anti-defensive-writing-Skill

Anti-Defensive Writing · 学术写作原则

阻止论文的防御性写作 —— 一个轻量级、零依赖的 AI 写作助手 Skill + 提示词，复制即用

English README · 中文 Skill · English Skill

---

Codex 插件安装

作者：Adkid-Zephyr。本仓库同时提供中文和英文 Skill。Codex 可以安装 Skill，也可以安装打包后的插件；仓库安装不等于官方插件目录已上架。

安装本仓库插件：

codex plugin marketplace add Adkid-Zephyr/anti-defensive-writing-Skill
codex plugin add anti-defensive-writing@adkid-zephyr-writing

安装后在新对话中选择 Anti-Defensive Writing 插件，并使用中文或英文 Skill 修改论文。

也可以把 中文 Skill 链接 交给 Codex，要求它安装这个 Skill。

插件发布包包含相同的两份 Skill，没有 MCP 服务或额外运行依赖。通过 python3 scripts/build_plugin.py 生成可上传到 OpenAI 发布后台的 ZIP；隐私政策。官方目录发布状态以 OpenAI 后台为准，GitHub release 不是官方目录上架记录。

这是什么

很多人用 AI 改论文，改完越看越心虚：审稿人还没开口，自己先把「遗憾的是」「效果有限」「仍明显落后」写满了。结构是工作汇报式的，语言是自我审查式的——等于提前把刀子递到审稿人手里，还附赠一份使用说明。

这叫防御性写作（Defensive Writing）。

解药只有一条原则：

论文是一场学术发布会，不是项目总结。 发布会只讲一件事：你最强的那个优势。

打不赢的指标不设为比赛，不占优的结果不说输，实验不是结果仓库、是论证工具。论文的说服力来自主张与证据高度一致，而不是比较项目最多。

效果对比

	防御性写作	发布会原则
劣势表述	虽然本方法在两个同领域数据集上提高了准确率，但必须承认，我们尚未测试跨领域泛化，因此应谨慎理解这些结果。	本方法在两个同领域数据集上提高了准确率；跨领域泛化尚未评估。
结构	我们首先尝试了 A，然后尝试了 B，最后选择了 C	问题 X 很关键，现有方法缺 Y，本文提出 Z，证据如下
实验	附录式堆结果，全覆盖无重点	每个实验承担一个论证职责
结论	最后一段突然自我否定	只强化记忆点：解决了什么、证明了什么、为什么重要

快速开始

方式一：复制提示词（任何 AI 都能用）

直接复制 prompts/精简版提示词.txt（English），粘贴到你和 AI 的对话开头，然后发给它你的论文段落。复制即用，零门槛。

方式二：安装 Skill（Claude Code / Cursor / 其他 Agent 工具）

把对应语言的 skill 目录整个拷进你的 skills 目录即可：

# 中文版
cp -r skills/anti-defensive-writing ~/.claude/skills/

# English version
cp -r skills/anti-defensive-writing-en ~/.claude/skills/

之后对 AI 说「帮我改改这段论文」「润色一下 abstract」，它会自动按发布会原则工作。

十二条规则速览

• 叙事：只围绕优势组织 · 不写工作汇报 · 打不过的维度不设为比赛 · 优势必须明说 · 控制比较范围 · 允许彻底重构故事
• 语言：禁用自我削弱表达 · 不占优的结果不说输 · 不把局部现象上升为对整体方法的否定
• 实验：每个实验必须有论证职责
• 结构：摘要引言 = 发布会开场 · 结论只强化记忆点

完整规则与决策流程见 SKILL.md。

仓库结构

anti-defensive-writing/
├── skills/
│   ├── anti-defensive-writing/      # 中文 Skill
│   └── anti-defensive-writing-en/   # English Skill
├── prompts/
│   ├── 精简版提示词.txt           # 中文精简提示词(复制即用)
│   └── quick-prompt-en.txt       # English quick prompt
└── README.md / README_EN.md

适用场景

• 写/改论文摘要、引言、结论
• 压缩论文篇幅（砍哪段，这里给了决策优先级）
• 组织实验章节（每个实验该承担什么职责）
• 写 rebuttal 前的自查（不给审稿人递刀子清单）

License

MIT。欢迎 PR，欢迎转发给你的同门。

Anti-Defensive Writing

Stop defensive writing in academic papers — a lightweight, zero-dependency Skill + prompt pack for AI writing assistants. Copy and go.

中文 README · English Skill · 中文 Skill

---

Install as a Codex plugin

Publisher: Adkid-Zephyr. This repository includes Chinese and English skills. A repository-installed plugin is separate from a published listing in the official plugin directory.

codex plugin marketplace add Adkid-Zephyr/anti-defensive-writing-Skill
codex plugin add anti-defensive-writing@adkid-zephyr-writing

After installation, start a new conversation, select Anti-Defensive Writing, and use its Chinese or English skill to revise your paper. You can also give Codex the Chinese skill URL or English skill URL and ask it to install the skill.

The plugin bundles the same two skills and has no MCP server or extra runtime dependencies. Run python3 scripts/build_plugin.py to build the ZIP for OpenAI submission. See the privacy policy. Official directory publication is determined by the OpenAI dashboard; a GitHub release is not an official directory listing.

What is this

Many people use AI to polish their papers and end up with something that reads like a self-audit: before the reviewer says a word, the text is already full of "unfortunately", "limited improvement", and "still lags behind". Work-report structure, self-censoring language — it hands the reviewer a knife, with a user manual attached.

This is defensive writing.

The antidote is one principle:

A paper is a press conference, not a project summary. A press conference says exactly one thing: your strongest advantage.

Never set a contest you cannot win. Never say you lost. Experiments are tools of argument, not a warehouse of results. Persuasion comes from tight claim-evidence alignment, not from the number of comparisons.

Before / After

	Defensive writing	Press-release principle
Weakness framing	Although our method improves accuracy on two in-domain datasets, we must acknowledge that cross-domain generalization has not been tested, so these results should be interpreted cautiously.	Our method improves accuracy on two in-domain datasets; cross-domain generalization has not been evaluated.
Structure	We first tried A, then B, and finally chose C	Problem X matters, existing methods lack Y, we propose Z, evidence follows
Experiments	An appendix-style pile of results	Every experiment carries one argumentative duty
Conclusion	Sudden self-negation in the final paragraph	Reinforces the takeaway only

Quick start

Option 1: Copy the prompt (works with any AI)

Copy prompts/quick-prompt-en.txt (中文) and paste it at the start of your AI conversation, then send your paper draft. ~400 words, zero setup.

Option 2: Install the Skill (Claude Code / Cursor / other agent tools)

Copy the skill folder of your preferred language into your skills directory:

# English
cp -r skills/anti-defensive-writing-en ~/.claude/skills/

# 中文版
cp -r skills/anti-defensive-writing ~/.claude/skills/

Then just ask "polish this paragraph" or "revise my abstract" — the agent applies the press-release principle automatically.

The 12 rules at a glance

• Narrative: organize around strengths only · no work-report chronology · never set a contest you can't win · state advantages explicitly · limit comparison scope · allow full story restructuring
• Language: ban self-weakening phrases · never say you lost · never turn a local observation into a verdict on the whole method
• Experiments: every experiment needs an argumentative duty
• Structure: abstract & intro = press-conference opening · conclusion only reinforces the takeaway

Full rules and decision flow in SKILL.md.

Repository structure

anti-defensive-writing/
├── skills/
│   ├── anti-defensive-writing/      # 中文 Skill
│   └── anti-defensive-writing-en/   # English Skill
├── prompts/
│   ├── 精简版提示词.txt           # 中文精简提示词
│   └── quick-prompt-en.txt       # English quick prompt (copy & paste)
└── README.md / README_EN.md

Use cases

• Writing / revising abstracts, introductions, conclusions
• Cutting paper length (the decision rules tell you what to cut first)
• Organizing the experiments section
• Pre-rebuttal self-check (the "don't hand the reviewer a knife" checklist)

License

MIT. PRs welcome — and feel free to share it with your labmates.