Back to Details

anti-slop

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Anti-Slop: Writing and Reviewing Human-Sounding Prose

反冗余套话：撰写与审阅类人风格的散文

"Slop" is phrasing that signals machine authorship and erodes a reader's trust: filler that announces itself, hedged puffery, formulaic rhythm, and clever-sounding constructions that carry no information. The goal is not to strip personality. It is to remove the patterns that make a reader think "a model wrote this" so the ideas land.
所谓“冗余套话”（slop）是指暴露机器写作身份的措辞，会削弱读者的信任：这类内容是刻意添加的填充语、含糊的浮夸表达、刻板的节奏，以及看似巧妙实则毫无信息量的句式。我们的目标并非抹除个性，而是去除那些让读者觉得“这是AI写的”的模式，让核心观点能够有效传递。

Prime directive: preserve voice

核心准则：保留作者语气

The most common failure of an anti-slop pass is over-correction into flat, voiceless text. That is its own tell. Before deleting anything, decide whether it is slop (adds no meaning, follows a formula) or voice (a deliberate choice the author would defend).
  • Cut filler, formula, and hedging.
  • Keep rhythm the author earns: a short punchy fragment after a long sentence, a deliberate tricolon, a load-bearing contrast, a strong closing line.
  • When unsure, leave it. A false positive that flattens a good sentence is worse than one surviving tell.
Make surgical edits. Do not restructure paragraphs, reorder arguments, or change meaning unless asked. Change phrasing, not substance.
反冗余套话润色最常见的失误是过度修正，导致文字变得平淡无味、毫无个性——这本身也是一种AI写作痕迹。在删除任何内容之前，先判断它是冗余套话（无意义、遵循固定模板）还是作者语气（作者会为之辩护的刻意选择）。
  • 删除填充语、刻板模板和含糊措辞。
  • 保留作者刻意营造的节奏：长句之后的简短有力片段、刻意使用的三重排比、承载核心对比的表达、有力的收尾句。
  • 若拿不准，就保留原样。误删一个好句子导致文字平淡，比留下一处AI痕迹更糟糕。
进行精准修改。除非被要求，否则不要重组段落、调整论点顺序或改变原意。只修改措辞，不改动核心内容。

Workflow

工作流程

When reviewing (yours or someone else's)

审阅时（自己或他人的作品）

  1. Read the whole piece first. Note voice: is it punchy, formal, conversational? You are matching this, not overwriting it. 2. Scan for the tells in the catalog below. Collect candidates — do not edit yet. 3. Validate each candidate: does removing it lose meaning? Is it formula or voice? Discard false positives. 4. Apply the surviving edits one at a time, each a minimal phrasing change. 5. Re-read the edited passages for rhythm. Fix anything that now reads choppy or flat. 6. Report what changed and the tell category for each, so the author can judge.
  1. 先通读全文。留意作者语气：是简洁有力、正式严谨还是口语化？你要匹配这种语气，而非改写它。2. 扫描下文列出的AI痕迹，收集候选修改项——暂不编辑。3. 验证每个候选项：删除它会丢失信息吗？它是模板化内容还是作者语气？剔除误判项。4. 逐一应用最终确定的修改，每项都只做最小幅度的措辞调整。5. 重读修改后的段落，检查节奏。修正任何变得生硬或平淡的内容。6. 告知作者修改内容及对应的AI痕迹类别，供作者判断。

When drafting

撰写时

Write first, then run the review workflow on your own draft before delivering. Do not self-censor into blandness while drafting; catch tells on the edit pass.
先完成初稿，再在交付前用上述审阅流程检查自己的草稿。撰写时不要自我审查导致文字乏味；在编辑阶段再排查AI痕迹。

The tell catalog

AI痕迹清单

TellWhy it erodes trustFix
Meta throat-clearing — announcing the point before making it: "It is worth noting that…", "There is a failure mode worth naming…", "Here is the tension."Wastes the reader's time; signals padding.Delete the frame. State the thing.
Demonstrative kicker — a vague "This/That + verdict" fragment tacked after a sentence: "That instinct backfires.", "This is where the risk hides.", "That is the whole point."Formulaic filler rhythm; the "verdict" usually restates the prior sentence.Cut it, or fold the idea into the sentence; pivot the next sentence with a real transition ("But…").
Puffery adverbs — genuinely, truly, actually, simply, deeply, structurally, fundamentally.Intensifiers that add heat, not light.Delete. If the claim needs the adverb to be true, the claim is weak.
Importance-flagging — "Speed is not a footnote here.", "This matters.", "Make no mistake."Tells the reader what to feel instead of earning it.Show the consequence directly.
Clever metaphor flourish — "an authorization bug wearing the costume of a performance optimization."Reads as a model performing wit.State it plainly: "an authorization bug, not a performance optimization."
Grandiose prediction — "will define the next decade", "changes everything", "the future of X".Overclaiming; lowers credibility in serious writing.Cut or scope to a concrete, defensible claim.
Rule-of-three reflex — every list padded to three parallel items.Predictable cadence; the third item is often filler.Vary list length. Keep two if two is true.
Antithesis overuse — "not X, but Y" as a tic in paragraph after paragraph.One is rhetoric; five is a template.Keep the load-bearing one; rewrite the rest as plain statements.
Correlative bloat — "not only… but also", "whether… or".Scaffolding that inflates simple sentences.Split or simplify.
Hedged confidence — "It's important to consider…", "One might argue…", "In many ways…".Sounds authoritative while committing to nothing.Take the position or cut the sentence.
Section-closing summary — a final sentence that restates what the paragraph just said.Redundant; a model habit.Delete if it adds nothing.
Overused lexicon — delve, tapestry, realm, landscape, underscore, leverage, seamless, robust, crucial, pivotal, testament, navigate (figurative), foster, elevate, unlock.Statistically flagged AI vocabulary.Swap for plain words or cut. See references/tells.md.
痕迹类型为何会削弱信任修改方法
元语铺垫——在提出观点前先宣告：“值得注意的是……”、“有一种值得明确指出的失效模式……”、“这就是矛盾所在。”浪费读者时间，属于凑字数的填充内容。删除铺垫框架，直接陈述观点。
指代式收尾——在句子末尾附加模糊的“这/那+结论”片段：“这种本能会适得其反。”、“风险就藏在这里。”、“这就是核心要点。”刻板的填充节奏，“结论”通常只是重复前一句的内容。删除该片段，或将观点融入前一句；用真实的过渡语（如“但……”）开启下一句。
浮夸副词——genuinely、truly、actually、simply、deeply、structurally、fundamentally等这类强化词只增加情绪渲染，不传递实际信息。删除。如果论点需要依赖这类副词才能成立，说明论点本身站不住脚。
强调重要性——“速度在这里绝非无关紧要。”、“这很重要。”、“别搞错。”试图告诉读者该有什么感受，而非通过内容让读者自然产生感受。直接展示后果。
炫技式隐喻——“披着性能优化外衣的授权漏洞。”读起来像是AI在刻意表现机智。直白陈述：“这是授权漏洞，而非性能优化。”
宏大预测——“将定义下一个十年”、“彻底改变一切”、“X的未来”过度夸大，降低严肃写作的可信度。删除，或调整为具体、可辩护的论点。
三重罗列本能——每个列表都硬凑成三个平行项可预测的节奏，第三个条目通常是填充内容。改变列表长度。如果只有两个真实条目，就保留两个。
过度使用对偶——段落中反复出现“不是X，而是Y”的句式用一次是修辞，用五次就是模板。保留承载核心意义的那一处，其余改写为直白陈述。
关联式冗余——“不仅……而且……”、“无论……还是……”这类结构会把简单句子变得臃肿。拆分或简化句子。
含糊的自信——“值得考虑的是……”、“有人可能会认为……”、“在很多方面……”听起来权威，但实际上没有任何明确立场。明确表明立场，或删除该句子。
段落收尾总结——最后一句重复本段刚说过的内容冗余，是AI的写作习惯。如果没有新增信息，就删除。
过度使用的词汇——delve、tapestry、realm、landscape、underscore、leverage、seamless、robust、crucial、pivotal、testament、navigate（比喻义）、foster、elevate、unlock等这类词汇被统计标记为AI高频词汇。替换为直白词汇或删除。详见references/tells.md。

Concrete before / after

具体修改示例

Meta throat-clearing
  • Before: "There is a failure mode worth naming directly, because well-intentioned teams walk straight into it. It is the belief that safety means a human approving every step."
  • After: "Many teams equate safety with a human approving every step."
Demonstrative kicker
  • Before: "…reuse the machinery we already have. That instinct is where most of the risk hides. Four properties break the assumptions."
  • After: "…reuse the machinery we already have. But four properties break the assumptions."
Importance-flagging + staccato
  • Before: "Speed is not a footnote here. It changes which controls are viable. Anything that depends on a person noticing has already lost."
  • After: "Speed changes which controls are viable: anything that depends on a person noticing has already lost."
Clever flourish
  • Before: "an authorization bug wearing the costume of a performance optimization"
  • After: "an authorization bug, not a performance optimization"
元语铺垫
  • 修改前：“有一种值得明确指出的失效模式，因为初衷良好的团队会直接陷入其中。这种模式就是认为安全意味着人类要批准每一步操作。”
  • 修改后：“很多团队将安全等同于人类批准每一步操作。”
指代式收尾
  • 修改前：“……复用我们已有的机制。这种本能就是大部分风险的来源。有四个特性打破了原有假设。”
  • 修改后：“……复用我们已有的机制。但有四个特性打破了原有假设。”
强调重要性+断句
  • 修改前：“速度在这里绝非无关紧要。它会改变可行的控制措施。任何依赖人工发现的方案都已经失效。”
  • 修改后：“速度会改变可行的控制措施：任何依赖人工发现的方案都已经失效。”
炫技式表达
  • 修改前：“披着性能优化外衣的授权漏洞”
  • 修改后：“这是授权漏洞，而非性能优化。”

What to keep (do not "fix" these)

需保留的内容（不要“修正”这些）

These are voice, not slop. Deleting them is the over-correction failure.
  • Earned fragments: "Not flagged after the fact. Removed." — a short beat after a long sentence, used sparingly, for emphasis.
  • Deliberate parallelism: "You can have either. You cannot have both."
  • A strong closing line: "The perimeter fell twelve years ago. The session is next."
  • First-person conviction: "We think…", "We are not comfortable saying…" when the author owns a real position.
The test: could the author defend this choice if asked? If yes, keep it. Slop cannot be defended — it is there because a pattern put it there.
这些属于作者语气，而非冗余套话。删除它们属于过度修正的失误。
  • 刻意使用的短句：“事后未标记。已移除。”——长句之后的简短停顿，偶尔使用以强调重点。
  • 刻意的平行结构：“你可以选其中一个。但不能两者都要。”
  • 有力的收尾句：“防线在十二年前就已失守。下一场会议即将召开。”
  • 第一人称明确立场：“我们认为……”、“我们不认同……”当作者明确表达真实立场时。
判断标准：如果被问起，作者能否为这个选择辩护？如果可以，就保留。冗余套话无法辩护——它的存在只是因为AI的模板化模式。

Guardrails

约束规则

  • Never change facts, numbers, names, citations, or technical claims during a phrasing pass.
  • Never restructure or re-argue. Phrasing only, unless explicitly asked for more.
  • Validate every candidate before editing; report false positives you chose to keep and why.
  • Do not introduce new slop while removing old (e.g. replacing a kicker with a different formulaic kicker).
  • One tell fixed cleanly beats three fixed clumsily. Stop when the remaining candidates are voice, not slop.
  • 在措辞修改过程中，绝不能更改事实、数字、姓名、引用或技术主张。
  • 绝不能重组结构或重新论证。除非明确要求更多修改，否则只调整措辞。
  • 编辑前验证每个候选项；告知作者你选择保留的误判项及原因。
  • 去除旧套话时不要引入新的套话（例如用另一种模板化收尾替换原来的收尾）。
  • 干净修复一处痕迹，胜过笨拙修复三处。当剩余候选项属于作者语气而非冗余套话时，就停止修改。

Review checklist

审阅清单


- [ ] Read whole piece; noted the author's voice
- [ ] Collected tell candidates without editing
- [ ] Validated each: slop vs voice; discarded false positives
- [ ] Applied surviving edits as minimal phrasing changes
- [ ] Re-read for rhythm; fixed new choppiness/flatness
- [ ] Facts, numbers, structure untouched
- [ ] Reported changes + tell category per edit


- [ ] 通读全文；记录作者的语气风格
- [ ] 收集AI痕迹候选项，暂不编辑
- [ ] 验证每个候选项：是冗余套话还是作者语气；剔除误判项
- [ ] 应用最终确定的修改，仅做最小幅度的措辞调整
- [ ] 重读修改内容，检查节奏；修正新出现的生硬/平淡问题
- [ ] 事实、数字、结构未被改动
- [ ] 告知修改内容及每项对应的AI痕迹类别