Home / All Skills / Content & Media Creation / humanizer-zh
English (EN)简体中文 (ZH)日本語 (JA)
Content & Media Creation · op7418/Humanizer-zh

humanizer-zh

Removes traces of AI-generated text. Ideal for editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's comprehensive guide on 'Signs of AI writing'. Detects and fixes patterns such as exaggerated symbolism, promotional language, superficial analysis ending in -ing, vague attributions, dash overuse, the rule of three, AI vocabulary, negative parallelism, and excessive conjunctive phrases.

L2 Local · Restricted workspaceV1 Commit-PinnedWrite FileDual-pass Audit
Author op7418GitHub op7418/Humanizer-zh ↗Stars 17300License MIT（仓库根 LICENSE，Copyright (c) 2026 歸藏；GitHub API spdx=MIT）Commit 91f3d394db
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Pure prompt rulebook: The skill directory (SKILL.md + README.md + LICENSE + .gitignore) contains no scripts or executable code; all capabilities derive from the checklist of 24 'AI writing' patterns and their rewriting benchmarks in the text.

humanizer-zh/SKILL.md
你是一位文字编辑，专门识别和去除 AI 生成文本的痕迹，使文字听起来更自然、更有人味。本指南基于维基百科的"AI 写作特征"页面，由 WikiProject AI Cleanup 维护。
What it does: Synthesizes observations from Wikipedia's Signs of AI writing into an actionable checklist for the model. The host only needs to load SKILL.md into the context to take effect, requiring no runtime dependencies.

Each pattern adopts a fixed four-part structure: 'Words to watch out for → Problem description → Before rewriting → After rewriting'. The model scans item by item and rewrites in place, rather than following a vague prompt to 'write more naturally'.

humanizer-zh/SKILL.md
### 1. 过度强调意义、遗产和更广泛的趋势 **需要注意的词汇：** 作为/充当、标志着、见证了、是……的体现/证明/提醒、极其重要的/重要的/至关重要的/核心的/关键性的作用/时刻、凸显/强调/彰显了其重要性/意义、反映了更广泛的、象征着其持续的/永恒的/持久的、为……做出贡献、为……奠定基础、标志着/塑造着、代表/标志着一个转变、关键转折点、不断演变的格局、焦点、不可磨灭的印记、深深植根于 **问题：** LLM 写作通过添加关于任意方面如何代表或促进更广泛主题的陈述来夸大重要性。
Pattern 1 also provides a Chinese trigger word list (e.g., 作为/充当, 标志着, 见证了…… 深深植根于), turning subjective 'AI flavor' judgments into an auditable wordlist.

Compresses the entire methodology into 5 core rules at the beginning (cut filler, break formulas, vary rhythm, trust the reader, drop punchlines), with the 24 patterns serving as expansions of these 5.

humanizer-zh/SKILL.md
1. **删除填充短语** - 去除开场白和强调性拐杖词 2. **打破公式结构** - 避免二元对比、戏剧性分段、修辞性设置 3. **变化节奏** - 混合句子长度。两项优于三项。段落结尾要多样化 4. **信任读者** - 直接陈述事实，跳过软化、辩解和手把手引导 5. **删除金句** - 如果听起来像可引用的语句，重写它
The 5-point quick reference serves as an executive summary for the model, preventing long texts from being treated as 'reference books' instead of 'actionable rules'.

Explicitly requires injecting tone in addition to 'de-AI-ing', acknowledging that sterile, opinionless clean text is equally obvious as fake.

humanizer-zh/SKILL.md
避免 AI 模式只是工作的一半。无菌、没有声音的写作和机器生成的内容一样明显。好的写作背后有一个真实的人。
The direct consequence is that the model cannot merely perform subtractive editing; it must also inject viewpoints, complex feelings, and allow a bit of messiness based on provided examples.

Includes quantitative quality gates before delivery: a 50-point rubric across 5 dimensions × 10 points and a 6-item quick checklist, turning 'does it sound human' into a self-evaluable score.

humanizer-zh/SKILL.md
- 45-50 分：优秀，已去除 AI 痕迹 - 35-44 分：良好，仍有改进空间 - 低于 35 分：需要重新修订
Scoring dimensions are Directness / Rhythm / Trust / Authenticity / Concision (10 points each); paired with a 5-step 'Processing Flow' (Read input → Identify patterns → Rewrite → Cross-check → Present) to form a reproducible pipeline.

The tool surface is narrowed to four tools in the frontmatter, with no shell, network, or scraping commands in the body.

humanizer-zh/SKILL.md
allowed-tools: - Read - Write - Edit - AskUserQuestion
Only reads and rewrites the user's own text; makes no external requests and fetches no external corpus. This accounts for the skill's minimal risk surface.

Lineage and sources are documented in the body and README: the Chinese version is translated from blader/humanizer, utility sections reference hardikpandya/stop-slop, and the theoretical foundation comes from Wikipedia's WikiProject AI Cleanup.

humanizer-zh/README.md
- 本项目的核心文件翻译自 [blader/humanizer](https://github.com/blader/humanizer/tree/main) > - 实用工具部分（核心规则、快速检查清单、质量评分）参考了 [hardikpandya/stop-slop](https://github.com/hardikpandya/stop-slop) > - 原项目基于维基百科的 [Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) 指南
The README's top 'Disclaimer' section lists the three sources; the 'References' section at the end of SKILL.md provides links to the same Wikipedia page (for attribution only, not runtime scraping).

2Core Capabilities

01Systematic identification of 24 types of AI writing traces (Content 6 / Language & Grammar 6 / Style 6 / Communication & Filler 6)
02Actionable, item-by-item 'Before → After' rewriting examples (including a Chinese trigger wordlist)
03Human touch injection: viewpoints, rhythm variations, acknowledging complexity, first-person perspective, allowing messiness, and grounding feelings
04Editing workflow: Identify patterns → Rewrite problematic fragments → Preserve meaning → Maintain tone → Inject soul
05Quick checklist (repetitive sentence lengths, single-line wrap-ups, dashes, explaining metaphors, conjunctions, rule of three)
0650-point quality rubric and passing threshold for pre-delivery self-evaluation
07Works on whole files (e.g., article.md) as well as pasted text snippets

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Restrict working directory
For Administrators
Retain operation logs
Evidence & Triggers (File Paths & Excerpts)
  • 本地工作目录: File I/O and artifact output: SKILL.md frontmatter declares allowed-tools including Write / Edit; README demonstrates 'Please humanize article'
Audit Notes: The target of evaluation is this skill directory (SKILL.md + README.md + LICENSE + .gitignore): zero scripts, zero network endpoints, zero credential access. The sole basis for classifying it as Blue tier is its explicit declaration and use of local write tools—frontmatter allowed-tools contains Write and Edit, and the README directly demonstrates operations on article.md, falling under 'local read/write only, no external transmission'. If the host does not grant write tools, it falls back to pure conversational rewriting (Green tier). Notable stylistic side effects: injecting opinions, colloquialisms, or first-person voice to remove AI flavor may alter original tone; mechanical enforcement of forbidden words might inadvertently remove legitimate conjunctions.

5Second-Pass Verification

  • [ok] 目录仅 4 个文件、无脚本 — find humanizer-zh -type f（排除 .git）＝SKILL.md、README.md、LICENSE、.gitignore；SKILL.md 全文无可执行代码。
  • [ok] 无网络调用 — SKILL.md 仅出现 https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing（「参考」节的出处标注）；无 fetch/curl/requests/端点调用指令。
  • [ok] external_deps 为空是否低估 — 唯一 URL 是书目出处（维基百科页面），不构成运行时调用；模型不需要访问该页即可执行全部规则，故不列入 external_deps，改在此说明。
  • [ok] 无凭证读取 — token 扫描 API_KEY/token/secret/密码/keychain 全零命中；frontmatter 只声明 Read/Write/Edit/AskUserQuestion 四工具。
  • [ok] 24 类模式计数 — SKILL.md 小节编号 1-24 连续（内容 1-6、语言语法 7-12、风格 13-18、交流/填充 19-24），与 README 图例一致。
  • [ok] 「注入灵魂」段确实存在（不是只看标题） — 原文含「避免 AI 模式只是工作的一半。」、缺乏灵魂的写作迹象清单（6 条）与「改写前/改写后（鲜活）」示例。
  • [ok] 50 分评分与合格线 — 质量评分表 5 行维度各 /10、总分 /50；分档 45-50 优秀、35-44 良好、<35 需重修订，原文逐字核对通过。
  • [ok] allowed-tools 的写工具声明（判蓝依据） — frontmatter 逐字为 allowed-tools 列表：Read / Write / Edit / AskUserQuestion；README「处理文件内容」演示 `/humanizer-zh 请人性化 article.md 文件中的内容`。

6Verdict & Scenarios

  • Deconstructs subjective 'AI flavor' into 24 check items with Chinese trigger words + fixed rewrite mappings; reproducible and batch-ready, ideal as execution standards for editorial pipelines.
  • Goes beyond trace removal by establishing positive criteria for human voice (viewpoints, cadence, nuance, first-person, concrete impressions) alongside counter-examples of 'clean yet soulless' text.
  • Includes pre-delivery quantitative gating: 50-point rubric + 6-item checklist + comprehensive examples (with step-by-step change explanations) for self-evaluation and iteration.
  • Minimal risk surface: zero scripts, zero network access, zero credential reading—pure local text rewriting.
  • Tailored for Chinese context without blindly forcing English rules (e.g., notes title case and curly quotes as non-applicable or behaving differently).
    • Best for: Ideal for editing/reviewing any Chinese text drafted by LLMs: blog posts, WeChat official accounts, product marketing copy, and report drafts. Also suitable as a 'de-AI' benchmark in team copywriting pipelines (the wordlist can be used directly for manual proofreading). On the model side, it only requires an agent that can read SKILL.md, with no installation dependencies.
    Not for: Not suitable for fact-checking or citation retrieval (it only alters writing style, even demonstrating replacement sentence structures with specific sources); not suitable for strictly formal official or academic documents unless reviewed paragraph by paragraph; not suitable for circumventing AI detectors (explicitly discouraged by the author); and not suitable for restructuring full-length articles—its granularity operates at the sentence and paragraph level.
    Install Agent-Ready
    ① Verified Mirror 2026-09-15
    Option A · Direct DownloadDownload humanizer-zh.tar.gz
    sha256: 5537a2e650012165…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit 91f3d394db.