avoid-ai-writting-flavor

Use when writing, editing, or reviewing academic papers (especially LLM-drafted) to eliminate significant AI-written patterns that reviewers spot instantly,...

Security Scan

The skill's instructions, scope, and requirements are coherent with its stated purpose (removing AI 'flavor' from academic writing) and it makes no disproportionate environment, install, or persistence demands.

Purpose & Capability

Name/description match the SKILL.md content: the checklist and formatting rules directly address eliminating AI-written patterns in academic papers (citations, punctuation, figures, math notation, hallucination checking). No unrelated credentials, binaries, or installs are requested.

Instruction Scope

Instructions stay within purpose (formatting, citation verification, figure and math style). They explicitly require manual verification of every citation on Google Scholar/publisher sites; this is appropriate for the goal but implies the agent (or human) must access external web resources or perform manual checks — ensure the agent's browsing capability and privacy constraints before relying on automated verification.

Install Mechanism

Instruction-only skill with no install spec and no code files. Lowest-risk install profile (nothing written to disk).

Credentials

No environment variables, credentials, or config paths are requested. Required access is limited to author-followable style checks and external citation lookups, which are proportional to the described purpose.

Persistence & Privilege

Flags indicate normal, non-persistent usage (always: false). The skill does not request permanent presence or modification of other skills or system settings.

Assessment

This skill is coherent and instruction-only, so risk is low. Before using: ensure whoever or whatever executes the checklist can legally and technically access Google Scholar/publisher sites (the SKILL.md relies on manually verifying citations), avoid sending entire manuscripts to untrusted external services for verification, and treat the 'verify every citation' step as a human review if your agent lacks reliable web access or has privacy constraints.

消除学术论文中的 AI 味

核心原则

经验丰富的审稿人能一眼识别 AI 生成的文本。以下每一条都是会被标记的信号。逐条检查，全部消除。

检查清单

1. 引用格式 & 真实性验证

• 引用括号前必须加空格：OpenVLA (Brohan et al., 2023) 而非 OpenVLA(Brohan et al., 2023)
• 多引用保持格式统一：(Brohan et al., 2022, 2023; Driess et al., 2023)
• 逐条在 Google Scholar 验证每一条引用。LLM 极易编造看似合理但不存在的文献，这是学术不端红线，无例外

2. 加粗与斜体

适当使用 bold / italic 标注关键论断和术语，节省读者注意力。每页几处即可，切忌整句加粗。

3. 去掉多余引号

AI写作最爱用夸张引号做强调，一眼暴露：

• Bad: The challenge is rarely "predicting the next action correctly."
• Good: The challenge is rarely predicting the next action correctly.

引号仅用于直接引述或首次出现的专有术语。

4. 禁用破折号连接句子

破折号（em-dash）连接从句是 AI写作 标志性句式，改写为独立短句或逗号分隔：

• Bad: ...scales up the policy itself—through larger data, more unified action spaces, and stronger representations—including RT-1, RT-2, ...
• Good: ...scales up the policy through larger data, more unified action spaces, and stronger representations. Representative works include RT-1 and RT-2.

5. 避免反复列举

反复罗列长串引用：在 Related Work 完整列一次即可，其他段落只引 2–3 个代表性工作。每段都堆 10+ 引用 = AI 味。

6. 图片格式

• 避免 AI 生成的位图质感（nanobanana 风）
• 导出为 PDF 矢量格式，确保文字可选中复制
• 图表应看起来像 TikZ / draw.io / PPT 制作，而非 AI 绘图工具

7. 数学符号规范

类型	格式	示例
向量	加粗	$\mathbf{x}$, $\mathbf{a}$
标量	不加粗	$t$, $\alpha$
矩阵	大写加粗	$\mathbf{W}$, $\mathbf{A}$

向量标量不区分 = 不专业或粗糙生成。

7. 实验图 Caption

• 图里面不要有大标题横幅（一眼nanobanana），只保留图下方的简洁描述性 caption（可以尽可能把细节描述详细些，只要文字本身避免AI味）
• 大号 "TITLE" banner 是 LLM 默认输出的典型特征

8. 幻觉引用——最危险的 AI 痕迹

单独强调：这是所有 AI 痕迹中后果最严重的一条。

• 每一条引用都必须手动在 Google Scholar / 出版商网站核实
• LLM 会用真实作者名编造不存在的论文
• 一条幻觉引用 = 学术造假认定风险
• 不要信任任何你没有亲自验证的引用。一条都不行

速查表

AI 信号	修复
Author(Year) 无空格	加空格 Author (Year)
引号做强调	删引号
破折号连接从句	拆成短句
每段堆 10+ 引用	Related Work 列一次，其余引 2–3 个
位图风格图表	矢量 PDF，可编辑格式
向量标量同样式	向量加粗，标量不加粗
图上大标题	仅保留下方 caption
未验证的引用	立刻查 Google Scholar