Pack · #336 trending skills repo · ↗ +4 stars this week

AI Check Skills

HoneyMeta/ai-check-skills

AI率检测报告生成、自动化降AI率、重复率/AI写作风格清理;支持 Codex、Claude Code、OpenCode,可生成 Word/DOCX 论文 AI率 PDF/HTML 报告。

Stars
35
This week
+4
Forks
3
Open issues
0
Skills inside
1
Star trend · last 9 days+4 ★
Oct 29, 2026: 31 starsOct 30, 2026: 31 starsOct 31, 2026: 31 starsNov 1, 2026: 32 starsNov 2, 2026: 32 starsNov 3, 2026: 33 starsNov 4, 2026: 33 starsNov 5, 2026: 33 starsNov 6, 2026: 33 starsNov 7, 2026: 35 stars

Get started

Install the skills

npx skills add HoneyMeta/ai-check-skills

Installs into Claude Code, Codex, Cursor, OpenCode and the other agents it finds on your machine.

Inside this repo

1 skill

Documentation

README

View on GitHub ↗

AI-check Skills

AI-check Skills 是一个面向 Codex、Claude Code、OpenCode 等 AI 编程助手的开源 Skill,用于 Word/DOCX 论文 AI 率检测报告生成、自动化降 AI 率、重复率/AI 写作风格清理与确认后改写。

该功能是 EasyIdea 科研工作台 的一小部分。如果你需要完整 的论文写作、DOCX 工作台、报告导入、标注修改、批量审阅和一键采纳/放弃能力,欢迎下载使用: https://honeymeta.com/easyidea

AI-check report preview

功能亮点

  • 从 .docx 中按文档顺序读取正文、摘要、结论等论文内容(含内容控件里的段落;表格默认跳过)。
  • 支持中文与英文论文。
  • 自动切分长文本,生成适合 AI 助手逐段判断的 chunk。
  • 提取可解释的 AI 写作风险信号,例如模板化表达、连接词密集、泛化套话、句式节奏过于均匀, 以及近年大模型常见的"赋能/助力/至关重要"、破折号堆叠、英文 delve / pivotal / underscore 等用语。
  • 同时记录缓和信号,例如引用编号、实验指标、模型/方法名、领域方法细节和定量结果支撑。
  • 由当前 AI 助手在可见对话中判断 aigcValue、verdict、reason 和 signals,脚本本身不黑箱打分。
  • 生成美观、可选择文本的 AI 率检测 PDF/HTML 报告。
  • 在用户确认后,将改写方案写入新的 DOCX,不会原地修改原文档;只改动有变化的字符, 原段落的加粗、上标引用、公式、图片、域代码都会保留。
  • 可选以 Word 修订(Track Changes) 形式写入,在 Word 里逐条接受/拒绝。
  • 改写校验:丢失引用编号(如 [4])或数字(如 2.1)的替换会被拒绝。
  • 支持 Codex、Claude Code、OpenCode,以及其他能运行本地脚本的 agent runtime。

独立 DOCX/PDF 管线

AI-check Skills 不依赖 LibreOffice、soffice、Microsoft Word 自动化、DOCX 转 PDF 渲染或截图渲染。

  • DOCX 解析和写回:python-docx
  • PDF 报告生成:reportlab
  • 确定性信号提取:仓库内置 Python 脚本

脚本不会调用 AI 模型,也不会自行输出最终 AI 概率。AI 助手负责逐段判断与改写,用户可以看到判断依据。

安装

在 Claude Code、Codex、OpenCode、Cursor、GitHub Copilot 等 AI 编程助手中直接发送:

帮我安装 https://github.com/HoneyMeta/ai-check-skills

安装完成后,重启客户端让新安装的 skill 生效。 也可以手动用通用的 skills 安装器:npx skills add HoneyMeta/ai-check-skills。

依赖(Python 3.10+,缺少时 AI 会按 SKILL.md 自动安装):

python -m pip install "python-docx>=1.1" "reportlab>=4"

目录结构

ai-check/
  SKILL.md
  agents/openai.yaml
  scripts/
    ai_check.py
    ai_rewrite_docx.py
  references/
    ai_rate_signals.md
    report_format.md
docs/
  assets/
    ai-check-report-preview.png
tests/
  test_ai_check.py

快速开始

在 AI 编程助手中发送:

使用 ai-check-skills 帮我为 XXXX(文档路径)生成 AI率检测报告

例如:

使用 ai-check-skills 帮我为 F:\papers\main.docx 生成 AI率检测报告

AI 助手会读取 DOCX,生成可解释的片段信号,并输出:

  • 可选择文本的 PDF 报告
  • HTML 报告
  • summary JSON

如果需要继续降 AI 率,可以在报告生成后让 AI 助手先列出改写方案;用户确认后再写入新的 DOCX。 原始 DOCX 不会被原地修改。

示例提示词

  • 使用 ai-check-skills 帮我为 F:\papers\main.docx 生成 AI率检测报告
  • 根据 AI-check 报告,先给我看需要修改的片段和改写方案,确认后再写出新的 Word。
  • 确认后用修订模式写入新的 Word,我要在 Word 里逐条接受。
  • 只检测,不改 Word。

注意事项

  • 报告仅供参考。
  • 不要宣称结果等同于知网、维普、Turnitin 或其他闭源检测平台。
  • 改写时应保留事实、术语、公式、数字、引用编号和结论方向。
  • 如果原文缺少证据,不要编造实验、数据、案例或引用;应降低语气、补充边界,或使用已有上下文。

License

MIT

English

AI-check Skills is an open-source Codex/OpenCode-style skill for AI-rate report generation, automated AI-writing reduction, and academic repetition/AI-writing style cleanup. It generates a polished, selectable-text PDF/HTML report from a Word document, then applies confirmed revisions to a new DOCX only after user approval.

This project is a small part of EasyIdea Research Workbench. For a full research-writing workflow with a live DOCX workbench, report import, annotations, batch review, and one-click accept/reject, download EasyIdea: https://honeymeta.com/easyidea

Highlights

  • Reads .docx files in document order (including content controls; tables skipped by default) and extracts body, abstract, and conclusion text. Works for Chinese and English papers.
  • Splits long academic text into reviewable chunks.
  • Extracts explainable AI-writing signals such as template phrases, dense connectors, generic claims, and uniform sentence rhythm.
  • Keeps mitigating evidence such as citations, metrics, method names, domain details, and quantitative results.
  • Lets the active AI assistant judge aigcValue, verdict, reason, and signals in the visible conversation.
  • Generates a polished AI-rate report as HTML and selectable-text PDF.
  • Applies user-confirmed replacements to a new DOCX without modifying the original document. Only the changed characters are rewritten, so bold text, superscript citations, equations, images and fields survive; --track-changes writes the edits as Word revisions you can accept or reject.
  • Rejects rewrites that drop a citation marker such as [4] or a number such as 2.1.
  • Supports Codex, Claude Code, OpenCode, and similar agent runtimes that can run local scripts.

Independent Pipeline

AI-check Skills does not require LibreOffice, soffice, Microsoft Word automation, DOCX-to-PDF rendering, or screenshot rendering.

  • DOCX parsing and rewriting: python-docx
  • PDF report generation: reportlab
  • Deterministic signal extraction: bundled Python scripts

The scripts do not call AI models and do not produce a final AI probability by themselves. The AI assistant remains responsible for judgment and rewriting so the user can see the reasoning.

Install

Ask your AI coding assistant (Claude Code, Codex, OpenCode, Cursor, GitHub Copilot, …):

Install https://github.com/HoneyMeta/ai-check-skills

Restart the assistant afterwards so the new skill is loaded. To install manually, use the cross-agent skills installer: npx skills add HoneyMeta/ai-check-skills. Dependencies (Python 3.10+): python -m pip install "python-docx>=1.1" "reportlab>=4".

Quick Start

Send a prompt like:

Use ai-check-skills to generate an AI-rate report for C:\papers\main.docx.

For Chinese workflows, you can use:

使用 ai-check-skills 帮我为 XXXX(文档路径)生成 AI率检测报告

The assistant will scan the DOCX, produce an explainable PDF/HTML report, and only write a rewritten DOCX after you approve the replacement plan.

Safety Notes

  • The report is for reference only.
  • Do not claim that it is equivalent to CNKI, VIP, Turnitin, or another closed detector.
  • Preserve facts, terminology, formulas, numbers, citation markers, and conclusion direction during rewriting.
  • If the source lacks evidence, do not invent experiments, data, cases, or citations.

Releases

Recent versions

All releases ↗
  1. v0.2.0