Back to Details

no-slop

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

No-Slop: write clean, audit hard

No-Slop：写得干净，审得严格

Two failure modes kill a piece of writing. It can be false (a quote that was never said, a borrowed idea claimed as original), or it can be slop (generic AI texture nobody finishes, trusts, or shares). This skill guards against both, in a fixed order: fidelity first, taste second. When polish ever pulls against truth, fidelity wins. A shareable lie is still a lie. Sharpen the framing, never the facts.
两种失败模式会毁掉一篇作品。它可能失实（一句从未有人说过的引语，或把借来的观点当作原创），也可能是低质套话（泛化的 AI 腔调，没人愿意读完、信任或分享）。这个 skill 按固定顺序同时防范两者：先看来源忠实度，再看品味。当润色与真实发生冲突时，来源忠实度优先。可分享的谎言仍然是谎言。锐化表达框架，绝不改写事实。

Two modes

两种模式

  • Write mode: drafting or rewriting. Apply the style rules below from the first sentence, not as a cleanup step; prose written slop-first and scrubbed later keeps the skeleton of slop. Before delivering, run both gates on your own draft and fix what fails.
  • Audit mode: the user hands you a draft. Run Gate 1, then Gate 2, and return the audit report (format below). Apply mechanical fixes when asked; flag judgment calls (an unverifiable-but-probably-real quote) for the user, who often knows the sources best.
This skill governs texture and truth, not voice, and composes with any voice skill. One precedence rule: the punctuation bans stay in force even under a voice skill, unless the user explicitly opts out.
  • 写作模式：起草或改写。从第一句话起就应用下述风格规则，不要把它当成事后清理步骤；先以低质套话起稿再事后清洗，文字仍会保留套话的骨架。交付前，对自己的草稿运行两道关卡，并修复所有未通过项。
  • 审计模式：用户交来草稿。运行关卡 1，再运行关卡 2，并返回审计报告（格式见下文）。用户要求时执行机械性修正；对于需要判断的情况（例如一句无法核验但很可能是真实的引语），将其标出交给用户判断，用户通常最了解来源。
这个 skill 约束的是文字质感和真实性，而不是声音风格，并且可与任何声音风格 skill 配合。只有一条优先级规则：即使启用声音风格 skill，标点禁用规则仍然有效，除非用户明确选择退出。

The style rules (always on)

风格规则（始终生效）

Banned: zero tolerance

禁用项：零容忍

One tell might survive review; three brand the piece as generated. Do not write:
  • em dashes or double-hyphen separators, in any form (use periods, colons, commas, parentheses)
  • "In today's rapidly evolving landscape..." and every cousin of it
  • "This chapter/article/post explores..."
  • "It is important to note..." / "It's worth noting..." / "We're excited to announce..."
  • the "not X, but Y" cadence, and its staccato variant "Not X. It is Y."
  • "That is..." as a sentence opener; "By understanding X, readers can..."
  • hedges: "arguably", "in many ways", "to some extent", "one could argue"
  • recap markers: "In short", "Ultimately", "At the end of the day", "In conclusion"
  • filler: 
    delve
    , 
    leverage
    (verb), 
    unpack
    , 
    seamless
    , 
    game-changer
    , 
    deep dive
    , 
    empower
    , 
    elevate
    , 
    supercharge
    , 
    unleash
  • filler only in its filler frame, literal use is fine: 
    unlock the potential
    (not a door), 
    robust framework
    (not robust standard errors), 
    the learning journey
    (not a journey home), 
    harness the power of
    (not harness the energy of the jet stream). Cutting the literal use is the over-correction this skill exists to prevent
  • stock flourishes: 
    a testament to
    , 
    plays a crucial role
    and every adjective in that slot, 
    navigate the complexities
    , 
    in an era of
    , 
    in a world where
    , 
    whether you're a...
    , 
    it's no secret
    , 
    look no further
    , 
    let's dive in
    , 
    treasure trove
    , 
    a beacon of
    , 
    tapestry of
    , 
    in the realm of
  • performed candor as filler: 
    let's be honest
    , 
    let's be real
    , 
    let's face it
    , 
    real talk
    , 
    here's the thing
    , 
    here's the kicker
    , 
    here's the catch
    (these are also injection failures; see the rewrite constraints). 
    let us be clear
    is formal register, not this
  • perfectly symmetrical sections; a generic recap after every section
  • fake-deep abstractions with no quote or case behind them; strategy-deck phrasing
When you feel the pull to smooth, resolve, or summarize, that is the instinct talking, not the reader's need. 
references/taste-gate.md
explains why each pressure exists, which is what lets you catch a tell the list does not name.
一个破绽可能侥幸通过审查；三个则会让作品带上生成痕迹。不要写：
  • 任何形式的 em dash 或双连字符分隔符（改用句号、冒号、逗号、括号）
  • In today's rapidly evolving landscape...
    及其所有同类变体
  • This chapter/article/post explores...
  • It is important to note...
    / 
    It's worth noting...
    / 
    We're excited to announce...
  • not X, but Y
    句式，以及其短促变体 
    Not X. It is Y.
  • 以 
    That is...
    开头的句子；
    By understanding X, readers can...
  • 模糊限定语：
    arguably
    、
    in many ways
    、
    to some extent
    、
    one could argue
  • 总结提示语：
    In short
    、
    Ultimately
    、
    At the end of the day
    、
    In conclusion
  • 填充词：
    delve
    、
    leverage
    （动词）、
    unpack
    、
    seamless
    、
    game-changer
    、
    deep dive
    、
    empower
    、
    elevate
    、
    supercharge
    、
    unleash
  • 填充词仅在填充语境中禁用，字面用法没问题：
    unlock the potential
    （不是指打开门）、
    robust framework
    （不是指稳健的标准误）、
    the learning journey
    （不是指回家之旅）、
    harness the power of
    （不是指利用喷流能量）。把字面用法也删掉属于过度纠偏，正是这个 skill 要防止的问题
  • 陈腐修辞：
    a testament to
    、
    plays a crucial role
    以及该位置上的所有形容词、
    navigate the complexities
    、
    in an era of
    、
    in a world where
    、
    whether you're a...
    、
    it's no secret
    、
    look no further
    、
    let's dive in
    、
    treasure trove
    、
    a beacon of
    、
    tapestry of
    、
    in the realm of
  • 充当填充语的表演式坦诚：
    let's be honest
    、
    let's be real
    、
    let's face it
    、
    real talk
    、
    here's the thing
    、
    here's the kicker
    、
    here's the catch
    （这些也是注入失败项，见改写约束）。
    let us be clear
    属于正式语域，不属于这一类
  • 完全对称的章节结构；每节结尾都有一段泛泛总结
  • 没有引语或案例支撑的假深刻抽象表述；战略演示文稿式措辞
当你感到想要把内容抹平、圆化或总结时，那是本能在作祟，不是读者需要。
references/taste-gate.md
解释了每种压力为何存在，这能让你识别清单未列出的破绽。

Narrative tells (fiction and story-driven prose)

叙事破绽（小说和故事型文字）

Recycled emotional choreography (
breath catching
, 
jaw clenching
, 
heart hammering
); named emotions ("she felt a surge of determination") instead of the action the feeling produces; tidy-summary endings ("For the first time, I understood...") instead of ending on action or image; sentiment skew (dark scenes silver-lined, tension resolved in the paragraph that raised it); one arc stamp on every scene; cluster metaphors (weight, light, drowning); zero subtext. The scanner catches the first three; the rest need judgment.
循环套用的情绪编排（
breath catching
、
jaw clenching
、
heart hammering
）；直接说出情绪（
she felt a surge of determination
），而不是描写情绪引发的行动；工整的总结式结尾（
For the first time, I understood...
），而不是以动作或画面收尾；情绪走向偏斜（黑暗场景总被涂上亮色，紧张感在制造它的同一段里就消解）；每个场景都套用同一种人物弧线；同类隐喻扎堆（重量、光、溺水）；完全没有潜台词。
扫描器能抓到前三种，其余需要判断。

Prefer

优先采用

  • short claims
  • primary-source-led sections
  • concrete consequences: name the market, the number, the person
  • cases before abstractions
  • blunt operator sentences
  • fewer transitions
  • compression over explanation
  • ambiguity where reality is ambiguous
The style test: if a paragraph could appear in a generic business ebook, cut or rewrite it. The repetition budget: rotate examples, include an instructive failure, and never let three sections share one skeleton.
  • 简短有力的主张
  • 以一手来源为主线的章节
  • 具体后果：点名市场、数字、人物
  • 先案例，后抽象
  • 像一线操盘手一样直白的句子
  • 更少的过渡
  • 宁可压缩，不要解释
  • 现实本身模糊时，保留模糊性
风格测试： 如果某段话可以原封不动地放进一本泛泛的商业电子书，就删掉或重写。 重复预算： 轮换例子，加入一个有启发性的失败案例，绝不让三个章节共用同一种骨架。

Rewrite constraints (govern what you produce)

改写约束（约束你的产出）

Gate 1 audits the draft you were handed. These govern the text you write, and no scanner can enforce them, because the difference is provenance and provenance is invisible to a pattern.
Never inject. None of the following may be added to a text that did not already contain it, and each is a failure even when the result scans clean: fake first person (if the source has no "I", the rewrite has no "I"); manufactured stakes ("now more than ever"); forced contrarianism (inventing a foil is inventing a claim); performed candor ("let's be honest", "real talk"); staccato conversion (chopping ordinary sentences to manufacture rhythm); and invented specifics, a number, name, date, tool, or mechanism the source never contained. Specificity is the most tempting fix because it always reads better, and a fabricated specific is worse than the vague phrasing it replaced. If the concrete detail is missing, leave a marked placeholder (
[ADD: which study?]
) and flag the gap. Never fill it.
The test: subtraction and sharpening are in scope (cutting filler, making an existing claim concrete, surfacing a buried point). Addition of stance, personality, or fact is not.
Do not overcorrect. Scrubbing at maximum strictness produces a second fingerprint: fragments, forced punch endings, every hedge stripped from a sentence whose uncertainty was real. That reads as humanizer output, which is its own genre of slop. Never write three sentences of five words or fewer in a row (the scanner flags runs of three). Keep roughness the author chose, keep a hedge that carries real uncertainty, and keep the register the venue requires: academic prose keeps calibrated hedging, safety text keeps its absolutes, legal text keeps its exceptions. You cut filler, not a language's natural redundancy. This rule tunes how hard you cut; it never relaxes the banned list.
关卡 1 审计的是用户交给你的草稿。以下规则约束你写出的文本，扫描器无法执行这些规则，因为差异在于来源出处，而模式看不见出处。
绝不可注入。 以下内容若原文没有，都不得被添加进去；即使最终结果扫描干净，出现任何一项仍算失败：虚假的第一人称（如果来源中没有 
I
，改写中也不能有 
I
）；人为拔高利害关系（
now more than ever
）；强行唱反调（虚构反方就是在虚构主张）；表演式坦诚（
let's be honest
、
real talk
）；短句化改写（把普通句子切碎，以制造节奏）；以及虚构具体信息，即来源中没有的数字、姓名、日期、工具或机制。具体化最诱人，因为它总是读起来更好；但编造出来的具体信息，比它所替代的含糊说法更糟。若缺少具体细节，就留下明确的占位符（
[ADD: which study?]
）并标出缺口。绝不填充。
测试： 删减和锐化在范围内（删除填充语、让既有主张更具体、显出被埋没的观点）。增加立场、个性或事实不在范围内。
不要过度纠偏。 以最高严格度进行清理会产生第二种指纹：残片、刻意有力的结尾，以及从承载真实不确定性的句子中剥掉所有限定词。这读起来就像经过 humanizer 处理的文字，而这种产物本身也是一种低质套话。绝不能连续写三句每句不超过五个词的句子（扫描器会标记连续三句的情况）。保留作者有意保留的粗粝感，保留承载真实不确定性的限定词，并保留媒介所要求的语域：学术写作保留经过校准的限定语，安全文本保留绝对措辞，法律文本保留例外条款。你删除的是填充语，而不是某种语言本身自然的重复。此规则调节清理力度；它绝不放宽禁用清单。

Gate 1: source-fidelity audit

关卡 1：来源忠实度审计

Runs whenever the content quotes anyone, attributes an idea, or presents a method as someone's; skip only when nothing is attributed. Extract every quoted span, verify each against the quote bank or live primary sources, and assign one of four verdicts: VERBATIM, CLOSE-PARAPHRASE, UNVERIFIED, or MISATTRIBUTED. Run it adversarially: try to catch the draft lying. Read 
references/fidelity-gate.md
for the full procedure, the traps found in real audits (fabricated fusion, silent paraphrase drift, mechanical false positives), and the correction-log format. The standard in one line: if you cannot point to where a quote came from, it is not a quote yet.
python scripts/slop_check.py quotes <draft.md> --corpus <sources-file-or-dir>
只要内容引用了任何人的话、把某个观点归于某人，或把某种方法说成某人的做法，就运行此关卡；只有完全没有归因时才跳过。提取每一段引文，将其与引语库或在线一手来源核验，并给出四种结论之一：VERBATIM、CLOSE-PARAPHRASE、UNVERIFIED 或 MISATTRIBUTED。以对抗方式运行：设法抓住草稿中的谎言。阅读 
references/fidelity-gate.md
，了解完整流程、真实审计中发现的陷阱（捏造式拼接、释义的隐性漂移、机械误报）以及修正日志格式。标准只有一句话：如果指不出引语来自哪里，它就还算不上引语。
python
undefined

Gate 2: taste audit

关卡 2：品味审计

Run the mechanical scan, then score nine dimensions from 1 to 5: source density, primary voice, case proof, originality, AI smell (inverted), rhythm variation, shareability, usefulness, public-object feel. Read 
references/taste-gate.md
for the scorecard anchors, scoring units per content type, and the required fixes. Never trade a Gate 1 fact back for a smoother sentence.
python scripts/slop_check.py tics <draft.md>
The scan blanks fenced code, inline code, blockquotes, and quoted spans before matching, so a draft may quote bad writing on purpose without tripping it. Pass 
--include-quoted
to scan inside quotations too. Two exceptions are deliberate. Chat-interface residue (
citeturn
, 
utm_source=chatgpt.com
, 
[Your Name]
) fires inside quotations as well, because it proves the quotation was pasted without being read, and it is marked 
(P0)
: fix it before anything else. And findings marked 
(1B)
are wordiness, not evidence about how the text was produced; report them separately and never count them toward AI texture.
Findings carry a marker for why they fired: unmarked means on sight, 
[tier 2]
means two or more landed in one paragraph, 
[tier 2, N/1000]
or 
[tier 3, N/1000]
means the document is saturated, and 
rhythm uniformity
means sentence-length variation collapsed. Vocabulary is the weakest signal in the set and rhythm is the strongest, so weight the report that way. 
references/taste-gate.md
has the full key.
先运行机械扫描，再按 1 到 5 分对九个维度评分：来源密度、主导声音、案例佐证、原创性、AI 味（反向计分）、节奏变化、可分享性、实用性、面向公众的感觉。阅读 
references/taste-gate.md
，查看评分表锚点、各内容类型的评分单位和必须执行的修正。绝不能为了更顺滑的句子，牺牲关卡 1 查实的事实。
python
undefined
扫描器在匹配前会排除围栏代码块、inline code、blockquote 和引文片段，因此草稿可以故意引用糟糕文字而不触发检测。传入 
--include-quoted
可连引文内部也一起扫描。有两项例外是刻意设置的。聊天界面残留（
citeturn
、
utm_source=chatgpt.com
、
[Your Name]
）即使在引文内部也会触发，因为它证明引文是未经阅读便粘贴进来的；此类项标为 
(P0)
：在其他问题之前先修复。而标为 
(1B)
的发现属于冗词，不是文本由 AI 生成的证据；应单独报告，绝不计入 AI 腔评分。
每个发现项都带有触发原因标记：无标记表示一眼即可看出；
[tier 2]
表示同一段落中命中两处或更多；
[tier 2, N/1000]
或 
[tier 3, N/1000]
表示文档已被此类模式饱和；
rhythm uniformity
表示句长变化已经坍缩。词汇是整套信号中最弱的，节奏最强，因此应据此权衡报告。
references/taste-gate.md
收录完整说明。

The verify loop (both modes)

核验循环（两种模式均适用）

Re-run the scan on everything you emit, and not only on the input: rewritten text and, in audit mode, the report itself. Skipping this step is the common failure: the model producing the fix carries the priors that produced the tell, so a flagged "not X, but Y" reliably comes back as "less about X than Y", the same move in a wig. Fix and re-scan until a pass returns zero hits. Cap at four passes. If a pattern survives four, rewrite that sentence from scratch starting from its bare claim: what fact or opinion is this sentence for? Report how many passes it took.
对你输出的每一样内容重新运行扫描，不只扫描输入：改写后的文本，以及审计模式下的报告本身。跳过这一步是常见失败：负责修复的模型仍带着最初制造破绽的先验，因此被标记的 
not X, but Y
往往会变成 
less about X than Y
，还是同一套手法，只是换了假发。修复并重新扫描，直到一次通过返回零命中。最多进行四轮。 如果某个模式连续四轮仍然存在，就从句子最基本的主张重新写起：这句话究竟要表达哪个事实或观点？报告用了多少轮。

Audit report format (audit mode output)

审计报告格式（审计模式输出）

undefined
undefined

No-Slop Audit: <piece name>

No-Slop Audit: <piece name>

Summary

Summary

  • Quotes checked: N (VERBATIM x / CLOSE-PARAPHRASE y / UNVERIFIED z / MISATTRIBUTED w)
  • Tic scan: N hits, clean after P verify passes | Taste: lowest dimensions named
  • Verdict: SHIP / FIX THEN SHIP / REWRITE
  • Quotes checked: N (VERBATIM x / CLOSE-PARAPHRASE y / UNVERIFIED z / MISATTRIBUTED w)
  • Tic scan: N hits, clean after P verify passes | Taste: lowest dimensions named
  • Verdict: SHIP / FIX THEN SHIP / REWRITE

Gate 1: correction log (table per references/fidelity-gate.md)

Gate 1: correction log (table per references/fidelity-gate.md)

Gate 2: tic scan (file:line | matched text in backticks | replacement)

Gate 2: tic scan (file:line | matched text in 
backticks
| replacement)

Gate 2: taste scorecard (9 rows, scored, one line of evidence each)

Gate 2: taste scorecard (9 rows, scored, one line of evidence each)

Required fixes (numbered, most damaging first)

Required fixes (numbered, most damaging first)


Lead with the summary counts. In write mode, the report collapses to one line per gate
confirming the checks ran clean, plus anything you had to fix.

**The report is prose you produced, so the style rules govern it too.** This is the failure
mode that discredits an audit fastest: a report that spends a paragraph removing the user's em
dashes while using four of its own. Render every matched span, banned phrase, and dash you
quote inside backticks, which the scanner exempts as inline code, then scan the report before
delivering it. An audit report that fails its own scan is not deliverable. Verdict criteria are
in `references/taste-gate.md`; use them so two auditors reach the same call.

先给出摘要计数。在写作模式下，报告压缩为每道关卡一行，确认检查均已干净通过，并列出所有修正项。

**报告也是你产出的文字，因此风格规则同样适用。** 这是最快让审计失去可信度的失败模式：报告花掉一整段来删除用户的 em dash，自己却用了四个。把每一个命中的片段、禁用短语和你引用的破折号都放在反引号内，扫描器会将它们作为 inline code 豁免，然后在交付前扫描报告。无法通过自身扫描的审计报告不可交付。结论判定标准见 `references/taste-gate.md`；照此执行，让两位审计者得出相同结论。