re-executions. In environments where Python cannot be executed, visually inspect from the same perspectives (phrasing shifts, added/deleted words, paragraph/bullet changes, logical connections).


3. Output Format

When performing conversational rewrites, present output in the following format. Do not use emojis or unnecessary parentheses.

### Rewritten Body

(Body text reconstructed into natural Japanese)

---

### Changes Made
- (List up to 5 points covering changes prone to shifting meaning, connection fixes, or supplemented words. Write line-by-line "Before → After" with a short reason. If none, write "None" in a single line)
- Original expression → Rewritten expression (Reason for change)

### AI-like Parts Retained (*Include this section only if introductory or concluding phrases were retained because they carry meaning or weight. Omit this heading entirely if none)
- (Retained expression and a question asking the writer whether to trim it)

### Points to Confirm with Writer (*Include up to 2 points only if judgment was uncertain during writing or if connections between sentences could not be determined. Omit this heading entirely if none)
- (Briefly list 1–2 premises or conditions to verify)