anti-slop-writing

反套路写作

任何重要的写作或改写任务都使用此技能。

仅在任务需要时加载本地参考资料：

references/anti-slop-writing-doctrine.md — read for a full prose review or when changing the doctrine.
references/flow-by-relation.md — read when paragraphs are locally clear but list-like, or when fixing transitions/conclusions.
references/rewrite-patterns.md — read when producing multiple concrete rewrites or adding a new failure pattern.

核心原则

Sharp detail beats inflated significance.

AI 写作常见的问题是：

less detail, more importance

这样改进：

more detail, earned importance

写作前

回答以下问题：

What is the exact point?
What concrete detail proves it?
What sentence machinery is making it feel true?
What should the reader remember?

如果答案不明确，请先提问或检查来源，再开始起草。

基于来源的检测器

改写前进行以下检查：

AI-writing signs: superficial analysis, undue significance language, canned emphasis, negative parallelisms, rule-of-three overuse, formulaic dashes, table/bold formatting as fake structure.
Rhetorical-style drift: watch noun-heavy abstractions, nominalizations, phrasal coordination, and informational density without mechanism.
Practical AI-editing tells: repeated “Not X. Y.” rhythm, symmetrical paragraph length, parallel headings, and bullet + bold-header + colon patterns.
Copula displacement: watch “serves as,” “stands as,” “features,” “marks,” and “represents” replacing plain “is/are.” Keep the verb when it does concrete work (enumerating, defining, locating); replace it when it only inflates a copula.
Hedged symmetry: watch “Whether you’re X or Y,” “While X, Y is also important,” and similar templates that address every reader or every value at once. Keep the structure only when it names a real branching condition with distinct downstream behavior.
Em-dash cadence: do not ban em-dashes; watch decorative clusters that use dashes for emphasis instead of for parenthetical or appositive insertion. Reduce count; keep dash insertions/pairs that bracket inline definitions or genuine asides.
Rhetorical staccato: do not ban antithesis/parallelism; watch antithetical parataxis where rhythm implies the relation before evidence is unpacked.
Hypotaxis preference: when the relation matters, prefer subordination and connective syntax over side-by-side clauses; use “because,” “although,” “when,” “while,” “where,” “once,” or an explicit summary noun to show which idea modifies which.
Flow-by-relation test: paragraphs should make the next question possible, not merely sit beside each other in a plausible order. At section boundaries, name the relation: cause, contrast, dependency, inference, resolution, scope change, or level-of-detail change.
Conclusion test: a conclusion should return to the concrete carrier, admit the limit if needed, and state the reusable structure. Avoid ending with a generic thesis sentence that could belong to any essay in the category. Reject outline templates such as “Despite challenges, X continues to thrive” and “Looking ahead, X will play an increasingly pivotal role” unless backed by a specific challenge or named next step.
Unseeing frame: unsee the polished sentence into the machinery that made it feel true; inspect cadence, contrast, omitted relation, evidence scope, and abstraction leaks.
Emphasis-source test: ask whether the emphasis comes from the idea or from a borrowed pattern. To apply, write the flattened version of the line in your critique (cadence removed, claim preserved) and judge whether the residual claim still names an actor, mechanism, or limit. If it collapses into a generic statement, the rhythm was carrying it; if it still lands, the idea was.
Syntax-relation test: ask whether the syntax clarifies the relation between ideas or only suggests one. To apply, restate the implied relation in plain prose using a connective (because, although, when, where, once, so that). If you cannot supply the connective without inventing it, the syntax was standing in for a relation that was not there.
Contribution bar: contribution must justify length; add judgment, framing, depth, or cut.
DevRel pressure test: durable prose needs judgment, evidence, runnable artifacts, and taste.

默认编辑流程

完成文稿前：

1. Delete generic opening.
2. Find prestige vocabulary clusters.
3. Replace abstract nouns with concrete mechanisms.
4. Remove “not just X but Y” unless it is the exact point.
5. Cut superficial “highlighting/underscoring” clauses.
6. Check every claim of importance against evidence.
7. Replace vague actors with named sources, named uncertainty, or an `ask-author` note that says what to ask. Do not invent a name, count, tool, or timing to fill a `Concrete rewrite` slot.
8. Collapse redundant bullets.
9. Vary sentence rhythm deliberately.
10. Audit staccato contrast cadence by identifying the hidden relation: contrast, evidence scope, cause, consequence, exception, or scope control.
11. Prefer hypotaxis to parataxis when the relation matters: subordinate the secondary idea instead of placing equal-weight clauses side by side.
12. Check paragraph flow: each paragraph should answer the prior question or make the next question necessary. Add hinge sentences when sections read like a list.
13. Check the conclusion: return to the concrete carrier, name what was made visible or solved, state what transfers, and avoid a generic final sentence.
14. Replace rhythm with relation when the line still sounds good but has not named the mechanism.
15. Replace displaced copulas (“serves as,” “stands as,” “features,” “marks,” “represents”) with plain “is” or a specific action verb, unless the verb is doing concrete enumerating, defining, or locating work.
16. Cut hedged symmetry (“Whether X or Y,” “While X, Y is also important”) and commit to a specific reader and a named tradeoff.
17. Reject outline conclusion templates (“Despite challenges, X continues to thrive,” “Looking ahead, X will play a pivotal role”); replace with a carrier-bound ending or cut.
18. Thin decorative em-dash clusters; keep at most one earned dash insertion/pair per sentence unless multiple insertions carry distinct definitions.
19. End with a concrete remembered line.

默认避免使用

短语：

In today's rapidly evolving landscape
In the realm of
When it comes to
At its core
Let's dive into
It's worth noting that
It's important to note that
A testament to
Not just X, but Y
Not only X, but also Y
Same X. Same Y. Different Z.
Not X. Y.
This is where X comes in
Whether you're X or Y
While X, Y is also important
Despite ongoing challenges, X continues to thrive
Looking ahead, X will play an increasingly pivotal role
In conclusion
Overall
Ultimately
I hope this helps

需要审查的词语：

delve
realm
landscape as metaphor
tapestry
testament
pivotal
crucial
underscore
intricate
meticulous
multifaceted
nuanced as filler
foster
bolster
garner
showcase
highlight
emphasize
encompass
utilize
facilitate
transformative
groundbreaking
seamless
robust outside engineering context

注意：以上词语和短语列表是带有时间标记的检测器，反映了当前几代模型中观察到的模式，之后会发生变化。添加或移除条目之前，请使用最新的人类与大模型语料重新分析；delve 是一个警示案例：这个高风险词在大模型输出中的出现频率于 2025 年大幅下降。

有意不把 serves as 和 stands as 这类系词结构列入词语列表。这些是由两个词构成的模板，是否合适取决于上下文：动词用于枚举、定义或定位时保留；若只是让系词表达显得冗长，则替换。

避免误报

将检测器命中视为待验证的假设，而非定论。如果同一句或邻近上下文提供了足以支撑该术语的机制、失效模式、测量结果或边界，就应判定为 Verdict: keep，并指出支撑依据。

示例：

Keep: "The queue is robust because each job has an idempotency key, a retry receipt, and a dead-letter cutoff."
Why: "robust" is an engineering qualifier earned by the deduplication, retry-tracking, and dead-letter mechanisms.

不要仅仅因为出现了观察名单中的词语，就提供只替换同义词的改写。

短促对比测试

不要把每个简短对比都视为套路化表达。先判断它的类型。测试标准是：对比的两侧是否都有前文支撑，而不只是对比主题是否被提及。

Earned antithesis: both sides of the contrast are evidenced in the prior sentences. The contrast lands a distinction the reader can already verify. Keep or use once.
Compressed antithesis: one side is evidenced; the other side is a leap or a new claim. The cadence implies the contrast before evidence is given. Expand into the relation by naming the unsupported side directly.
Decorative antithesis: neither side is evidenced. The contrast supplies closure without content. Cut or replace.

同一来源中的实例分析（joe.dev/posts/thinking-out-loud）：

有依据：文章证明了平台的失效（Medium 转型、Twitter、LinkedIn），也展示了替代方案（你拥有自己的域名，也运行了自己的服务器）。结尾的 That's what I want. My words, at a URL I own, without anyone in the middle. 点明了一个对比，而对比的两侧都有所展示。

压缩过度：前文证明了可迁移性——ATproto lets you store structured records in a data server (your PDS) that you own and can move.——但没有证明这是有意设计的。结尾的 That's not incidental. It's the design. 仅凭节奏感就断言了设计意图。改写时应指出可移植性为何是设计的一部分（即在 PDS 迁移过程中保持记录身份稳定的协议属性），而非直接断言。

装饰性对比：The point is not the pelicans. The point is the process. 对比之前既没有说明鹈鹕载体，也没有说明流程。改写时应点明两者的关系：The pelican is useful because it gives the process a small, inspectable carrier.

改写规则：

replace rhythm with relation
prefer hypotaxis to parataxis

常见改写方式：

Although X, Y.
Because X, Y.
When X changes, Y no longer means the same thing.
Where X only shows final state, Y records process.
X matters because it changes Y.
Once X is visible, Y becomes inspectable.
Because the concrete example is small/memorable/inspectable, it can carry the larger claim.

改写模式

不好：

This underscores the importance of durable execution.

更好：

The workflow can fail on step 4, retry only that step, and keep the previous outputs.

不好：

Cloudflare is not just a CDN, but a platform.

更好：

Cloudflare turns the network boundary into a programmable runtime.

不好：

This empowers teams to build seamless experiences.

更好：

The team can ship a WebSocket room without running a room server.

较弱的结尾：

A benchmark is stronger when you can inspect the run that produced it.

更好的结尾：

Because the pelican project is small enough to inspect and strange enough to remember, it works as a compact carrier for the larger claim: a benchmark is stronger when you can inspect the run that produced it.

最终自检

在宣布编辑完成前：

Did I answer the user's actual request: review, rewrite, draft, or all three?
Did every flagged slop tell get a concrete rewrite or a reason to cut?
Does each rewrite name the mechanism, relation, actor/action/result, or concrete carrier?
Did I preserve earned compression instead of expanding it into bland explanation?
Did I avoid introducing new prestige abstractions or banned fallback phrases?
Did I run the same detectors on my rewrite that I ran on the source? In particular, does my rewrite reuse the cadence I just flagged (em-dash antithesis, "X isn't A, it's B", short parallel clauses) under different punctuation?
When a rewrite needs a fact not in the source (a name, a number, a tool, a mechanism), did I ask the author or recommend cutting, rather than inventing a confident-sounding specific?

对于高风险文稿，进行一次有明确范围的评审与改进：按 1 到 5 分评价具体性、证据契合度、关系清晰度和节奏；只改进得分最低的一项。

批评输出格式

审阅现有文案时，请使用：

Verdict: keep / revise / ask-author / reject
Slop tells:
Specificity missing:
Inflated claim:
Flow break:
Concrete rewrite:
Rewrite check:
Remembered line:

如果某一行确实可以改进，但修改需要原段落未提供的信息，例如工具名称、人物、数量、时间说法、具名机制或具体事件，请使用 ask-author。在 Concrete rewrite 栏中说明要询问什么，例如： Ask author: which coding tool? Cursor, Claude Code, Copilot, Aider, other? 并提供一个备用方案（删掉，或保留该内容，让后续句子承载相关信息）。不要编造缺失信息来填入 Concrete rewrite 栏。ask-author 区块内的任何备用改写也遵循相同规则：编造事实的备用方案还不如不提供备用方案。

Rewrite check 为必填项。说明你的 Concrete rewrite（包括任何 ask-author 备用方案）是否包含以下任一内容：三段式结构、X 而非 Y／否定式平行结构、破折号式对立、默认禁用的规避措辞、彰显地位的形容词、装饰性收尾（“That was the point”“In conclusion”“Overall”“Ultimately”），或编造的事实。如果该改写本应被标记为源段落的 revise，那它自身也应标记为 revise：重新改写或升级为 ask-author。如果改写通过检查，请写 passes self-detectors。

判断对立表达或短促对比时，先阅读同一段落的前一句。如果前一句交代了对比所指向的机制，那么这种对比就有依据。不要仅凭节奏感，就把有依据的对比判定为过度压缩或装饰性表达。

目标语气

倾向于：

specific
source-grounded
direct
varied rhythm
concrete mechanisms
named tradeoffs
memorable line

避免：

safe essay voice
prestige abstraction
balanced filler
fake profundity
marketing fog