# ai-forgets-itself

Adopt this skill: https://proskills.md/skills/other/er-go-proxy-ai-forgets-itself-ai-forgets-itself

Page: https://proskills.md/skills/other/er-go-proxy-ai-forgets-itself-ai-forgets-itself
This document is the public skill bundle for that page. It contains SKILL.md and every helper file the catalog carries. No login is required.

## File: ai-forgets-itself/SKILL.md
size: 5934
upstream: https://github.com/er-go-proxy/ai-forgets-itself/blob/main/ai-forgets-itself/SKILL.md

```text
---
name: "ai-forgets-itself"
description: "Refine user-owned academic writing so the author's real viewpoint, reasoning, and voice come back to the front. Use when the user already has their own notes, outline, or draft and wants help with essays, reports, literature reviews, or thesis sections in one of three modes: (1) expand and deepen arguments, (2) reduce generic AI-like phrasing and rewrite toward natural student or author voice, or (3) combine de-AI revision with further development. Preserve the user's stance, avoid fabricating evidence, citations, or personal experience, and prefer claim-evidence-analysis over template-like summary prose."
---

# AI Forgets Itself

## Overview

Turn user-provided viewpoints, notes, outlines, or draft paragraphs into stronger academic prose without ghostwriting a brand-new position. Preserve the user's intellectual ownership, make the reasoning more specific and evidence-aware, and reduce generic or over-smoothed phrasing that reads like stock AI output.

## Workflow Decision Tree

1. Identify the mode.
- Use `expand` when the user asks to expand, deepen, refine, enrich, or make a section more substantial.
- Use `de-ai` when the user asks to lower AIGC rate, remove AI flavor, humanize the wording, reduce repetition, or make the text sound more like a student or author.
- Use `hybrid` when both goals appear or when the safest path is "de-AI first, then expand."

2. Extract the author signal before rewriting.
- Reduce the user's input to 1-3 sentences that state the core claim, intended section function, and any explicit stance.
- If the author signal is unclear, infer the narrowest defensible version from the provided text or ask for the missing viewpoint.

3. Load only the references that fit the mode.
- Read [ai-style-signals.md](references/ai-style-signals.md) for `de-ai` and `hybrid`.
- Read [chinese-academic-voice.md](references/chinese-academic-voice.md) when matching natural student or Chinese academic prose.
- Read [expansion-patterns.md](references/expansion-patterns.md) for `expand` and `hybrid`.
- Read [revision-checklist.md](references/revision-checklist.md) before finalizing any rewrite.

4. Use the helper script when useful.
- Run `scripts/flag_ai_style_terms.py` on a draft or paragraph when you want a quick heuristic scan for clustered high-risk phrases, repetitive transitions, or overly even sentence rhythm.
- Treat the script output as a heuristic, not a verdict.

5. Return a structured answer.
- Match the output format specified below for the selected mode.

## Non-Negotiables

- Preserve the user's stance; do not replace it with a cleaner but different argument.
- Do not fabricate citations, evidence, data, quotations, personal experience, or case studies.
- Do not promise to "beat" detectors or guarantee a specific AIGC score.
- Do not use mechanical synonym swapping, random clause inversion, or deliberate grammatical degradation.
- Prefer narrower and more defensible claims over inflated conclusions.
- Replace empty summary language with claim-evidence-analysis.
- Mark missing evidence or unverifiable details as `需作者补充`.
- If the user provides only a topic and no viewpoint, help structure the problem but do not pretend to have recovered a personalized voice.

## Mode: expand

1. State the paragraph or section's job.
- Explain what it is trying to prove, compare, explain, or qualify.

2. Expand the reasoning instead of merely expanding the surface length.
- Prefer one or more of these moves:
  - define a concept more precisely
  - explain the causal or analytical link
  - compare with an alternative interpretation or case
  - add limits, conditions, or counterarguments
  - clarify the practical, theoretical, or methodological consequence

3. Keep the prose authorial.
- Avoid abstract noun stacking and ceremonial transitions.
- Retain uncertainty if the draft legitimately contains uncertainty.

4. Output in this order:
- `观点提炼`
- `可扩写点`
- `扩写版本`
- `需作者补充`

## Mode: de-ai

1. Compress before rewriting.
- Reduce the input to 1-3 sentences that capture the author's real point.

2. Diagnose AI-like signals.
- Flag generic openings, repetitive transitions, stacked evaluation words, conclusion-heavy paragraphs, over-even sentence shapes, and abstract claims unsupported by evidence.

3. Rewrite toward human academic rhythm.
- Prefer specific judgments over grand claims.
- Prefer natural sentence variation over uniform cadence.
- Prefer direct logical links over stock transition words.
- Keep the text clear and disciplined; "human" does not mean colloquial or sloppy.

4. Output in this order:
- `核心观点`
- `AI腔诊断`
- `去模板后的版本`
- `修改理由`
- `需作者确认`

## Mode: hybrid

1. Extract the author's core point.
2. Remove AI-like padding and generic phrasing.
3. Expand only after the argument becomes clear and local.

4. Output in this order:
- `核心观点`
- `去AI后的骨架`
- `深化扩写版本`
- `修改说明`
- `需作者补充`

## Style Targets

- Write like a thoughtful student or emerging scholar, not a marketing brochure.
- Use scoped judgments such as "more convincing here," "mainly reflects," or "works better for this part" instead of constant "important," "comprehensive," or "groundbreaking."
- Let some sentences stay plain if plain wording is the clearest option.
- Use transitions sparingly and specifically.
- Prefer concrete verbs over empty abstract nouns.
- Keep the level of confidence proportional to the evidence.

## Final Check

Before sending the final answer, apply [revision-checklist.md](references/revision-checklist.md).
- If several high-risk phrases remain clustered in the same paragraph, revise again.
- If the rewrite drifts away from the user's stance, roll it back.
- If the paragraph sounds smoother but not smarter, deepen the analysis instead of polishing further.
```

## File: README.md
upstream: https://github.com/er-go-proxy/ai-forgets-itself/blob/main/README.md

````text
<div align="center">

# ai-forgets-itself.skill

### 让 AI 忘掉它自己.skill

![Platforms](https://img.shields.io/badge/Platforms-Claude%20Code%20%7C%20Codex%20%7C%20OpenClaw-2563eb)
![Modes](https://img.shields.io/badge/Modes-Expand%20%7C%20De--AI%20%7C%20Hybrid-16a34a)
![Languages](https://img.shields.io/badge/Languages-ZH%20%7C%20EN-f59e0b)
![Type](https://img.shields.io/badge/Type-Skill-7c3aed)

<p>
  <strong>一个面向论文、课程作业、综述和毕业论文写作的 Skill，适用于 Claude Code、Codex、OpenClaw 以及其他支持 Skill 的 Agent / IDE / 大模型。</strong><br/>
  <strong>帮助你在保留自己观点的前提下扩写、修订、重组，并削弱模板化 AI 腔，让文本更像真实作者写出来的学术表达。</strong>
</p>

<p>
  <strong>A portable academic-writing skill for Claude Code, Codex, OpenClaw, and other skill-enabled agents.</strong><br/>
  <strong>It expands, refines, and de-AIs user-owned drafts while keeping the author's actual viewpoint in control.</strong>
</p>

</div>

---

## 中文

> “不是把 AI 写过的东西换几个词，而是让文本重新带回作者自己的思路、判断和论证痕迹。”

**ai-forgets-itself** 是一个给“已经有自己的观点，但不想让文本一股模板味”的人准备的 Skill。  
它不追求“骗过某个检测器”，而是把空泛、均匀、过度平滑的 AI 式表述，改成更像学生、作者或研究者本人会写出来的文字。

它特别适合这几类场景：

- 你已经有自己的核心观点，但段落太短、太薄，想扩写得更扎实。
- 你已经有草稿，但里面有明显的 AI 腔、模板感、空话、套话，想改得更自然。
- 你想先把 AI 味压下去，再继续深化、扩展和重组。

### Features

- `Expand`：围绕你的原观点补足论证链、概念界定、比较对象、边界条件和意义。
- `De-AI`：提炼你的真实观点，删除模板化表达，削弱 AI 常见的套话和过度平滑语气。
- `Hybrid`：先去 AI 腔，再在不偏离立场的前提下继续深化扩写。
- `Author Signal First`：先抓作者真正想说什么，再开始改写，而不是只改表面措辞。
- `Heuristic Scanner`：附带一个轻量脚本，能标记常见 AI 高风险表达簇、重复过渡词和过于平均的句式节奏。
- `Cross-Agent Friendly`：对自动触发和显式调用都做了适配，适合多种支持 Skill 的 Agent。

### Why This Skill

- 保留作者自己的立场和思考轨迹
- 优先改善论证质量，而不是机械同义替换
- 降低“流畅但空”的模板化表达
- 兼顾中文学术写作习惯与英文草稿修订
- 显式调用稳定，跨 Agent 迁移成本低

### Compatibility

- Claude Code
- Codex
- OpenClaw
- 其他支持 Skill 的 Agent、IDE 或工作流

### Install

#### 方式一：直接对话安装

在 Claude Code、Codex、OpenClaw 等支持 Skill 的 Agent 中，直接说：

```text
安装这个 skill：https://github.com/er-go-proxy/ai-forgets-itself
```

如果你最终上传到其他仓库地址，把上面的 URL 替换成你的真实 GitHub 仓库地址即可。

#### 方式二：手动添加

```bash
git clone https://github.com/er-go-proxy/ai-forgets-itself.git
```

把 [`ai-forgets-itself/`](./ai-forgets-itself) 这个文件夹复制到你的本地 Skill 目录，并保留目录名 `ai-forgets-itself`。

示例：

```text
~/.codex/skills/ai-forgets-itself
```

Windows:

```text
C:\Users\<YourName>\.codex\skills\ai-forgets-itself
```

其中：

- [`ai-forgets-itself/SKILL.md`](./ai-forgets-itself/SKILL.md) 是最通用的核心入口
- [`ai-forgets-itself/references/`](./ai-forgets-itself/references) 是配套参考资料
- [`ai-forgets-itself/scripts/flag_ai_style_terms.py`](./ai-forgets-itself/scripts/flag_ai_style_terms.py) 是可选的辅助脚本
- [`ai-forgets-itself/agents/openai.yaml`](./ai-forgets-itself/agents/openai.yaml) 是部分生态会读取的可选元数据

不同 Agent 的自动触发机制不完全一样，但通常主要看 `SKILL.md` 里的 `description`。  
所以这个 skill 既可以靠相关任务描述自动触发，也可以显式写 `$ai-forgets-itself` 来更稳定地调用。

### Usage

```text
用 $ai-forgets-itself 按 expand 模式扩写这段论文内容，不要改变我的立场，重点补足论证链、比较对象和限制条件。
```

```text
用 $ai-forgets-itself 按 de-ai 模式处理这段文字，先提炼我的核心观点，再去掉模板化 AI 腔，改成更像学生论文的表达。
```

```text
用 $ai-forgets-itself 按 hybrid 模式处理这段内容，先去 AI 味，再继续深化扩写；如果有需要我补充的事实、论据或引用，请单独列出来。
```

### Repo Layout

```text
ai-forgets-itself/
├─ README.md
├─ LICENSE
├─ .gitignore
└─ ai-forgets-itself/
   ├─ SKILL.md
   ├─ agents/
   │  └─ openai.yaml
   ├─ references/
   │  ├─ ai-style-signals.md
   │  ├─ chinese-academic-voice.md
   │  ├─ expansion-patterns.md
   │  └─ revision-checklist.md
   └─ scripts/
      └─ flag_ai_style_terms.py
```

---

## English

**ai-forgets-itself** is a portable writing skill for Claude Code, Codex, OpenClaw, and other skill-enabled agents.  
It is designed for users who already have their own viewpoint and want help expanding, refining, and de-AI-ing academic prose without outsourcing the argument itself.

### Features

- Expand thin paragraphs into stronger claim-evidence-analysis writing
- Reduce templated AI-like phrasing and over-smoothed prose
- Preserve the author's original stance and intent
- Support `expand`, `de-ai`, and `hybrid` workflows
- Include an optional heuristic scanner for clustered AI-style phrases

### Install

#### Option 1: install by chatting with your agent

```text
Install this skill: https://github.com/er-go-proxy/ai-forgets-itself
```

#### Option 2: add it manually

```bash
git clone https://github.com/er-go-proxy/ai-forgets-itself.git
```

Then copy [`ai-forgets-itself/`](./ai-forgets-itself) into your local skills directory and keep the folder name as `ai-forgets-itself`.

The portable core lives in [`ai-forgets-itself/SKILL.md`](./ai-forgets-itself/SKILL.md).  
[`ai-forgets-itself/agents/openai.yaml`](./ai-forgets-itself/agents/openai.yaml) is optional ecosystem metadata for platforms that read it.

Explicit invocation with `$ai-forgets-itself` is the most reliable cross-agent option.

### Usage

```text
Use $ai-forgets-itself to expand this thesis paragraph without changing my viewpoint. Focus on claim-evidence-analysis instead of filler.
```

```text
Use $ai-forgets-itself to de-AI this draft, keep my core argument, remove stock phrasing, and rewrite it in a more natural student voice.
```
````
