← 返回 Skills 市场

Humanize Chinese

Name: Humanize Chinese
Author: erongcao

作者 Yirong · GitHub ↗ · v2.1.0 · MIT-0

cross-platform ✓ 安全检测通过

137

总下载

当前安装

版本数

在 OpenClaw 中安装

/install erong-humanize-chinese

功能描述

检测并去除中文文本中的AI写作痕迹。当用户说"去AI味"、"改得自然一点"、"太机器了"、"帮我润色"、"去掉AI感"时使用。支持文件输入或直接粘贴文本，输出改写后文本和对比报告。适用于论文、文案、公众号、社交媒体等场景。

使用说明 (SKILL.md)

Humanize Chinese v2.1

中文 AI 文本人性化改写工具。检测 AI 写作特征，去除机器腔调，保留原意。

工作流

用户给我文本 → 我运行脚本 → 输出改写后文本

方式一：直接粘贴（最常用）

用户直接把文字发给我，我调用脚本处理：

# 对比 + 改写（激进模式，适合明显AI味的内容）
python3 ~/.openclaw/skills/humanize-chinese-2-0-0/scripts/compare_cn.py --scene social -a

# 纯检测，不改写（想先看看AI味有多重）
python3 ~/.openclaw/skills/humanize-chinese-2-0-0/scripts/detect_cn.py -v

# 指定场景改写
# --scene general  通用（默认）
# --scene social   社交媒体、短文
# --scene tech     科技文章
# --scene formal   正式文章/报告
# --scene chat     对话/聊天

方式二：文件处理

# 改写并保存到新文件
python3 ~/.openclaw/skills/humanize-chinese-2-0-0/scripts/humanize_cn.py 输入.txt -o 输出.txt --scene social -a

# 检测AI分数
python3 ~/.openclaw/skills/humanize-chinese-2-0-0/scripts/detect_cn.py 输入.txt -v

方式三：批量处理

# 批量改写所有 markdown 文件
for f in *.md; do
  python3 ~/.openclaw/skills/humanize-chinese-2-0-0/scripts/humanize_cn.py "$f" --scene tech -a -o "${f%.md}_clean.md"
done

检测评分体系

分数	等级	含义
0-24	🟢 低	读起来像真人
25-49	🟡 中	有一些AI特征
50-74	🟠 高	大概率AI写的
75-100	🔴 很高	基本确定是AI

核心检测类别

类别	权重	示例
🔴 三段式套路	高	首先…其次…最后
🔴 机械连接词	高	值得注意的是、综上所述
🔴 空洞宏大词	高	赋能、闭环、新质生产力
🟠 引流开场白	高	让我们一起、揭秘、划重点
🟠 模糊概括	高	有研究表明、专家认为
🟠 AI高频词	中	助力、彰显、底层逻辑
🟠 套话废话	中	值得一提的是、毫无疑问
🟡 节奏单一	低	句子长度高度一致

改写原则（基于研究）

打破固定句式 — AI 喜欢"首先/其次/最后"，改成自然过渡
增加具体细节 — "有研究表明" → "XX团队在2025年发表的数据"
隐藏中心句 — AI 开头就给结论，改写到段中或段尾
注入个人视角 — 加"我觉得"、"查了资料后发现"
句式变化 — 长短句交替，主动被动互换

规则配置

所有检测词和替换规则在：

~/.openclaw/skills/humanize-chinese-2-0-0/scripts/patterns_cn.json

可自行添加新的 AI 特征词或调整替换映射。

安全使用建议

This skill appears to do what it says: local detection and rewrite of Chinese text using the included scripts and patterns. Before installing or running it, consider: 1) The SKILL.md uses a hard-coded installation path (~/.openclaw/skills/humanize-chinese-2-0-0) and there is a small version/ownerId mismatch in metadata — confirm the actual file path or update the commands. 2) The humanizer may insert fake-looking specifics (e.g., invented study citations) to make text feel more 'human' — review outputs for factual accuracy. 3) Run the scripts on non-sensitive text first (they are deterministic/non-networked but include randomness) and inspect patterns_cn.json if you want to customize what gets detected/replaced. 4) Because the skill executes local Python scripts (exec allowed in SKILL.md), only install/run it from a source you trust; if you cannot verify the source, run it in a sandboxed environment.

功能分析

Type: OpenClaw Skill Name: erong-humanize-chinese Version: 2.1.0 The skill bundle is a legitimate utility for detecting and 'humanizing' AI-generated Chinese text. It uses a series of Python scripts (detect_cn.py, humanize_cn.py, style_cn.py) to perform regex-based pattern matching and phrase replacement. There is no evidence of data exfiltration, network activity, or malicious execution; subprocess calls are limited to internal script communication without shell-injection vulnerabilities, and the logic relies entirely on the provided patterns_cn.json configuration.

能力评估

✓ Purpose & Capability

Name/description match the included scripts: detection, comparison, style transforms, and humanization. The declared lack of required env vars/credentials is consistent with local text-processing functionality. One minor mismatch: SKILL.md and scripts reference a specific install path (~/.openclaw/skills/humanize-chinese-2-0-0) and the _meta.json/registry metadata show version 2.1.0 / different ownerId — this looks like a packaging/path/version inconsistency, not a security issue, but it may break the exact commands in SKILL.md unless the skill is installed to that path.

ℹ Instruction Scope

Runtime instructions explicitly run the included Python scripts via exec and read/write local files (stdin/stdout and optional output files). The scripts only operate on local text and the shipped patterns JSON. They do not read unrelated system config paths or environment variables, nor do they call external network endpoints. Note: the tool will insert fabricated-looking concrete details (e.g., templated 'XX 在2024年...' replacements) as part of humanization—this is a behavioral risk (possible introduction of invented citations), not a covert exfiltration issue.

✓ Install Mechanism

No install spec (instruction-only) and included code files run locally. There are no downloads or external installers, so no high-risk install behavior. The SKILL.md assumes files live under a hard-coded ~/.openclaw path which may not match actual install location; that is a usability/consistency issue rather than an installation security risk.

✓ Credentials

The skill requests no environment variables, no credentials, and no config paths beyond its own patterns_cn.json. That is proportionate for a purely local text-processing tool.

✓ Persistence & Privilege

always is false and the skill does not request any elevated persistence or attempt to alter other skills or global agent settings. It performs file I/O only on user-specified inputs/outputs and its own pattern file.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install erong-humanize-chinese
安装完成后，直接呼叫该 Skill 的名称或使用 /erong-humanize-chinese 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v2.1.0

- 新增场景支持（通用、社交、科技、正式、对话），可指定不同改写风格 - 优化检测与改写流程，支持批量处理和文件输入 - 引入详细的AI痕迹检测评分体系与类别说明 - 提供具体的改写原则，提升文本自然度与人性化 - 支持自定义检测规则和替换映射，便于扩展

元数据

Slug erong-humanize-chinese

版本 2.1.0

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 1

常见问题