I’ve researched how to remove the AI flavor, and found 7 different GitHub projects: Chinese rewriting, English rewriting, technical writing, and complete writing workflows — there are options for each.

1. qu-ai-wei (299 stars) — A Chinese de-AI skill that handles clichés, abstract expressions, sentence rhythm, and register, without making up facts during revision.
   https://github.com/LifelongLazyLearner/qu-ai-wei
2. 说人话 (818 stars) — Great for writing READMEs, development updates, and tech articles. It preserves version numbers, commands, and attribution, making the text more natural without losing key information.
   https://github.com/MrGeDiao/shuorenhua
3. De-AI Prompt Enhancer (566 stars) — Uses 7 real articles as style references; you can swap in your own old articles to teach AI your writing style, then clean up residual machine traces.
   https://github.com/OUBIGFA/De-AI-Prompt-Enhancer-Writer-Booster-SKILL
4. Humanizer (31k stars) — A basic English de-AI tool that specifically cleans up fixed openings, empty emphasis, and templated endings. Simple rules, easy to integrate into various agents.
   https://github.com/blader/humanizer
5. Stop-Slop (14k stars) — Suited for already-finished English content, targeting unnecessary preambles, business jargon, structural clichés, and pretentious choppy sentences.
   https://github.com/hardikpandya/stop-slop
6. no-ai-slop (2696 stars) — Works well when the original draft already has a personal tone. It can either just detect or make minimal edits, trying not to erase the author’s voice.
   https://github.com/petergyang/no-ai-slop
7. Writing Agent (335 stars) — If you want to go from topic selection and evidence gathering all the way to proofreading and export, this complete writing workflow is ready to use; de‑AI is just one part.
   https://github.com/dongbeixiaohuo/writing-agent

What’s your most-used de-AI skill? If you have better ones, let me know in the comments.

LifelongLazyLearner/qu-ai-wei

Source: https://github.com/LifelongLazyLearner/qu-ai-wei

去 AI 味(qu-ai-wei)

Version (https://github.com/LifelongLazyLearner/qu-ai-wei/releases)
License: MIT
Language
GitHub stars (https://github.com/LifelongLazyLearner/qu-ai-wei/stargazers)

Language: 简体中文 | English | 日本語 | 한국어 | Español

⚠️ 0.x development version (current v0.8.2): qu-ai-wei is still iterating; rules, categories, and APIs may change. Issues (https://github.com/LifelongLazyLearner/qu-ai-wei/issues) / discussions (https://github.com/LifelongLazyLearner/qu-ai-wei/discussions) / PRs are welcome.

Only Simplified Chinese is supported. Traditional Chinese has its own AI-tone features, character preferences, and typographic rules — these need separate maintenance and will be left for later versions.

Remove the traces of AI writing from Simplified Chinese, making the draft read like it was written by a human. Currently natively supports 8 agents: Cursor, Claude Code, OpenAI Codex CLI, OpenCode, Kiro, Factory Droid, Slate, Hermes.

https://github.com/user-attachments/assets/24513c20-968d-437b-8ceb-1ac1f77f6ad6

⚠️ What this skill can and cannot do

This is a floor cleaner + polisher, not a carving knife.

Current v0.8.2 behavior is still “floor cleaner + polisher,” but all conflicts are uniformly handled by the six-level “Conflict Arbitration Order” from SKILL.md:

• 1 Floor Cleaner — Removes obvious AI pollution (e.g., “赋能 / 助力 / 让我们一起 / 璀璨画卷 / 🚀 emoji bullet points / **bold mechanical ** / 希望对您有帮助”), making the draft go from “one look and you know it’s AI” to “doesn’t look like AI anymore.”
• 2 Polisher (added in v0.6.0) — Proactively polishes without inventing facts: verb strengthening (e.g., 进行讨论 → 讨论), rhythm reshaping (breaking machine-gun rhythm), filler removal (e.g., 一些 / 实际上 / 在一定程度上), abstract-to-concrete (only if details exist in the original), word order normalization, register matching. The bar moves from “doesn’t look like AI” to “clean and precise.”

But still not a carving knife. Carving requires viewpoint, context, and voice. If these aren’t in the original, the polisher can’t invent them — because §1 of the Conflict Arbitration Order is don’t invent facts, which takes priority over §6 polishing elevation. So:

• Original mediocre opinion + polishing → clean mediocrity (better than original, still not amazing)
• Original clear opinion + polishing → a decent draft (normal expectation for this skill)
• Original unique opinion + concrete details + polishing → close to readable level

Turning an ordinary draft into a truly sharp, voice-driven piece — that’s beyond its capability.

Removing AI tone from a draft without opinion and then polishing it results in a clean but still opinion-less draft. The opinion has to come from you.

This skill is suitable for

• ✅ Writers who want to make an AI-generated first draft (ChatGPT/DeepSeek/Claude/Qwen first version) sound less AI-like
• ✅ Everyday scenarios: work emails, PRDs, reports, tech blogs, product copy, public accounts (WeChat)
• ✅ Editors and operators who need to batch-clean obvious AI tone from drafts
• ✅ Students, product managers, operators, programmers, entrepreneurs — anyone who wants their daily writing to be less AI-ish
• ✅ Scenarios like Chinese teaching, AI writing detection research, text comparison analysis

This skill is not suitable for

• ❌ People who want to replace thinking with a tool — if the problem is “no unique observations / mediocre opinion / plastic details,” qu-ai-wei can’t fix it
• ❌ Serious in-depth writing (feature articles at the level of Sanlian Life Weekly / Renwu / GQ Reports ) — this skill can help with basic text cleanliness, but real writing quality depends on human observation, interviews, and thought, not a floor cleaner
• ❌ Academic paper writing — this skill may misjudge necessary academic expressions like 进行 + V, 然而 / 此外, 性 / 化 suffixes (though there is register recognition protection, risk remains)
• ❌ Legal documents, official documents — officialese is a compliance requirement, not AI tone; this skill should leave such texts almost untouched
• ❌ Writing that needs to preserve narrative buildup and control textual rhythm — this skill may mistake “pacing” (environmental description, pauses before character voices, digressive background paragraphs) as “redundancy” and delete it; it also cannot manage the overall momentum and flow (strong start weak finish, losing breath at key points, argument stalling). For such writing, qu-ai-wei can at most run one pass for basic cleaning; structure and rhythm must be controlled by the author.
• ❌ Using it to “launder AI-generated content to avoid detection” — please don’t use this skill to bypass school/journal/company AI detection policies. qu-ai-wei is not a cheating tool; it’s a tool to make your own real writing cleaner.

One-sentence positioning (v0.8.2)

Feed it a draft you’ve written, and it outputs a clean, sharp final version. It’s not for pretending you wrote something you didn’t, nor for turning any draft into a master-quality article.

Inspiration for this skill came from humanizer (https://github.com/blader/humanizer) (author Siqi Chen, MIT license, 2025). humanizer’s official positioning is language-agnostic, but its original set of rules targets English — Title Case, em dash, hyphenated pairs are all English-specific.

The Chinese version borrowed the skeleton from humanizer: three-pass workflow, voice calibration approach, the “Personality & Soul” chapter, and about half of the AI-tone patterns (Chinese adaptation of humanizer’s equivalent rules). The other half was reorganized in the Chinese context: a pre-step for register recognition, Chinese grammar basis, full-width punctuation, de-Europeanized word order, and a three-question check to avoid over-correction. Patterns #39–#42 have another source, from original articles on yage.ai.

Installation

Supported agents (same installation method)

All 8 agents above can be installed using this repository’s script.

Installation via external skills CLI (recommended)

If you already have Node/npm, you can install directly from GitHub using the external skills CLI. By default, skills CLI auto-detects available agents on your machine; if none are detected, it will prompt you to select an installation target.

npx skills add https://github.com/LifelongLazyLearner/qu-ai-wei

This is not qu-ai-wei publishing its own npm package; the command uses skills CLI to pull this GitHub repository and install it.

To explicitly install to a specific agent, use -a:

# Install only to Codex
npx skills add https://github.com/LifelongLazyLearner/qu-ai-wei -a codex

# Install to multiple agents simultaneously
npx skills add https://github.com/LifelongLazyLearner/qu-ai-wei -a codex -a claude-code -a cursor

# Install to global directory and skip confirmation prompts
npx skills add https://github.com/LifelongLazyLearner/qu-ai-wei -g -a codex -y

Installation via repository script (full control path)

If you need to install all at once, use symlink/copy mode, set a custom directory, or repeat with overwrites, continue using the repository script:

git clone https://github.com/LifelongLazyLearner/qu-ai-wei.git ~/qu-ai-wei
cd ~/qu-ai-wei

# Install per platform (repeatable)
bash scripts/install-skill.sh --platform cursor
bash scripts/install-skill.sh --platform claude
bash scripts/install-skill.sh --platform codex
bash scripts/install-skill.sh --platform opencode
bash scripts/install-skill.sh --platform kiro
bash scripts/install-skill.sh --platform factory
bash scripts/install-skill.sh --platform slate
bash scripts/install-skill.sh --platform hermes

# Install all at once
bash scripts/install-skill.sh --platform all

Optional parameters:

• --name <name>: custom installation directory name (default qu-ai-wei)
• --mode copy: copy instead of symlink (default symlink)
• --to <path>: install to an arbitrary custom skill directory (repeatable)
• --host: alias for --platform (compatible with old habits)

Example:

# Custom directory name
bash scripts/install-skill.sh --platform codex --name qu-ai-wei-cn

# Custom platform directory
bash scripts/install-skill.sh --to ~/.my-agent/skills --name qu-ai-wei

Other models supporting custom instructions (ChatGPT / DeepSeek / Kimi / Tongyi etc.)

Paste the body of SKILL.md directly into the model’s custom instructions or system prompt, skipping the YAML frontmatter section wrapped by --- at the top.

Upgrade

After installation, to get the latest rules:

# Update repo (recommended)
cd ~/qu-ai-wei && git pull

# If you are using symlink mode (default), you're done.
# If you are using copy mode, reinstall to overwrite:
bash scripts/install-skill.sh --platform all --mode copy --force

You can also run directly from an installed directory (example):

bash ~/.claude/skills/qu-ai-wei/update.sh
# or
bash ~/.codex/skills/qu-ai-wei/update.sh

update.sh will print old version → new version and a link to the release page for this version; if there’s no new version, it will say “already up to date.” For specific changes in each version, see Releases (https://github.com/LifelongLazyLearner/qu-ai-wei/releases).

Usage

Natural language trigger (recommended)

Simply ask the model to process text in plain language, and qu-ai-wei will trigger automatically:

帮我去 AI 味:[paste Chinese text]
改得说人话:[paste Chinese text]
这段中文太 AI 了,润色一下:[paste Chinese text]
让它更像人写的:[paste Chinese text]
humanize 这段中文:[paste Chinese text]

Explicit invocation (slash command)

/qu-ai-wei [paste the Chinese text you want to rewrite]

Character range

Only processes Simplified Chinese. If Traditional Chinese is input, it will prompt “The current version only supports Simplified Chinese. It is recommended to use OpenCC or similar tools to convert to Simplified first” and will not perform automatic conversion.

Voice calibration (optional)

To make the rewrite more closely match your own writing style, you can paste a sample of your own Chinese writing as a reference:

/qu-ai-wei Here is my own writing sample for style reference:
[paste 2–3 paragraphs of your own writing]
Now please rewrite this text:
[paste AI-written Chinese]

The skill will first extract a 5-item lightweight list from the sample (high-frequency words, average sentence length, sentence-initial words, punctuation preference, overall register), ask you to confirm, and then rewrite. Compared to a full style analysis, this approach is more honest: 2–3 paragraphs of Chinese can only provide limited signals.

Design Philosophy

Relationship with humanizer

The skeleton and methodology are borrowed from humanizer: three-pass workflow, voice calibration, the “Personality & Soul” chapter, and roughly half of the pattern concepts. The rest was reorganized in the Chinese context.

Core capability: Register recognition

If all text were rewritten toward “everyday spoken Chinese,” academic papers’ “进行深入分析” would become “好好分析一下” and official documents’ “依法予以处理” would become “按法律办.” Therefore, qu-ai-wei first performs register recognition and then selects rules:

The complete rule activation matrix is in SKILL.md. When uncertain, qu-ai-wei will ask: “Is this text from Moments / WeChat official account / work email / feature article / brand website / academic report / contract / exam essay?”

Core detection principle: Density matters, not occurrence

Antithesis, parallelism, four-character idioms — human writers use them all the time. Just using them doesn’t mean AI tone. Repeated piling up, unrelated to surrounding content, and applicable to any other topic — that’s the sign. A single occurrence doesn’t count; repeated occurrence within a short paragraph (under 200 characters) counts.

Avoiding “over-correction”

These 51 pattern categories are detection heuristics, not rewriting templates. De‑AI ≠ remove all normalized expressions. Common, widely accepted Chinese writing patterns (e.g., 支持 X 等平台, 用于 X, 基于 X) are not AI tone in themselves. Before rewriting, ask three questions:

1. Naturalness check: Is this expression common in real human Chinese writing? Uncommon + I just invented it → dangerous.
2. Function check: Does it serve a function in the original text, such as topic introduction, term fidelity, rhythm pause? If so, don’t force-delete it just for “more colloquialism.”
3. Register check: Does the rewritten text match the original document’s register? Turning academic rigor into everyday chat, or official document solemnity into Moments banter — both are register downgrade failures.

About positive references

The skill lists Sanlian Life Weekly, Reader, Chinese National Geography, WeChat UX, and Wang Zengqi, Lu Xun, A Cheng as “positive references.” v0.4 added Apple Greater China / Nike Greater China as a reverse checkpoint for brand advertising register (“Would this text look jarring at Apple’s hero copy or Nike ad?”). Usage note: They are diagnostic negative references (“Would this sentence feel awkward in Sanlian Life Weekly?”) not style imitation targets (“Write in Sanlian style” / “Write an Apple-style slogan”). Just giving an LLM a name to learn a style usually results in stereotyped imitation. One-click style switching is not planned for now.

Chinese-English mixing: Keep only proper nouns, ignore spaces

The only hard rule: English proper nouns, product names, person names, abbreviations — keep them as-is during rewriting. Don’t translate, paraphrase, or replace. “Codex CLI” cannot become “科德克斯命令行”; “Transformer” cannot become “变换器.” This is an extension of “information integrity.”

Half-width spaces between Chinese and English/numerals (so-called “Pangu’s White”) — This skill doesn’t manage them. AI and humans both vary in this regard; adding or not adding spaces is indistinguishable and cannot serve as a signal for AI tone removal. It’s also not a national standard — GB/T 15834-2011 only covers punctuation; W3C CLREQ says “character spacing or whitespace” is optional and not mandatory for U+0020; even the original author of Pangu’s White (sparanoid) calls it an “informal typographic convention.” If we added spaces for users, we’d fall into a pit: the original text might not have spaces (common in Moments, WeChat official accounts, Xiaohongshu), and adding them would introduce a “tech circle tone.”

Rule: If the original has spaces, keep them; if not, don’t add them. If you need uniform formatting or pursue Chinese typographic aesthetics, use pangu.js (https://github.com/vinta/pangu.js) for automatic processing, or refer to sparanoid’s Chinese Copywriting Guidelines (https://github.com/sparanoid/chinese-copywriting-guidelines/blob/master/README.zh-Hans.md) — that’s a typographic standard, with a different goal from this skill’s “AI trace detection” (typography manages aesthetics; we manage mechanical feel).

Punctuation: Two levels of norms

First level (full-width vs. half-width): In a Chinese context, punctuation should always be full-width, not mixed with half-width. English training data uses half-width punctuation, and models often directly reuse that when outputting Chinese, creating a visual “translationese” signal.

Second level (functional division): Full-width vs. half-width solves “the right character”; the next level is “what each punctuation does.” SKILL.md covers functional division and common AI mismatches for punctuation like enumeration commas, semicolons, book title marks, ellipses, colons, quotation marks, and brackets (e.g., replacing colons that introduce information with dashes, using quotation marks instead of book title marks, abnormally frequent semicolons and ellipses in serious text). Semicolons in daily/business register have a baseline close to 0, but in editorials/commentary/long arguments they are a legitimate parallel rhythm tool — they don’t trigger AI-tone judgment.

Comparison tables, exception lists, typical errors, and baseline punctuation frequencies from serious media are all written in the two “Punctuation” sections of SKILL.md.

Word order: Fully Sinicized, avoid English-style inversion

LLMs often apply English syntax directly to Chinese. Individual sentences may read like Chinese, but when read together they feel off — likely due to English residue in word order. Basic Chinese habits: subject first, short attributives, adverbials before verbs, conditional and reason clauses before the main clause. Common ailments (overlong prepositional phrases at the beginning, starting with “作为一个 X, …”, stacked attributives, direct translation of English relative clauses) and a four-question self-check are written in the “Word Order” section of SKILL.md.

Information integrity

Rewrite only the expression; do not delete information. Every concrete fact in the original (numbers, names, times, product names, key phrasing) should have a corresponding presence in the final draft. If an expression feels redundant, you can compress it, but you cannot quietly drop facts. When in doubt, retaining the original information is more important than brevity.

51 Pattern Categories at a Glance (abbreviated)

For execution order, register matrix, and rule index, see SKILL.md; core rule details are in references/patterns.md, and platform/style rules for Group H are in references/platform-patterns.md. This README only keeps a directory-level overview for quick judgment of whether your scenario is covered.

Common entry points:

• Execution order and rule index: SKILL.md
• Rule trigger details: references/patterns.md + references/platform-patterns.md
• Full process: SKILL.md Processing Flow
• Complete examples: SKILL.md Complete Examples

Complete Example

Original text (heavy AI tone):

随着协同办公模式的不断发展,远程周报机制在团队管理中发挥着越来越重要的作用。通过标准化模板,我们能够从多维度、全方位地掌握项目进展,提升协作效率,赋能组织管理。

值得一提的是,一个高质量周报通常包括以下几个方面:

1. 本周完成情况:系统梳理关键成果。
2. 风险与挑战:识别潜在问题并提出解决思路。
3. 下周计划:明确重点任务,确保高效推进。

总体来看,周报不仅是信息同步工具,更是推动团队持续成长的重要抓手。让我们共同打造更加透明、高效、可持续的协作机制。

Example output (following current workflow):

[Gate Check] Judgment: AI-generated text | Evidence: [“随着...不断发展”, “值得一提的是”]

First draft

远程周报有用,但前提是别写成模板作文。

真正有价值的周报通常就三件事:这周做成了什么、卡在哪里、下周准备怎么推进。写清楚这三件,团队同步就够用了。

周报不是“看起来很完整”的文档,更像一个对齐工具:让大家知道项目现在到哪一步、问题在哪、谁在跟进。

Self-review

• AI afraid to make assertions: Present. “周报有用,但前提是别写成模板作文。” This contains a judgment, not customer service tone.
• Full polishing self-check five questions:
  • Precise wording: Pass (removed empty words like 多维度、全方位、赋能、抓手).
  • Rhythm variation: Pass (long sentences broken into short; paragraph length varies).
  • Verb-driven: Pass (condensed 梳理/识别/提出 into 做成/卡住/推进).
  • Concrete replaces abstract: Pass (turned 持续成长机制 into 项目到哪一步、谁跟进).
  • Consistent voice: Pass (unified as written/general, neutral direct narration).
  • Residual traces: Slight slogan feel mostly eliminated.

Final draft

远程周报有用,但前提是别写成模板作文。

一份能用的周报,核心就三件事:这周做成了什么、卡在哪里、下周怎么推进。把这三件写清楚,团队同步效率通常不会差。

周报本质上是对齐工具,不是展示文笔的材料。它要回答的是:项目到哪一步了,当前风险是什么,下一步由谁跟进。

Polishing report

• Verb strengthening: “发挥...作用” → “有用”, “推动团队持续成长” → “回答...问题”.
• Rhythm reshaping: First paragraph changed from one long sentence to two short sentences, reducing press release tone.
• Filler removal: Deleted “不断发展 / 多维度 / 全方位 / 赋能 / 重要抓手 / 共同打造”.
• Abstract to concrete (using only original info): Turned “高质量周报包括三个方面” into “做成了什么、卡在哪里、下周怎么推进”.
• Word order normalization: Changed “通过标准化模板,我们能够...” to “核心就三件事...写清楚即可”.
• Register matching: Maintain “written/general” register, not downgraded to chatty tone.

Change summary (optional)

• Main hits: #1/#2 (hollow elevation and background opening clichés), #7 (high-frequency business words), #30/#31 (redundancy and abstract empty words), #37 (templated structure).

For more complete examples, see SKILL.md.

Reminders for Contributors / Self-Maintenance

For the review and change record against agentskills.io (https://agentskills.io) skill creation best practices, see docs/agentskills-review.md (v0.8.1).

1. After modifying SKILL.md / references/ / README, feed the prose paragraphs you changed back to qu-ai-wei for self-check. The structured metadata in rule entries (problem / keywords / original / revised / register restrictions) is the skeleton for the model to read, don’t humanize it — removing symmetry would make it harder for the model to understand the rule’s structure. Only run self-check on explanatory prose.
2. When adding new rules, check for conflicts/overlaps with existing rules. For example, “abstract universal verbs” are adjacent to #7 AI high-frequency words and #30 redundant written forms; clarify what each governs. Core A-G/I rules go into references/patterns.md; platform/style H group rules go into references/platform-patterns.md. Also add a quick-reference line in the “51 AI-Tone Patterns · Index Table” at the end of SKILL.md.
3. All examples must have “original / revised” pairs. Providing only judgment criteria without demonstration makes it hard for the model to grasp.
4. Whenever the version number is bumped, update the version record in README, CHANGELOG.md, and SKILL.md frontmatter together. After changes, run bash tests/check-version-sync.sh to align README / flat build / changelog at once.
5. After modifying SKILL.md or references/, run scripts/build-flat.sh before committing — it will flatten SKILL.md + references/ to generate .cursorrules and WARP.md automatically. Cursor / Windsurf / Warp don’t support progressive disclosure and must use a single file. Drift between the three files is the most common bug; the script is the single source of truth (place this line after rule 4, not after 6, because its refresh frequency is higher than 咬文嚼字). If you forget to run it, it’s okay — tests/check-flat-sync.sh will regenerate a comparison and fail on drift.
6. When encountering classic human texts (Jin Yong / Wang Shuo / Wang Zengqi / real interview transcripts etc.) don’t modify them. This is the most common disaster for qu-ai-wei, and it’s explicitly written in the “🛑 Step Negative One” gate check.
7. After changes, run tests/check-snapshot-smoke.sh for regression smoke testing. tests/fixtures/ covers long-term samples, negative contrast pairs, real-person stop scenarios, brand advertising, academic/tech, old Xiaohongshu templates, skincare ingredient whitelists, etc.; before-and-after comparisons and targeted outputs are in tests/after/, and 01-03 can be compared with tests/baseline/ to see if structural changes introduced behavioral drift. Any major change (new rules, rule merges, register matrix adjustments) should be preceded by a full smoke test.
8. Use manifest for real model output verification. First run bash tests/check-runs.sh tests/after to verify existing anchor outputs; future captured runs should go into tests/runs/<date>-<description>/, write metadata according to tests/runs/README.md, then run the same script.
9. Each December, after 咬文嚼字 releases the top ten buzzwords, and the following January the top ten language errors, refresh the “fresh words” / “dead words” lists in rule #49. Add new words to the “official list” line; demote old words that have been off the list for 3+ years and are no longer used by the community to the “dead” line. The update modifies the #49 section in references/platform-patterns.md + the “Language Timeliness Anchor” subsection in references/sources.md, then run scripts/build-flat.sh to sync to .cursorrules / WARP.md, and also update the corresponding sections in the README. The cost of