write-well

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

write-well

Great prose has a person behind it: a voice, a body, opinions, a sense of rhythm. The patterns that make writing "sound like AI" (significance inflation, rule-of-three, em-dash spray, flat reporting) are symptoms of one disease: no voice. So the move is positive. Add craft and voice, and the tells disappear on their own. Subtracting tells from voiceless prose just gives you cleaner voiceless prose.

优秀的散文背后是真实的人：有独特风格、真实体感、个人观点，还有节奏韵律感。让写作“听起来像AI”的那些模式（意义膨胀、三连式表达、破折号滥用、平淡陈述）其实是同一种问题的表现：缺乏独特风格。所以解决思路是积极的——注入写作匠心和独特风格，AI特征自然就消失了。仅仅去除无风格文本中的AI特征，得到的只是更工整但依然无灵魂的文本。

Two modes

两种模式

Mode	For	Starting point
draft (default)	writing from scratch	a topic, an outline, a blank page
edit	text that's flat, generic, or AI-flavored	an existing draft to bring alive

Both modes serve the same core. They just enter from different ends.

模式	适用场景	起始基础
draft（默认）	从头创作	主题、大纲或空白页面
edit	修改平淡、通用或AI风格的文本	需要赋予生命力的现有草稿

两种模式服务于同一个核心，只是切入方式不同。

The core — what good prose has (both modes)

核心要素——优秀散文必备（两种模式通用）

Voice. A real person reacting, not a neutral reporter. Opinions ("I genuinely don't know how to feel about this"), acknowledged complexity ("impressive but unsettling"), first person when it fits, a little mess (tangents, self-correction). Voiceless-but-clean is as obvious as slop.
Reliving, not reporting (VAKS). Most weak writing reports ("I met her at a bar, did a trick, won her over"). Good writing relives it with sensory detail: Visual, Auditory, Kinesthetic, Smell. Every scene or section needs ≥2 VAKS ingredients, at least one non-visual. Diagnostic: "am I watching a movie or reading a news summary?" Summary means add VAKS.
Rhythm — vary sentence length. The single biggest lever for "music." Short sentences punch. Then a longer one builds and breathes and carries the reader somewhere before it sets them down. Then a fragment. If every sentence is the same length, the prose is a monotone (Provost). Read it aloud. Your ear catches what your eye skims.
Specific over abstract. "I moved schools four times because I kept getting bullied" beats "I had a difficult childhood." Dense, concrete words carry more than vague ones. Name the thing.
Story-first structure. Hook, Struggle, Breakthrough, Application. Drop into a moment, relive the difficulty, show the turn, link to the point last. The AI default inverts this (state the lesson, backfill with anecdotes), and the result reads like a sermon. Lead with life; arrive at the principle. (Register-gated; see below. A performance review is not a personal essay.)

独特风格。是真实的人在表达，而非中立的陈述者。要有个人观点（“我真的不知道对此该作何感受”）、承认事物的复杂性（“令人印象深刻但也不安”），适合时使用第一人称，带点“小瑕疵”（题外话、自我修正）。工整但无风格的文本和粗糙的文字一样显眼。
重现而非陈述（VAKS原则）。大多数劣质写作只是陈述（“我在酒吧遇到她，表演了个小把戏，赢得了她的好感”）。优秀的写作会用感官细节重现场景：V视觉、A听觉、K触觉、S嗅觉。每个场景或段落至少需要2种VAKS元素，且至少有一种非视觉元素。判断标准：“我是在看电影还是在读新闻摘要？”如果是摘要，就需要补充VAKS元素。
节奏——变换句子长度。这是打造文字“韵律感”最有效的方法。短句有冲击力，长句则能铺陈、舒缓情绪，带领读者进入状态后再收尾，接着用一个碎片句。如果所有句子长度一致，散文就会单调乏味（Provost）。大声朗读，耳朵能捕捉到眼睛忽略的问题。
具体而非抽象。“我因为一直被霸凌，转了四次学”比“我有一个艰难的童年”更有力量。密集、具体的词汇比模糊的表达承载更多信息，要明确说出事物的名称。
故事优先的结构。钩子、挣扎、突破、应用。从某个瞬间切入，重现困境，展示转折，最后再点明核心观点。AI的默认结构是反过来的（先讲结论，再用轶事填充），结果读起来像布道。先讲生活场景，再引出原则。（受语体限制；详见下文。绩效评估不属于个人随笔。）

Punctuation budget — a generation rule, not a cleanup

标点预算——创作时遵循的规则，而非事后修正

Set this while writing, not in a later pass. The em-dash is the loudest AI tell, and the model's instinct is to oversupply it.

Em-dash budget: aim for zero, cap at ~2 per 1,000 words of nonfiction. Don't type the em-dash in the first place. Reach for a comma, a period (start a new sentence), a colon (when the second half delivers on the first), or parentheses (a true aside) instead. Reserve the em-dash for a genuine interruption or hard reversal, and spend it deliberately.
Count in the final pass. Search the draft for every
```
—
```
, and replace each one that isn't an earned interruption. If a draft of 600 words has eight em-dashes, seven are wrong.
This is a default with one exception: a
```
--voice
```
/
```
--like
```
whose own fingerprint runs on em-dashes (DFW, some essayists) restores them, because the voice outranks the house default. Absent that signal, stay near zero.
Same discipline, lighter touch, for the rule-of-three and the short-fragment closer below: countable habits, not reflexes.

要在写作时就设定好，而不是事后再调整。破折号（em-dash）是最明显的AI写作特征，模型本能地会过度使用它。

破折号预算：目标为零，非虚构类每1000字最多约2个。一开始就不要输入破折号。改用逗号、句号（开启新句子）、冒号（后半句承接前半句内容）或括号（真正的题外话）。破折号只用于真正的打断或强烈转折，且要谨慎使用。
最终检查时计数。在草稿中搜索所有
```
—
```
，替换掉每一个并非必要打断的破折号。如果一篇600字的草稿有8个破折号，其中7个都是不合理的。
这是默认规则，仅有一种例外：如果
```
--voice
```
/
```
--like
```
指定的风格本身频繁使用破折号（比如DFW或某些随笔作家），则可以保留，因为风格优先级高于默认规则。没有这类信号时，尽量保持零使用。
对于三连式表达和下文提到的短句碎片收尾，也要遵循同样的原则：有意识地使用，而非本能反应。

Register & tone — set this first

语体与语气——先设定好

The core has a bias toward punch: short kicks, high-contrast rhythm, the dramatic short-fragment closer ("She did that."), "don't narrate the landing." That register is right for marketing, fiction, and personal essays, and wrong everywhere else. Set the register before you reach for those tools. Ask or infer it from the task if it isn't given.

Register	Wants	Avoid
Punchy / dramatic (marketing, fiction, personal essay)	short kicks, the earned closer, contrast, voice turned up	flatness, hedging
Measured / professional (performance feedback, business, diplomatic, reference, academic)	specific, balanced, kind; even rhythm; substance-led; claims you'd stand behind	the zinger ending, manufactured drama, the fragment closer

The punchy short-fragment closer is now its own AI tell. Don't reach for it by reflex. In measured registers, close on substance, not a kick: the last line lands a point, it doesn't perform one. Performance feedback specifically: lead with the specific behavior and its impact, balance strength and growth, stay warm and direct, never end on a one-liner.

核心要素偏向冲击力：短句、高反差节奏、戏剧性的短句碎片收尾（“她就那么做了。”）、“不要叙述结局”。这种语体适合营销文案、小说和个人随笔，但不适用于其他场景。在使用这些技巧前，先设定好语体。如果任务未明确说明，就自行判断或询问。

语体	适用技巧	避免事项
有冲击力/戏剧性（营销、小说、个人随笔）	短句、恰当的收尾、反差、强化风格	平淡、含糊其辞
沉稳/专业（绩效反馈、商务、外交、推荐信、学术）	具体、平衡、友善；节奏平稳；内容导向；观点有依据	俏皮结尾、刻意制造戏剧冲突、短句碎片收尾

戏剧性的短句碎片收尾现在本身就是一种AI写作特征。不要本能地使用它。在沉稳语体中，要以内容收尾，而非冲击力：最后一句要传递观点，而非刻意表演。尤其在绩效反馈中：先讲具体行为及其影响，平衡优势与成长空间，保持真诚直接，永远不要以俏皮话收尾。

Voices & personas

风格与人设

By default write-well calibrates a voice to the piece, audience, and register (above). You can also set one explicitly, and the LLM adopts it with no config needed:

```
--voice <name>
```
adopts a named voice. Two libraries: pick a register voice by what you're writing (
```
plain-professional
```
,
```
warm-feedback
```
,
```
diplomatic
```
,
```
executive-brief
```
,
```
conversational
```
,
```
academic
```
,
```
punchy-marketing
```
,
```
storyteller
```
), no author knowledge needed, or an author voice if you know it (Hemingway, Didion, and so on).
```
--persona <name|description>
```
is a voice plus a stance/worldview (the implied author, e.g. "skeptical engineer"). A persona writes as someone.
```
--like <sample path or paste>
```
clones a voice: read the sample's fingerprint and write toward it. Use this to match your habits, including whether you use em-dashes at all.

Every voice is one 6-axis fingerprint: diction · syntax (parataxis↔hypotaxis, clause-nesting) · rhythm (sentence-length variance) · POV/distance · tone/punctuation · stance (persona only). The first five are clonable from a sample; stance is the rhetorical layer you describe, not sample. Full model plus a starter library (Hemingway, McCarthy, DFW, Vonnegut, Morrison, Didion, Thompson, Faulkner, Ogilvy) with 2–4 exemplars each: references/voices.md.

Two failure modes — engineer against both (a voice tuned away from one drifts into the other):

Regression to generic (the default gravity, where the LLM falls back to bland web-average). Guard: make the voice's most distinctive axis a hard constraint (e.g. Hemingway means no subordinating conjunctions) and pass 2–4 exemplars.
Caricature (maxing every axis into parody: "and…and…and" every line). Guard: treat the fingerprint as a center of gravity with variance, not a rule applied to 100% of sentences. Real voices vary.

edit

/rewrite, separate meaning from style: re-skin the voice, keep the content, and watch for meaning drift.

默认情况下，write-well会根据文本类型、受众和语体（上文）调整风格。你也可以明确设定风格，LLM无需额外配置就能直接采用：

```
--voice <name>
```
采用指定名称的风格。两个库：根据写作类型选择语体风格（
```
plain-professional
```
、
```
warm-feedback
```
、
```
diplomatic
```
、
```
executive-brief
```
、
```
conversational
```
、
```
academic
```
、
```
punchy-marketing
```
、
```
storyteller
```
），无需了解作者；或者如果你熟悉作者，选择作者风格（比如Hemingway、Didion等）。
```
--persona <name|description>
```
是风格加立场/世界观（隐含作者，比如“持怀疑态度的工程师”）。人设代表以某个人的身份写作。
```
--like <sample path or paste>
```
克隆风格：读取样本的风格特征，以此为基准写作。用这个功能匹配你自己的写作习惯，包括是否使用破折号。

每种风格都对应一个6维度特征：措辞·句法（并列vs主从、从句嵌套）·节奏（句子长度变化）·视角/距离·语气/标点·立场（仅人设包含）。前5个维度可以从样本中克隆；立场是你描述的修辞层面，无需样本。完整模型加初始风格库（Hemingway、McCarthy、DFW、Vonnegut、Morrison、Didion、Thompson、Faulkner、Ogilvy），每个风格有2-4个范例：references/voices.md。

两种失败模式——需同时规避（调整一种风格时，容易陷入另一种）：

回归通用风格（默认趋势，LLM会退回到平淡的网络平均风格）。防范措施：将风格最独特的维度设为硬性约束（比如Hemingway风格意味着不使用从属连词），并提供2-4个范例。
夸张模仿（将所有维度拉满，变成拙劣模仿：每句都用“and…and…”）。防范措施：将风格特征视为带有变化的核心趋势，而非适用于每一句的规则。真实的风格是有变化的。

在

edit

/重写模式中，区分内容与风格：重新塑造风格，保留内容，注意避免内容偏离。

The craft canon

写作权威准则

The five-part core is the spine. The full positive ruleset, deduped across Strunk, Zinsser, Orwell, Williams, Lanham, Clark, Pinker, Le Guin, McPhee, and Provost, lives in references/craft-canon.md. The load-bearing rules beyond the core:

Characters as subjects, actions as verbs; kill nominalizations (Williams). The strongest positive clarity engine: it tells you what to build, not just what to delete.
Cohesion: open with old information, end with the new (Williams). This is what makes paragraphs flow, not just sentences shine.
The Paramedic Method (Lanham) for edit mode: circle prepositions and be-verbs, find the action, make the doer the subject, start fast.
Right-branch; emphatic word last; keep related words together (Clark; Strunk & White R22/R20). Information order within the sentence: don't split the subject from its verb.
Express coordinate ideas in similar form (Strunk & White R19). Like content in like grammatical form, most load-bearing in lists, goals, tenets, and headings, where a broken parallel reads as a stumble.
Make the paragraph the unit; front the point in exposition/argument (Strunk & White R13). The register counterpart to the core's story-first rule: narrative delays the point, argument fronts it. A memo paragraph opens with its claim; an essay paragraph opens with a moment.
Structure from the material, kept invisible; write the lead first (McPhee). The nonfiction structure model.
Classic style (Pinker): prose as a window. Show the reader what you see, as an equal. The default stance for essay/newsletter/blog.
Beat the curse of knowledge (Pinker). You're too close to the subject to see what the reader doesn't know, so use a test reader.

Rules are defaults with reasons, not commandments. Keep the positive maxims, but the grammar superstitions (never the passive, never split an infinitive, never open with "however") are folklore. Know why a rule exists, then break it when it serves the reader. Inviolable: voice, specificity. Calibration-dependent: em-dash budget, tricolon limit, passive voice (use it when the acted-upon is the reader's focus).

五要素核心是基础。完整的正向规则集来自Strunk、Zinsser、Orwell、Williams、Lanham、Clark、Pinker、Le Guin、McPhee和Provost，已去重，详见references/craft-canon.md。核心之外的关键规则：

以角色为主语，以动作作动词；消灭名词化表达（Williams）。这是最有效的正向清晰表达工具：它告诉你要构建什么，而不只是删除什么。
连贯性：以已知信息开头，以新信息结尾（Williams）。这是让段落流畅的关键，而非仅仅让句子出彩。
急救法（Paramedic Method）（Lanham），适用于编辑模式：圈出介词和be动词，找到核心动作，将执行者设为主语，快速切入主题。
右分支结构；重点词汇放最后；相关词汇放在一起（Clark；Strunk & White规则22/20）。句子内的信息顺序：不要将主语和动词分开。
并列观点采用相似形式表达（Strunk & White规则19）。同类内容用相似的语法形式，在列表、目标、原则和标题中尤为重要，形式断裂会显得生硬。
以段落为单位；说明/论证类文本先点明观点（Strunk & White规则13）。这是核心故事优先规则的语体对应原则：叙事类文本延迟点明观点，论证类文本先点明观点。备忘录段落开头就提出主张；随笔段落开头则是某个瞬间。
从素材出发构建结构，保持结构隐形；先写开头（McPhee）。非虚构类文本的结构模型。
经典风格（Pinker）：散文如窗口。向读者展示你所看到的，平等交流。这是随笔/通讯稿/博客的默认立场。
克服知识诅咒（Pinker）。你对主题过于熟悉，看不到读者不知道的内容，所以要找测试读者。

规则是有原因的默认选项，而非戒律。保留正向准则，但语法迷信（永远不用被动语态、永远不拆分不定式、永远不用“however”开头）只是民间说法。要明白规则存在的原因，当对读者有利时就打破它。不可违背的原则：风格、具体性。可调整的内容：破折号预算、三连式表达限制、被动语态（当动作承受者是读者关注的焦点时可以使用）。

draft mode

draft模式

Write the piece, holding the core (and the chosen voice). Plus:

Open on a moment, not exposition. Sensation, dialogue, or a question, never a concept the reader doesn't care about yet. The opening earns or loses the reader.
POV-colored vocabulary (fiction): a painter sees composition, an engineer sees load-bearing walls. Narration should taste like the character.
Dialogue does the work. Don't narrate exposition; let characters say it, argue it, get it wrong. Each line answers the actual previous line. Don't prop up "said" with adverbs or explanatory verbs ("he said consolingly," "she consoled"); let the line disclose the manner (White, don't explain too much).
Don't narrate the landing. Never state the emotional conclusion ("that changed everything"). If you have to name it, the scene didn't earn it.

Craft detail and examples: references/prose-craft.md.

创作文本时，遵循核心要素（以及选定的风格）。另外：

从瞬间切入，而非说明性内容。用感官细节、对话或问题开头，永远不要用读者还不关心的概念开头。开头决定了读者是否会继续读下去。
视角化词汇（小说）：画家关注构图，工程师关注承重墙。叙事语言要符合角色的身份。
对话承担叙事功能。不要叙述说明性内容；让角色说出来、争论、说错。每一句对话都要回应上一句的真实意图。不要用副词或解释性动词来修饰“said”（比如“he said consolingly”、“she consoled”）；让对话本身体现语气（White，不要过度解释）。
不要叙述结局。永远不要直接说出情感结论（“那改变了一切”）。如果必须点明，说明场景还不足以支撑这个结论。

写作细节和范例：references/prose-craft.md。

edit mode

edit模式

The text exists but it's flat or AI-flavored. Bring it alive:

Add voice first (the positive half). Inject opinions, complexity, first-person where it fits, varied rhythm, specific feeling. This is what most "humanize" passes skip, and it's the half that matters.
Then fix the patterns. Work the catalog of 29 recurring AI patterns (significance inflation, promotional language, superficial -ing phrases, vague attributions, rule-of-three, copula avoidance, em-dash overuse, inline-header lists, hedging, filler, sycophancy, and the rest) with before/after fixes: references/edit-catalog.md.
Run the Paramedic Method on bloated sentences (see the craft canon) for a mechanical concision pass.
Match the target voice. If
```
--voice
```
/
```
--like
```
is set, rewrite toward that fingerprint, not toward generic "natural."
Cool off and check facts (anything that matters). Sleep on it, reread aloud, get a second reader; verify quotations, names, and numbers. See the craft canon's cooling-off pass (Ogilvy).

文本已经存在，但平淡或带有AI风格。赋予它生命力：

先注入风格（积极的一步）。加入个人观点、复杂性、适合的第一人称、多变的节奏、具体的感受。这是大多数“人性化”步骤会忽略的部分，但却是最重要的。
然后修正模式问题。处理29种常见的AI写作模式（意义膨胀、宣传语、表面化的-ing短语、模糊归因、三连式表达、避免系动词、破折号滥用、内嵌标题列表、含糊其辞、填充内容、奉承等），包含修改前后的示例：references/edit-catalog.md。
对冗长句子使用急救法（见写作权威准则），进行机械性的精简。
匹配目标风格。如果设置了
```
--voice
```
/
```
--like
```
，要向该风格特征靠拢，而非通用的“自然风格”。
冷却后检查事实（重要内容）。先放一放，大声朗读，找他人帮忙审阅；核实引文、姓名和数字。见写作权威准则中的冷却检查步骤（Ogilvy）。

The final pass (both modes)

最终检查（两种模式通用）

Cheap, high-yield. After drafting or editing, run three prompts:

"Count the em-dashes; replace every one that isn't a genuine interruption." The budget is enforced here if it slipped during drafting.
"What makes this sound AI-generated or flat?" Name the remaining tells honestly.
"Now fix exactly those." Revise.

This catches what the first pass missed.

低成本、高收益。创作或编辑完成后，运行三个提示：

“统计破折号数量；替换掉每一个并非必要打断的破折号。” 如果创作时没遵守预算，在这里强制执行。
“哪些地方听起来像AI生成的或平淡乏味？” 诚实地找出剩余的AI特征。
“现在只修正这些问题。” 进行修改。

这能找出第一轮检查遗漏的问题。

Banned words & phrases

禁用词汇与短语

Hard-ban list (delve, tapestry, testament to, "It's not just X, it's Y", "at its core", bold-colon bullets, and so on): references/banned-words.md. Quick gut-check: if a sentence "says everything and means nothing," cut it.

严格禁用列表（delve、tapestry、testament to、"It's not just X, it's Y"、"at its core"、粗体冒号项目符号等）：references/banned-words.md。快速判断：如果一句话“说尽一切却毫无意义”，就删掉它。

Self-review checklist

自我检查清单

Domains

适用领域

Generic domain notes (apply the core, weight differently): references/domains.md covers fiction, articles/essays, marketing/testimonials, and faith/values content. For a specific project, layer that project's own style guide on top. This skill is the foundation, not the override.

通用领域说明（遵循核心要素，权重不同）：references/domains.md涵盖小说、文章/随笔、营销/推荐语、信仰/价值观内容。针对特定项目，要叠加该项目自身的风格指南。本技巧是基础，而非替代方案。

Gotchas

注意事项

The em-dash rule only works if you obey it while writing. "Near-zero" as a review-pass note failed in practice: the model drafts in em-dashes and the cleanup doesn't catch them all. Treat the budget as a generation rule (don't type it), then count in the final pass. The skill's own files are held to the same budget, and its eval fails if they regress, because the prose write-well is written in is the example the model imitates.
Over-correcting kills voice. Applying every rule rigidly produces sterile prose, as obviously AI as the tells. The core is the goal; the rules are guardrails. When in doubt, keep the voice.
Set the register before the punch. The dramatic levers (short kicks, the fragment closer, manufactured drama) belong to marketing, fiction, and personal essay. In performance feedback, professional, or diplomatic writing they read wrong, so go measured, specific, and kind, and never end on a zinger. The default is not punchy.
A voice sits between generic and caricature. Both are failures. Generic means no fingerprint; caricature means the fingerprint on every sentence. Aim for a center of gravity with variance.
Rules are defaults, not commandments. Don't apply the grammar superstitions robotically; that makes write-well the very rule-bot it warns against (Pinker).
Add, don't just subtract. Edit mode's first step is add voice. A draft scrubbed of every tell but still flat has not been fixed.
Chat is not prose. This is for long-form content. Don't apply it to quick replies.
"Too much" is usually unfamiliar, not wrong. The risk is rarely going too far; it's not going far enough.

破折号规则只有在写作时遵守才有效。将“接近零”作为检查步骤在实践中失败了：模型创作时会大量使用破折号，事后清理无法全部发现。要将预算作为创作时的规则（不要输入破折号），然后在最终检查时计数。本技巧自身的文件也遵循同样的预算，如果违反，评估就会失败，因为write-well自身的写作就是模型模仿的范例。
过度修正会扼杀风格。僵化地应用所有规则会产生生硬的文本，和AI特征一样明显。核心要素是目标，规则只是护栏。不确定时，保留风格。
先设定语体，再考虑冲击力。戏剧性技巧（短句、碎片收尾、刻意制造冲突）属于营销、小说和个人随笔。在绩效反馈、专业或外交写作中使用会显得不合时宜，所以要采用沉稳、具体、友善的风格，永远不要以俏皮话收尾。默认风格不是有冲击力的。
风格要介于通用和夸张模仿之间。两者都是失败的。通用风格意味着没有独特特征；夸张模仿意味着每句都带有同样的特征。要追求带有变化的核心趋势。
规则是默认选项，而非戒律。不要机械地应用语法迷信；否则write-well就会变成它所警告的那种规则机器人（Pinker）。
要添加，不要只删除。编辑模式的第一步是注入风格。仅仅去除所有AI特征但依然平淡的草稿并没有被修复。
聊天不属于散文。本技巧适用于长篇内容。不要应用于快速回复。
“过度”通常只是不熟悉，而非错误。风险往往不是做得太过，而是做得不够。

Changelog

更新日志

1.0 (2026-06-10) — first release

1.0（2026-06-10）——首次发布

First real version of write-well. Lightly tested in real prose so far (a few passes), so expect iteration; the dated entries below are the pre-release development log, not a maturity claim. 1.0 ships the five-part core, the voices & personas system (6-axis fingerprint, register + author libraries, the two failure-mode guards), the deduped craft canon (Strunk, Zinsser, Orwell, Williams, Lanham, Clark, Pinker, Le Guin, McPhee, Provost, plus Ogilvy and Strunk & White 4th-ed), the 29-pattern AI-tell edit catalog, and a countable em-dash budget the skill dogfoods.

The fix that triggered cutting 1.0: write-well's output ran heavy on em-dashes. Two root causes, both addressed:

The rule was advisory and post-hoc. "Near-zero" lived in the checklist and Gotchas, a cleanup pass the model was trusted to run after drafting. Replaced with a countable generation-time budget (aim zero, cap ~2 per 1,000 words nonfiction; don't type it, then count
```
—
```
in the final pass and replace each non-interruption). New "Punctuation budget" section, hardened checklist item, and a first final-pass prompt that counts.
The skill contradicted itself. Its own SKILL.md ran ~30 em-dashes per 1,000 words, full spray, so the in-context exemplar taught the opposite of the rule. Rewrote SKILL.md to meet its own budget, and added a dogfood eval (
```
tests/eval.sh
```
counts em-dash density in the skill's files and fails if SKILL.md regresses). The voice-override exception (DFW and the like) is preserved.

Development log (pre-1.0) — what was built, and when. The V-numbers are the build passes, not shipped releases.

write-well的第一个正式版本。目前已在真实散文中进行初步测试（几次修改），所以会持续迭代；以下日期条目是预发布开发日志，并非成熟度声明。1.0版本包含五要素核心、风格与人设系统（6维度特征、语体+作者库、两种失败模式防范措施）、去重后的写作权威准则（Strunk、Zinsser、Orwell、Williams、Lanham、Clark、Pinker、Le Guin、McPhee、Provost，加上Ogilvy和Strunk & White第4版）、29种AI写作特征编辑库，以及本技巧自身遵循的可计数破折号预算。

促使发布1.0版本的修复：write-well的输出破折号使用过多。两个根本原因均已解决：

规则是建议性的，且在事后执行。“接近零”只出现在检查清单和注意事项中，是模型创作后被信任执行的清理步骤。现在替换为可计数的创作时预算（目标为零，非虚构类每1000字最多约2个；不要输入破折号，然后在最终检查时统计
```
—
```
的数量，替换掉非必要打断的破折号）。新增“标点预算”章节、更严格的检查清单条目，以及第一个统计破折号的最终检查提示。
本技巧自相矛盾。其自身的SKILL.md每1000字约有30个破折号，大量滥用，所以上下文范例传递的是与规则相反的信息。重写SKILL.md使其符合自身预算，并添加自我评估（
```
tests/eval.sh
```
统计本技巧文件中的破折号密度，如果SKILL.md违反规则则评估失败）。风格覆盖例外（比如DFW等）依然保留。

开发日志（1.0版本前）——构建内容及时间。V编号是构建版本，非发布版本。

V2.3 (2026-06-07) — Strunk & White studied in full (4th edition, the source of record)

V2.3（2026-06-07）——完整研究Strunk & White（第4版，权威来源）

Studied The Elements of Style, 4th ed. (the source provided), and drew the composition principles from it (copyrighted text consulted and paraphrased, not reproduced). Strunk was already credited five times (omit-needless-words, active voice, positive form, specific/concrete, emphatic-word-last); the dedup had hidden him. Three genuinely-new rules added, plus a resolved tension. All numbering follows the 4th edition, matching six-pager (an earlier draft used the 1918 numbers, corrected against the 4th-ed text):

Express coordinate ideas in similar form (Rule 19). New canon rule; load-bearing in lists, goals, tenets, headings.
Keep related words together (Rule 20). New; proximity shows relationship, so don't split subject from verb.
Make the paragraph the unit of composition (Rule 13). New, register-gated: exposition/argument fronts the point; the core's story-first rule still governs narrative. Precision note: the 4th ed dropped the explicit "begin with a topic sentence" (that's Strunk's 1918 Rule 9, condensed into Rule 13), so it's prescribed by name, not cited as a 4th-ed rule.
Voice vs. White's "do not inject opinion." Stated and resolved: White cuts gratuitous opinion in exposition; write-well's voice is earned, relevant stance. Register decides.
Folded "avoid a succession of loose sentences" (Rule 18) into the vary-rhythm rule. Left the Elementary Rules of Usage (Rules 1–11: comma/colon/dash/agreement/case) out, since those are correctness, not craft.
Then studied White's Chapter V ("An Approach to Style"). Finding: it adds no new positive rule; it supplies the restraint half of voice. Added a voice-discipline note (place yourself in the background, don't overwrite or overstate, no breezy manner, prefer standard to offbeat, don't explain too much) tied to the existing caricature failure mode, and sharpened draft mode's dialogue rule (no adverb-propped "said"). White independently corroborates the AI-tell catalog from the stylist's side.

研究《风格的要素》第4版（提供的来源），从中提炼写作原则（参考并改写受版权保护的文本，未复制）。Strunk此前已被引用五次（删除冗余词汇、主动语态、正向表达、具体/具象、重点词汇放最后）；去重过程中隐藏了他的贡献。新增三条真正的新规则，以及一个解决的矛盾点。所有编号遵循第4版，与六页文档一致（早期草稿使用1918年版编号，已根据第4版文本修正）：

并列观点采用相似形式表达（规则19）。新增准则；在列表、目标、原则和标题中至关重要。
相关词汇放在一起（规则20）。新增； proximity体现关联，不要将主语和动词分开。
以段落为写作单位（规则13）。新增，受语体限制：说明/论证类文本先点明观点；核心的故事优先规则仍适用于叙事类文本。精确说明：第4版删除了明确的“以主题句开头”（这是Strunk1918年版的规则9，已浓缩到规则13中），所以明确规定该要求，但不引用为第4版规则。
风格与White的“不要注入观点”的矛盾。明确并解决：White反对说明性文本中无意义的观点；write-well的风格是有依据的、相关的立场。由语体决定。
将“避免连续使用松散句”（规则18）并入变换节奏规则。省略了基础用法规则（规则1-11：逗号/冒号/破折号/一致/格），因为这些是正确性问题，而非写作技巧。
随后研究White的第五章《风格入门》。发现：没有新增正向规则；它补充了风格的克制层面。新增风格约束说明（保持低调，不要过度写作或夸张，不要轻浮，优先使用标准表达而非另类表达，不要过度解释），与现有的夸张模仿失败模式关联，并强化了draft模式的对话规则（不要用副词修饰“said”）。White从文体学家的角度独立验证了AI写作特征库。

V2.2 (2026-06-07) — Ogilvy's "How To Write" memo, incorporated three ways

V2.2（2026-06-07）——整合Ogilvy的《如何写作》备忘录，三个方面

From Ogilvy's 1982 staff memo (the 10 rules plus the Roman & Raphaelson lineage it points to). Sorted by what kind of rule each is, not bolted on whole, since the memo is mostly a business-register playbook, not new core rules:

Core canon. Ogilvy added as a cited authority (short words, no jargon already lived in the canon; he corroborates, doesn't duplicate).
Edit-process. A new cooling-off pass in the craft canon: sleep on it, then reread aloud and edit, get a colleague to improve it, check your quotations and facts. Surfaced as edit-mode step 5.
Register layer. Rules 2/5/9/10 folded into the business registers (
```
plain-professional
```
,
```
executive-brief
```
,
```
punchy-marketing
```
): write the way you talk, ≤2 pages, make the desired action unmistakable, and don't-write-when-you-should-talk as the register's boundary.
Author voice. Added
```
ogilvy
```
to the voices library (the ad-man's prose: concrete, benefit-led, zero jargon), distinct from applying his rules.

来自Ogilvy1982年的员工备忘录（10条规则及其指向的Roman & Raphaelson传承）。按规则类型分类，而非整体添加，因为备忘录主要是商务语体指南，而非新的核心规则：

核心准则。Ogilvy作为权威被引用（短词汇、无行话已在准则中；他是佐证，而非重复）。
编辑流程。在写作权威准则中新增冷却检查步骤：先放一放，然后大声朗读并修改，找同事帮忙改进，核实引文和事实。作为编辑模式的第5步呈现。
语体层面。规则2/5/9/10并入商务语体（
```
plain-professional
```
、
```
executive-brief
```
、
```
punchy-marketing
```
）：像说话一样写作，不超过2页，明确期望的行动，以及该语体的边界：该沟通时不要写作。
作者风格。在风格库中添加
```
ogilvy
```
（广告人的写作风格：具体、以利益为导向、无行话），与应用他的规则区分开。

V2.1 (2026-06-04) — register-first + register voices (real-use feedback)

V2.1（2026-06-04）——语体优先+语体风格（实际使用反馈）

From using V2 on performance feedback, where the punchy default fought the task. Four fixes:

Register & tone, set first. The punch bias (short kicks, the dramatic closer, "don't narrate the landing") is now explicitly a register tool for marketing/fiction/essay, not the default. Measured registers (feedback, professional, diplomatic, reference, academic) get even rhythm, balance, and substance-led endings.
Named the punchy short-fragment closer as its own AI tell. Don't reach for it by reflex; close on substance in measured prose.
Em-dashes → near-zero by default (a top AI tell now), not the old ~4/1,000 budget; triples tightened. Budget rises only if a
```
--like
```
sample uses them.
Register voices (
```
plain-professional
```
,
```
warm-feedback
```
,
```
diplomatic
```
,
```
executive-brief
```
,
```
conversational
```
,
```
academic
```
,
```
punchy-marketing
```
,
```
storyteller
```
), pickable by what you're writing, since the author-name voices are useless if you don't know the author. Author voices kept as a power-user library.

在绩效反馈中使用V2版本时，默认的有冲击力风格与任务冲突。进行四项修复：

语体与语气，先设定。有冲击力偏向（短句、戏剧性收尾、“不要叙述结局”）现在明确是语体工具，仅适用于营销/小说/随笔，非默认。沉稳语体（反馈、专业、外交、推荐信、学术）采用平稳节奏、平衡、内容导向的结尾。
将戏剧性短句碎片收尾列为AI写作特征。不要本能地使用它；沉稳语体中要以内容收尾。
破折号默认接近零（现在是顶级AI写作特征），而非之前的每1000字约4个的预算；三连式表达更严格。只有当
```
--like
```
样本使用破折号时，预算才会提高。
语体风格（
```
plain-professional
```
、
```
warm-feedback
```
、
```
diplomatic
```
、
```
executive-brief
```
、
```
conversational
```
、
```
academic
```
、
```
punchy-marketing
```
、
```
storyteller
```
），可根据写作类型选择，因为如果你不熟悉作者，作者名称风格毫无用处。保留作者风格作为高级用户库。

V2 (2026-06-02) — voices & personas + the craft canon

V2（2026-06-02）——风格与人设+写作权威准则

Two research-backed expansions (deep-research on the style canon plus on voice/persona/stylometry):

Voices & personas system.
```
--voice
```
/
```
--persona
```
/
```
--like
```
, built on a 6-axis voice fingerprint (diction, syntax, rhythm-variance, POV, tone/punctuation, stance) with a starter library plus 2–4 exemplars each. Engineered against the two documented failure modes: regression-to-generic (arXiv 2509.14543, "LLMs default to a generic style… stripping the personal touch") and caricature/stereotype-amplification. The persona/voice split follows the rhetoric distinction (ethos/persona is the implied author, separate from style mechanics).
The craft canon. A deduped positive ruleset across ten authorities, adding what the anti-AI references lacked: Williams' characters-as-subjects plus cohesion (old→new), Lanham's Paramedic Method, McPhee's structure law, Pinker's classic style plus curse-of-knowledge, and the rules-as-defaults-not-commandments principle (Pinker vs Strunk/Orwell: keep the positive maxims, demote the grammar superstitions). Deliberately did NOT add the no-passive/no-split-infinitive folklore.

两项基于研究的扩展（对写作准则以及风格/人设/文体计量学的深度研究）：

风格与人设系统。
```
--voice
```
/
```
--persona
```
/
```
--like
```
，基于6维度风格特征（措辞、句法、节奏变化、视角、语气/标点、立场）构建，包含初始风格库及每个风格的2-4个范例。针对两种已记录的失败模式进行优化：回归通用风格（arXiv 2509.14543，“LLMs默认采用通用风格……失去个人特色”）和夸张模仿/刻板印象放大。人设/风格的区分遵循修辞学定义（ethos/人设是隐含作者，与风格机制分离）。
写作权威准则。来自十位权威的去重正向规则集，补充了反AI参考资料缺失的内容：Williams的角色为主语+连贯性（已知→新）、Lanham的急救法、McPhee的结构法则、Pinker的经典风格+知识诅咒，以及规则是默认选项而非戒律的原则（Pinker vs Strunk/Orwell：保留正向准则，降级语法迷信）。刻意未添加禁止被动语态/拆分不定式的民间说法。

V1 (2026-05-27)

V1（2026-05-27）

Merged

human-writing

(drafting craft) plus

humanizer

(29-pattern edit catalog) into one positively-framed skill: lead with what good writing has (voice, VAKS, rhythm, specificity, story-first) and treat AI tells as the absence of those, fixed in

edit

mode. Added Gary Provost's sentence-variety/read-aloud principle. Genericized domains (no project-specific names). Sources: VAKS plus story arc (Vinh Giant); edit catalog (Wikipedia "Signs of AI writing"); rhythm (Provost).

将

human-writing

（创作技巧）和

humanizer

（29种模式编辑库）合并为一个正向框架的技巧：先关注优秀写作具备的要素（风格、VAKS、节奏、具体性、故事优先），将AI特征视为这些要素的缺失，在

edit

模式中修复。添加Gary Provost的句子变化/大声朗读原则。通用化领域（无特定项目名称）。来源：VAKS+故事弧（Vinh Giant）；编辑库（维基百科“AI写作特征”）；节奏（Provost）。

References

参考资料

references/craft-canon.md — the deduped positive ruleset from the ten authorities, with sources
references/voices.md — the 6-axis fingerprint model plus named voice/persona library
references/voice-examples.md — one passage rewritten across voices (Hemingway, McCarthy, DFW, Didion, Ogilvy, executive-brief)
references/prose-craft.md — fiction prose craft plus anti-AI rules
references/edit-catalog.md — the 29 AI-tell patterns, before/after
references/banned-words.md · references/domains.md

references/craft-canon.md — 来自十位权威的去重正向规则集，附来源
references/voices.md — 6维度特征模型+命名风格/人设库
references/voice-examples.md — 同一段落用不同风格改写（Hemingway、McCarthy、DFW、Didion、Ogilvy、executive-brief）
references/prose-craft.md — 小说写作技巧+反AI规则
references/edit-catalog.md — 29种AI写作特征，修改前后示例
references/banned-words.md · references/domains.md