open-seo deslop 技能实战：Before/After 改写示例全解析，一次看懂 15 种 AI 写作套路的修复方法

open-seo deslop 技能实战：Before/After 改写示例全解析，一次看懂 15 种 AI 写作套路的修复方法

【免费下载链接】open-seoOpen source alternative to Semrush and Ahrefs项目地址: https://gitcode.com/GitHub_Trending/op/open-seo

在 open-seo 仓库的 Agent 技能目录中，deslop是一套用于"去除 AI 写作痕迹"的写作技能：它定义了核心规则、快速检查清单和评分标准，而 references/examples.md 则把这套规则落成了 15 组可直接对读的 Before/After 改写示例，覆盖学术论文、求职信、博客、摘要、基金申请书等多种文体。本文以该文档为主体，逐例继承原文的改写前后对照与修改说明，并结合 SKILL.md 的 10 条核心规则与 Quick Checks 清单，讲清每个示例到底在修哪一类问题、修复手法可以怎么迁移到你自己的写作和评审流程中。读完本文，你应当能独立识别诸如 throat-clearing 开场、被动语态假主语、"尽管存在挑战"式转折等常见 AI 文体信号，并按照示例给出的手法把它们改写成具体、直接、有人味的文字。

一、examples.md 在 deslop 技能中的位置

deslop 技能的目录结构（见 README.md）如下：

deslop/ ├── SKILL.md # 核心规则、快速检查项、评分标准 ├── README.md # 使用说明 └── references/ ├── phrases.md # 应删除或替换的短语目录 ├── structures.md # 应避免的结构性模式 ├── tropes.md # AI 写作套路完整目录 └── examples.md # Before/After 改写示例（本文主题）

SKILL.md 在 "Reference Files" 一节明确将references/examples.md定位为 "Before/after transformations showing how to fix common patterns"（展示如何修复常见模式的改写前后对照）。也就是说，phrases.md、structures.md、tropes.md 负责"给出清单"，而 examples.md 负责"示范动手"。文档中的 15 个示例按文体标注了适用场景（Scientific、Cover Letter、Blog Post、Grant Narrative、General、Abstract、Discussion Section），每个示例统一采用三段式结构：

• Before：包含 AI 套路的原始句子；
• After：按 deslop 规则改写后的版本；
• Changes：一句话说明做了哪几处修改、分别对应哪条规则。

这种"对照 + 修改说明"的格式本身就是可复用的：无论是人工改稿，还是让 Agent 按 README.md 中描述的方式触发技能（例如说 "deslop this text"、"de-AI"、"make it sound human"、"check for slop"），示例都提供了可直接模仿的改写模板。

下面按缺陷类型把 15 个示例重新组织成 6 组讲解，原文示例编号（Example 1~15）全部保留，便于回查原文。

二、开场与修辞问题：去掉 throat-clearing、反问与自我设问

这一组对应 SKILL.md 核心规则 1（Cut filler phrases）和 Quick Checks 中的 "Any 'here's what/this/that' throat-clearing? Cut to the point"。

Example 1：Throat-Clearing + Binary Contrast（Scientific）

Before：

"Here's the thing: forecasting infectious disease is hard. Not because the models are complex. Because the data is complex. Let that sink in."

After：

"Forecasting infectious disease is hard. The models are tractable. The data, collected under shifting surveillance definitions and reporting lags, is not."

Changes：删除了开场白（"Here's the thing:"）、二元对比结构（"Not because X. Because Y."）和强调拐杖（"Let that sink in."），并把数据问题具体化（点名了"不断变化的监测口径和报告延迟"）。

这个示例同时修掉了 phrases.md 中 "Throat-Clearing Openers" 与 "Emphasis Crutches" 两节所列的条目，以及 structures.md 中被标记为"最常被识别的 AI 写作信号"的二元对比（"Not because X. Because Y."）。

Example 5：Self-Posed Rhetorical Question（Blog Post）

Before：

"What if I told you that most bioinformatics pipelines break in production? The result? Wasted compute and silent errors. The worst part? Nobody checks the intermediate outputs. Here's why that matters:"

After：

"Most bioinformatics pipelines break in production, and the failures are quiet. The FASTQ passes QC. The alignment runs. But the reference genome version changed between runs, and the variant calls shift without warning."

Changes：删除了反问铺垫和三个自问自答句，替换成一个读者可以在脑中具体还原的故障场景。

对应 SKILL.md Quick Checks 中 "Any self-posed rhetorical question answered immediately? Fold into a statement."，即自问自答应折叠成陈述句。

Example 13：False Vulnerability + Meta-Commentary（Blog Post）

Before：

"And yes, since we're being honest: I've run plenty of analyses where the p-value was borderline and I squinted at it until it cooperated. I want to explore why that impulse is so common. In this post, I'll walk you through what I've learned."

After：

"I've nudged a borderline p-value along by trying one more covariate. You probably have too. The question is what makes that feel acceptable in the moment, and the answer is usually that the rest of the analysis already 'looks right.'"

Changes：用具体的、诚实的承认替代了表演式的"虚假脆弱"（"squinted at it until it cooperated"），删掉元评论（"In this post, I'll walk you through"），直接陈述观点而不是宣布自己要讲什么。

这里对应两条规则：phrases.md 中 "False Vulnerability" 一节的 "And yes, since we're being honest..." 原样条目，以及 "Meta-Commentary" 一节的 "Let me walk you through..."。

Example 15："Imagine a World" + Patronizing Analogy（General）

Before：

"Imagine a world where every meeting had a clear agenda. Think of it like a recipe: you wouldn't start cooking without knowing the ingredients. That's the promise of async-first communication. Let's unpack why this matters."

After：

"Meetings without agendas waste time. A 15-person sync with no written agenda averages 47 minutes and produces no decisions (Atlassian, 2019). Writing the agenda forces the organizer to decide whether the meeting is necessary at all."

Changes：删除 "imagine" 式开场、没有信息量的烹饪类比，以及教学腔的 "let's unpack"；替换为一个具体论断、一个数字，以及议程为何有效的机制说明。

对应 phrases.md "Pedagogical Hand-Holding" 一节的 "Imagine a world where..."、"Think of it like..."、"Let's unpack this" 三条原样条目。

三、填充语与公式化转折："Despite These Challenges" 家族

Example 2：Filler + "Despite These Challenges"（Cover Letter）

Before：

"It's worth noting that these findings have important implications for how we navigate the challenges of forecast ensembling moving forward. Despite these challenges, this work contributes meaningfully to the growing body of literature, highlighting the need for continued evaluation and underscoring the importance of robust benchmarking."

After：

"If individual model rankings are unstable across geography and time, ensemble methods that weight models by past performance may not improve on equal-weight approaches."

Changes：用研究发现的具体含义替换了填充式转折（"It's worth noting"）、模糊宣言（"have important implications"）、"despite these challenges" 公式句，以及两个浅层分词短语（"highlighting the need for..."、"underscoring the importance of..."）。

这一个示例几乎是一次集齐：phrases.md 的 "It's worth noting"、"Moving forward"、"Vague Declaratives"，structures.md 的 "Despite Its Challenges..." 与 "Superficial Participle Analyses"。值得注意的是，SKILL.md 末尾的 Quick inline example（scientific writing）用的正是同一组 Before/After 文本，说明这是技能作者认为最典型的一组信号。

Example 14："It's Worth Noting" + Invented Concept Label（Scientific）

Before：

"It's worth noting that this creates what might be called the 'calibration paradox': models that are well-calibrated at the national level may be poorly calibrated at the state level, reflecting broader trends in the tension between aggregation and granularity."

After：

"National-level calibration does not guarantee state-level calibration. A model can produce well-calibrated 90% intervals for the US overall while consistently undercovering in states with smaller populations and noisier surveillance data."

Changes：删掉填充式转折和自造概念标签（"calibration paradox"），用具体机制（人口规模更小、监测数据更噪的州）替换浅层分词分析（"reflecting broader trends in..."）。

"自造概念标签"在 tropes.md 的 "Invented Concept Labels" 一节有完整说明：AI 喜欢把抽象问题名词（paradox、trap、creep、divide、vacuum、inversion）拼接到领域词上，当成既定术语使用；修复原则是"如果概念需要名字就先定义它，不需要名字就用平实语言描述"。

四、宏大叙事与"景深"式夸张：把 stakes 拉回实际尺度

Example 3：Grandiose Stakes + Landscape（Scientific）

Before：

"In today's rapidly evolving genomic landscape, single-cell RNA sequencing has fundamentally reshaped how we think about cellular heterogeneity. This paradigm shift has far-reaching implications for our understanding of disease."

After：

"Single-cell RNA sequencing reveals cell-type-specific expression patterns that bulk methods average out. In tumor samples, this distinction matters: rare resistant subpopulations visible in single-cell data disappear in bulk profiles."

Changes：消除 "landscape"、"paradigm shift"、"fundamentally" 和模糊的 stakes 论断，替换成一个具体例子说明该方法为何重要。

对应 phrases.md 三处条目："In today's [X]" 填充短语、Business Jargon 表中 "Landscape → Situation, field"、"Paradigm → Model, approach"，以及 tropes.md "Grandiose Stakes Inflation" 与 "Tapestry and Landscape" 两节。

Example 10：Historical Analogy Stacking（General）

Before：

"Apple didn't build Uber. Facebook didn't build Spotify. Stripe didn't build Shopify. AWS didn't build Airbnb. Every major technological shift follows the same pattern."

After：

"AWS sold infrastructure. Airbnb built a business on top of it. That split between platform and product repeats across the industry."

Changes：用一次深挖的具体案例替代四个名字堆砌，并点名具体模式（platform 与 product 的分工）。

structures.md 对这类"历史类比堆砌"的修复建议是 "Use one example, examine it in depth. One well-analyzed case beats five name-drops."；SKILL.md 核心规则 10（Do not dilute）同样禁止 "stack historical analogies for false authority"。

五、主语与视角：被动语态、假主语与远距离叙述者

这一组对应 SKILL.md 核心规则 4（Use active voice with human subjects）。

Example 4：Passive Voice + False Agency（Discussion Section）

Before：

"It was observed that model performance degraded at longer forecast horizons. The uncertainty naturally increased as the prediction window expanded. These results emerged from our analysis of 54 state-level forecasts."

After：

"We observed that model performance degraded at longer forecast horizons. Each additional week of lead time added roughly 15% to the mean WIS. We saw this pattern across all 54 state-level forecasts."

Changes：点名行为主体（"we"）；把假主语表达（"uncertainty naturally increased"、"results emerged"）替换为具体论断并给出数字。

"假主语"（False Agency）在 structures.md 中列为独立小节："Giving inanimate things human verbs. AI loves this because it avoids naming the actor."，修复原则是 "Name the human"；学术论文中若没有具体的人，用 "we"。

Example 7：Narrator Distance + Vague Declarative（Blog Post）

Before：

"Nobody designed the current state of genomic data sharing. The reasons are structural. This is genuinely hard, and the implications are significant."

After：

"You inherit a genomic data sharing system that grew by accretion: dbGaP for controlled access, GEO for expression, SRA for raw reads, ClinVar for variants. Each solves one problem. None talks to the others without custom glue code."

Changes：用 "you" 替代了远距离叙述者（"Nobody designed..."），用具体的结构性问题（逐个列出各库的分工与互不联通）替换了两个模糊宣言。

对应 structures.md "Narrator-from-a-Distance" 小节（"Nobody designed this." 是该表中的原样条目）和 phrases.md "Vague Declaratives" 小节（"The reasons are structural"、"The implications are significant" 均为表内原样条目）。SKILL.md 核心规则 6（Match register to context）进一步区分了语境：博客/通讯中 "You" 优于 "People"；科学写作中用 "we" 指代自己的工作、引用具体作者而不是 "researchers have shown"。

六、结构与节奏：列表化长文、稀释、头韵与断句表演

Example 8：Listicle in a Trench Coat（Discussion Section）

Before：

"The first limitation of our study is the relatively short evaluation window. The second limitation is the lack of a standardized baseline across all forecasters. The third limitation is the potential for overfitting to the specific characteristics of the 2021-22 and 2022-23 seasons."

After：

"Our evaluation covers only two seasons, and both followed the unusual post-COVID influenza dynamics that may not generalize. We also lacked a standardized baseline: each forecaster chose their own training data and feature set, making apples-to-apples comparison difficult."

Changes：把编号式列表结构改写成连贯散文，合并相关要点，并补充了局限为何重要的具体说明。

structures.md 对该模式的修复建议是二选一："If the content is a list, present it as a list. If it should be prose, weave the points together without numbering."

Example 9：One-Point Dilution（Blog Post）

Before：

"Reproducibility matters. In fact, reproducibility is the most underrated quality in computational biology. When you make your analysis reproducible, you reduce errors. Fewer errors mean more trust. More trust means faster adoption. The lesson? Make it reproducible. Irreproducible analyses are the enemy of science. The best papers are the reproducible ones."

After：

"Reproducible analyses catch errors earlier. When Beaulieu-Jones and Greene re-ran 68 papers from the PharmacoGenomics Knowledge Base, only 40% reproduced. The ones that failed had no shared code or pinned dependency versions."

Changes：八句重复同一个论点的文字压缩为三句，并附带一个具体支撑案例和一个数字。

对应 SKILL.md 核心规则 10（Do not dilute："One point per section. Do not restate the same argument in ten different ways"）和 tropes.md "One-Point Dilution" 一节（"An 800-word argument becomes 4000 words of circular repetition"）。

Example 11：Anaphora Abuse（Grant Narrative）

Before：

"We will develop novel computational methods. We will apply these methods to large-scale genomic datasets. We will validate our findings using independent cohorts. We will disseminate our tools through open-source repositories. We will train the next generation of computational biologists."

After：

"We will develop and validate statistical methods for multi-ancestry fine-mapping using UK Biobank and TOPMed cohorts, then release them as an R package with documentation and tutorials suitable for graduate training."

Changes：把五句头韵式排比压缩成一句，点名了具体方法、数据集和交付物。

Example 12：Dramatic Fragmentation（General）

Before：

"Speed. Quality. Cost. You can only pick two. That's it. That's the tradeoff."

After：

"Speed, quality, cost: pick two."

Changes：单句搞定。没有表演性强调。

对应 structures.md "Dramatic Fragmentation" 小节。该小节给出了这一模式的成因解释：RLHF 训练把模型推向了"为最低阅读门槛写作"的方向，一句一个念头、不需要读者做任何心智维护；修复原则是 "Complete sentences. Trust content over presentation."

Example 6："Serves As" + Superficial Participle Analysis（Abstract）

Before：

"The FluSight initiative serves as a foundational framework for influenza forecasting in the United States, contributing to public health preparedness and underscoring the importance of collaborative forecasting efforts."

After：

"The FluSight initiative coordinates influenza forecasting across dozens of modeling groups in the United States. Since 2013, it has standardized targets, submission formats, and evaluation metrics."

Changes：用 FluSight 实际做什么替换了 "serves as a foundational framework"，用一个具体事实替换了两个分词短语。

对应 phrases.md "The 'Serves As' Dodge" 表格（serves as / stands as / marks / represents → is）和 structures.md "Superficial Participle Analyses"（修复原则："Either make a specific analytical claim or delete the participle phrase."）。

七、模式索引：15 个示例映射到 deslop 规则体系

把 15 个示例映射回 SKILL.md 的 10 条核心规则，可以得到这样一张索引表：

示例中反复出现的修复手法可以归纳为三类：一是点名主体与数字（Example 4 的 "we observed"、Example 14 的 "90% intervals"、Example 9 的 "68 papers, 40% reproduced"）；二是把抽象名词换成机制（Example 3 的 "bulk methods average out"、Example 7 逐个列出 dbGaP/GEO/SRA/ClinVar）；三是删除铺垫、直陈要点（Example 2、12）。这三类手法与 phrases.md、structures.md 中每个模式的 "Fix" 建议一一对应。

八、评分与落地：如何在 open-seo 仓库中触发这套规则

deslop 技能的 README 给出了评分标准：交付前对文本按五个维度各打 1-10 分，低于 35/50 就回去修改：

仓库内实际调用这个技能的入口是 openseo-review-web-content 技能：它的第 4 条规则 "Sound like a person" 明确指向 deslop——"The deslop skill is the reference for what to hunt and how to fix it"，并要求把 deslop 评审作为子 Agent 的一路（voice/deslop）来跑，产出精确的 old → new 修改提案。也就是说，examples.md 中的 15 组 Before/After 不只是教学材料，而是 open-seo 网站内容（web/content下的博客与营销页）在评审流程中实际执行的改写基准。

触发方式（摘自 README.md）包括自然语言指令，例如 "Review this draft using the deslop checklist"、"Score this text on the deslop rubric"、"Rewrite this paragraph to pass the deslop quick checks"；技能也可以在用户要求 "deslop"、"de-AI"、"make it sound human"、"check for slop" 时自动触发。技能声明其内容来自两个开源素材的整合：tropes.fyi 的 AI 写作套路目录（对应references/tropes.md）与 stop-slop 项目的短语清单、结构模式和评分标准，许可证为 MIT（见 LICENSE）。

九、把示例库用于自己的稿件：一份可复制的工作流

结合本文前面的拆解，评审任意一篇稿件时可以按以下顺序执行，每一步都有对应的仓库文件可查：

1. 过 Quick Checks 清单：SKILL.md 的 "Quick Checks" 一节约 20 条二元问题（副词是否过多、有无被动语态、有无 "here's what" 式开场、有无 "not X, it's Y" 对比、连续三句是否等长、有无 em dash、有无 "It's worth noting" 等），逐条打勾。
2. 对号入座到具体模式：命中的信号在 phrases.md 找短语条目，在 structures.md 找结构条目，在 tropes.md 找语体与排版条目；三个文件都给出 "Fix" 建议。
3. 参照最近的 Before/After 示例改写：本文第二节到第六节的 15 组示例覆盖了从科学摘要到基金申请书的各类文体，选择文体最接近的一组作为改写模板。
4. 按五维评分表复核：Directness、Rhythm、Trust、Authenticity、Density 各 1-10 分，低于 35/50 回到第 1 步。
5. 注意文体匹配：科学写作允许并期望领域术语（SKILL.md 举例："Weighted interval score" 是精确语言而非黑话），真正要清理的是商业词汇（"leverage"、"landscape"、"ecosystem"）和 AI 词汇信号（"delve"、"tapestry"、"nuanced"）混入技术文本。

deslop 技能的适用边界也值得说明：它处理的是"文体信号"，不替代事实核查。open-seo 的评审技能明确要求 deslop 一路（voice/deslop）与 claims accuracy 一路并行，且每条修改提案都要先对照实际文件与代码验证后再应用——改写手法再标准，也不能拿它去修正事实错误。

参考文件

• Before/After 示例原文（本文主体，15 组示例）
• deslop 核心规则与 Quick Checks
• 技能使用说明与评分标准
• 短语目录
• 结构模式目录
• AI 写作套路完整目录
• 仓库内实际调用 deslop 的评审技能
• 许可证

【免费下载链接】open-seoOpen source alternative to Semrush and Ahrefs项目地址: https://gitcode.com/GitHub_Trending/op/open-seo