# De AI Writing > Rewrite Chinese or English prose so it reads more naturally and less like AI-generated text while preserving facts, evidence, logic, genre, and author intent. Use when the user asks to 去 AI 腔, 去除 AI 味, humanize writing, make it less AI, less ChatGPT-like, less generic, more natural, more human, more authorial, less templated, or when Chinese or English prose feels over-structured, repetitive, connector-heavy, vague, over-polished, lexically flat, voice-less, translation-like, or machine-smoothed. Works across English manuscripts, academic writing, thesis prose, Chinese medical writing, translated drafts, emails, application letters, reports, PRDs, README files, product copy, presentations, newsletters, and other Chinese or English writing tasks. - Skill: `qiyanghong2020/de-ai-writing` (Agent Skill, multi-file: 6 files) - Install (CLI): `npx skillmds@latest add qiyanghong2020/de-ai-writing` - Raw SKILL.md: https://api.skillmd.com/api/skills/qiyanghong2020/de-ai-writing/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - License: MIT - Author: qiyanghong2020 (https://skillmd.com/u/qiyanghong2020) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/qiyanghong2020/de-ai-writing --- # De-AI Writing Use this skill to turn stiff, templated, or machine-smoothed Chinese or English into natural human prose without flattening the substance, faking personality, or chasing detector scores. ## Choose the language lane Load only the references that fit the task: - Read `references/chinese.md` for Chinese drafts or mixed drafts where the main prose is Chinese. - Read `references/english.md` for English drafts or mixed drafts where the main prose is English, especially academic manuscripts, section-level manuscript editing, translation-like English, sentence polishing, emails, application letters, narrative prose, or requests that mention tools such as Paperpal, DeepL Write, Grammarly, Wordtune, Hemingway Editor, ProWritingAid, QuillBot, or "AI humanizer". - Read `references/execution-patterns.md` for long documents, layout-sensitive documents, staged rewrites, files with automatic fields, or full-manuscript structural audits. - Read `references/thesis-medical.md` for Chinese theses, dissertations, medical papers, academic discussion sections, or formal Chinese scholarly prose. For Chinese work, use `references/chinese.md` first; it points to `references/markers.md`, `references/rewrite_patterns.md`, and `references/risk_words_quicklist.md` when concrete examples or quick phrase checks are needed. For English work, use `references/english.md` first; it contains the English-specific AI-tone markers, phrase risks, and genre lanes. For long English academic papers, Viewpoints, framework papers, or conceptual manuscripts, also read `references/execution-patterns.md`. If the user has authorized whole-manuscript revision, run the document-level structural audit before local rewriting; if the request is local, do not expand it into a manuscript-wide restructure. For English biomedical manuscripts or English tables in medical papers, keep the work in the English lane and use `references/english.md`, especially the scientific/medical table-language guidance. Use `references/thesis-medical.md` only when the main prose is Chinese medical or formal Chinese scholarly writing. ## Core workflow ### 1. Lock the meaning Before rewriting, preserve: - factual claims - source role or document purpose - sample size, dates, numbers, and named entities - effect direction and comparison structure - evidence boundary, limitation, uncertainty, or scope - citation relationship and attribution Do not add examples, anecdotes, data, citations, author emotions, dialect markers, or first-person claims unless they are already present or the user explicitly asks for them. ### 2. Diagnose before rewriting Make a quick local diagnosis: - Which language lane applies? - What genre is this: academic, public article, report, product doc, email, speech, personal essay, or README? - Which edit lens applies: academic manuscript, naturalized translation, tone and clarity, expression variants, brevity, long-form style, or humanizer-caution? - Is the AI feel mostly lexical, structural, tonal, or caused by missing specificity? - Does the user provide a writing sample that should calibrate rhythm and stance? - Are there protected zones: references, formulas, tables, citations, figure titles, legal wording, automatic fields, code, or UI strings? ### 3. Rewrite the whole local block AI tone usually appears in clusters. If the user points to one awkward sentence, inspect and rewrite the local block, not only the quoted sentence. At minimum, check: - the quoted sentence - the paragraph it sits in - the paragraph before and after when available - any heading, caption, summary box, or bullet list that carries the same tone For long or structure-sensitive documents, use paragraph or local block as the default unit. Skip short structural labels unless the user explicitly wants them rewritten. ### 4. Calibrate to the writer If the user provides a sample of their own writing, extract a compact style memory: - sentence length tendency: short, medium, or long - paragraph opening habit: direct entry, contextual setup, or mixed - punctuation habit - stance level: flat explanatory, mildly assertive, personal, formal, or visibly authorial - register: academic, business, technical, public-facing, conversational, or local English variety Apply rhythm and stance without copying verbal tics mechanically. Preserve a writer's legitimate Chinese style, non-native English style, or World English variety when it carries identity or pragmatic meaning. ### 5. Remove machine polish first Prioritize fixes that restore specific human judgment: - replace vague abstractions with concrete subjects, verbs, conditions, and outcomes - reduce stock transitions when the relation is already clear - break repeated sentence skeletons and paragraph openings - vary sentence pressure according to meaning, not for cosmetic irregularity - cut meta-commentary about the writing itself - remove teacher-like reader management and over-helpful scaffolding - replace generic enthusiasm with evidence, constraint, consequence, or choice - preserve field-typical terms when they are precise Do not make the prose longer just to seem more human. Do not introduce mistakes, slang, or awkwardness as camouflage. If the request is framed around a direct "AI humanizer" or detector bypass tool, convert it into a meaning-preserving genre edit. Avoid synonym churn, random imperfection, and any promise that the output will evade Turnitin, GPTZero, or another detector. Do not infer authorship or the extent of AI use from prose style. If disclosure accuracy matters, distinguish language editing, substantive rewriting, framework design, and conceptual synthesis using the actual workflow and the target journal's policy. Without that evidence, state uncertainty rather than alleging undisclosed or improper use. ### 6. Match the genre Academic and thesis writing: - keep evidence-reporting verbs and field-typical nouns - distinguish exploratory findings from confirmed conclusions - avoid adding first person unless the genre or user allows it - keep citations, methods, results, and limitations conservative Reports, PRDs, READMEs, and technical docs: - name the owner, object, action, constraint, and result - prefer measurable or observable wording over strategy fog - keep commands, API names, UI labels, and code terms exact Public articles, newsletters, and explainers: - speak directly to the topic, not to the act of reading - keep the reader relationship flat and respectful - add warmth only where the genre already allows it Emails and workplace messages: - make the ask, context, deadline, and next step clear - avoid corporate padding and over-politeness - keep tone appropriate to the relationship ### 7. Final human pass Before output, check: - meaning and evidence boundaries are unchanged - no invented facts, citations, examples, or personal details - no detector-evasion claim or promise - the rewrite sounds like a plausible person in that genre - the paragraph no longer depends on stock phrasing to feel organized - any retained high-risk words are there because they are precise - for an authorized whole-manuscript academic edit, cross-section restatements add evidence, a boundary condition, an operational consequence, or a genuine inference - retained taxonomies and conceptual labels perform distinct analytical work rather than renaming the same distinction - manuscript length remains proportionate to the evidence and conceptual contribution - necessary definitions, Methods and Results detail, key limitations, and Abstract-Introduction-Conclusion correspondence were not removed as false redundancy When returning a rewrite, output the rewritten content directly unless the user asks for explanation, diagnostics, or before/after notes.