# Japanese Naturalization > Naturalize the current or explicitly referenced assistant output into context-appropriate Japanese while preserving meaning. Use when the user asks to rewrite an existing assistant response so it sounds natural, removes AI-like Japanese or translationese, or fits a different register. For standalone evidence-backed term selection, use term-translation-research. - Skill: `japboy/japanese-naturalization` (Agent Skill, multi-file: 11 files) - Install (CLI): `npx skillmds@latest add japboy/japanese-naturalization` - Raw SKILL.md: https://api.skillmd.com/api/skills/japboy/japanese-naturalization/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - Author: japboy (https://skillmd.com/u/japboy) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/japboy/japanese-naturalization --- # Japanese Naturalization ## Purpose Use this skill to naturalize assistant-generated Japanese so it sounds like something a native speaker in the inferred target context would plausibly write. This is a response-naturalization and localization skill, not a fact-checking, summarizing, or copywriting skill. Preserve the target output's meaning, factual claims, intent, structure, numbers, citations, links, code, and named entities unless the user explicitly asks to change them. ## Core Principle Define "natural Japanese" as context fit: the wording should match the likely audience, domain, channel, register, and purpose. Do not equate naturalness with plain language, pure Japanese vocabulary, fewer katakana words, or more kanji. ## Invocation Model When invoked: 1. Treat the current user message as an instruction for how to naturalize the target output, not as the target text itself. 2. Treat the immediately previous assistant response as the default target. 3. If the user explicitly references another assistant response or earlier generated draft in the conversation, use that referenced output as the target. 4. Do not assume that quoted, fenced, or pasted text in the current user message is the target text. Treat it as instruction or context unless the user explicitly says it is the assistant output to naturalize. 5. If the target context is explicit, use it. 6. If the target context is missing, infer it from the target output and current conversation. State assumptions only when they affect the output or the user asked for notes. ## Workflow 1. Infer the communication context. - Identify domain, audience, channel, register, purpose, source language, and terminology risk. - Classify the target as casual chat, technical explanation, product UI, business document, support text, academic prose, marketing copy, or another explicit finite category. 2. Preserve meaning before style. - Keep claims, uncertainty, scope, examples, ordering, code blocks, URLs, file paths, citations, numbers, and proper nouns intact. - Do not add new evidence, promises, caveats, or persuasive claims. - Do not silently remove nuance to make the sentence shorter. 3. Choose terminology by context. - Use [terminology-traps.md](references/terminology-traps.md) for context-sensitive term choices that can produce unnatural or misleading Japanese. - If a term choice could change meaning and the reference does not cover it, inspect available local project evidence first. - For substantial term research, use the `term-translation-research` skill when available. - Prefer official, primary, product-maintained, or domain-specific sources over dictionaries and search snippets when external evidence is needed. 4. Check audience-sensitive register risks. - Use [audience-register-risks.md](references/audience-register-risks.md) when casual expressions, jokes, slang, age/generation cues, hierarchy, identity or attribute references, or harassment sensitivity could affect how the wording lands. - Prefer widely understood, neutral wording when the audience is mixed, unknown, external, or workplace-broad. - Preserve intentionally community-specific wording only when local evidence or user instruction shows it fits the target context. 5. Check notation variation and consistency. - Use [notation-variation-risks.md](references/notation-variation-risks.md) when kanji/kana choices, okurigana, same-reading kanji, katakana/English variants, punctuation, number style, or full-width and half-width forms vary within the target text. - Preserve intentional variation when it carries meaning, register, readability, quoted/source form, or local product terminology. - Normalize confusing variation when it makes one concept look like multiple concepts, hides a meaning distinction, or breaks document-level predictability. 6. Apply Japanese style defaults. - Use [japanese-style-defaults.md](references/japanese-style-defaults.md) for orthography and lightweight style defaults. - Treat these as defaults, not rules. Project style guides and target publication norms override them. 7. Remove translationese and AI-like stiffness. - Use [ai-like-japanese-patterns.md](references/ai-like-japanese-patterns.md) as warning signs, not banned words. - Replace only when the alternative is more natural in the inferred context. - Keep technical katakana, kanji compounds, and passive constructions when they are normal for the domain. 8. Control register explicitly. - Match the user's requested tone if present. - Otherwise, preserve the source's communicative intent while making the Japanese less stiff, less literal, and less templated. - Avoid making a serious document casual or a casual answer overly formal. ## Output Contract Default output: - Return only the rewritten Japanese. - Preserve Markdown structure when it carries meaning. - Keep code blocks, commands, tables, links, citations, and file paths stable. When the user asks for explanation or when a key ambiguity affects the result, use: ```markdown ## Rewritten Japanese ... ## Notes - Context: ... - Terminology: ... - Ambiguity: ... ``` ## Ambiguity Handling - If one conservative wording preserves meaning and sounds natural, use it. - If two plausible term choices imply different concepts, ask one concise question before rewriting. - If the user requested no questions, choose the safest wording and include a short note when notes are allowed. ## Quality Checklist Before answering, check that: - The rewritten text still says the same thing. - The terminology matches the inferred audience and domain. - The register fits the channel. - Audience-sensitive register and harassment risks were checked when relevant. - Natural notation variation and confusing notation inconsistency were separated when relevant. - AI-like patterns were inspected but not mechanically removed. - No new factual claim, citation, example, or recommendation was added. - The result sounds like a plausible Japanese speaker wrote it for that context. ## Evaluation Fixtures For examples and regression prompts, use [evaluation-prompts.csv](references/evaluation-prompts.csv). ## Reference Map For a complete map of runtime references, maintenance records, and source hierarchy, use [REFERENCE.md](references/REFERENCE.md). ## Skill Improvement Feedback Apply this section only when the user explicitly asks to update, maintain, or incorporate reusable feedback into this skill. A correction to the current rewrite alone is not permission to propose or edit skill files; fix the rewrite and stop. In an explicit skill-maintenance task, use [improvement-workflow.md](references/improvement-workflow.md). Keep updates bounded, evidence-backed, and validated. If the update involves Agent Skills structure, metadata, validation, or SkillOpt-style review discipline, use the `agent-skill-authoring` skill when available.