① 工程：动手写 AI agent 系统前先写设计文档——文档不是给"已完成的设计"留档，而是设计在被写出来时才暴露出它是坏的：当你被迫把多个 agent 的调用时序逐步写成句子，那个在脑中一直被跳过的竞态条件才终于现形。② 育儿：给学龄孩子讲清一个概念前，先逼自己用三句话写下来——你会精确地发现自己理解里最虚的那块，往往正是孩子会卡住的那块。写不出，等于还没真懂。
English Summary
Writing Is Thinking — writing isn't transcribing finished thoughts; it's the process that makes fuzzy thought precise. Before you write, you feel you understand — that's the illusion of explanatory depth. Writing breaks it by forcing serialization: thought is a parallel associative graph, prose is a single line, and flattening the graph forces you to answer what your intuition glossed over — what comes first, what causes what, where the chain breaks. The page is an external working memory that doesn't decay (your head holds only ~4 chunks), a compiler that won't pass vague sentences, and an append-only log you can replay. Practice: write to discover, not to record — let the first draft be ugly, and split "generate" from "edit."
AI Prompts
中文提示词
这是我对 [主题/决策] 的初步想法（粗糙、未整理）：[贴上你的草稿/要点]。
请把它当成"我在纸上思考"，帮我用写作戳破理解的错觉：
① 指出 2-3 处我以为讲清了、其实有逻辑断裂或缺定义的地方；
② 对每一处，提出一个我必须回答才能想清楚的尖锐问题；
③ 不要替我润色成漂亮成稿，目标是让我看见自己还没想透的地方。
English Prompt
Here are my rough, unsorted thoughts on [topic/decision]: [paste your draft/notes].
Treat this as me "thinking on paper" and use writing to break my illusion of understanding:
1. Point out 2–3 places I think I explained but where the logic actually breaks or a term is undefined.
2. For each, pose one sharp question I must answer before I can think it through.
3. Don't polish it into clean prose — the goal is to expose what I haven't actually figured out.
① 工程：技术评审里，一个逻辑改动一个 commit、一句话一个论断。一个改了三件事的 200 行 commit，正是散文里那种塞了三个想法的长句——评审者被迫并行解析，看不动。把它拆成三个小 commit，等于把长句拆成短句。② 育儿：给学龄孩子下指令，"先收玩具，然后洗手，再来吃饭"一次说完，孩子的工作记忆装不下三步，多半只执行最后一步。一次一件，做完再说下一件——和写作完全同理。你不替读者/听者切小，认知超载就由他们买单。
English Summary
One Idea at a Time — one idea per sentence, one point per paragraph. The bottleneck isn't the writer but the reader's working memory (~4 chunks). A sentence packing three ideas forces the reader to parse, buffer, and relate them at once → overload → reread, and throughput collapses. Key idea: complexity is conserved — tangles you don't resolve as writer are dumped on every reader, multiplied by readership (compress at write-time, or pay decompression cost on every read). The idea boundary is the unit of revision: a sentence that resists splitting often signals a still-tangled thought. Follow the given-new contract — start each sentence with old information (an anchor), end with the new. Test: can the reader restate the sentence in one short clause?
AI Prompts
中文提示词
这是我写的一段文字：[贴上段落]。
请按"一次只说一件事"帮我做认知带宽审计：
① 找出装了 2 个以上想法、会让读者回读的句子，逐句标出来；
② 把每个这样的句子拆成若干"一句一事"的短句，并尽量让每句以旧信息开头、新信息结尾；
③ 指出这一段到底想证明哪一个论点——如果它其实在证两件事，建议怎么拆成两段。
English Prompt
Here is a passage I wrote: [paste paragraph].
Audit it for reader cognitive bandwidth, using "one idea at a time":
1. Flag every sentence carrying 2+ ideas that would force a reread.
2. Split each into "one idea per sentence" clauses, and where possible start each with old info and end with new.
3. Name the single point this paragraph is meant to prove — if it's actually proving two, suggest how to split it into two paragraphs.
"Mistakes were made（出现了一些失误）"是英语世界政治辞令的标本——承认了有错，却让犯错的人神秘缺席。乔治·奥威尔在《政治与英语》中早已点破：含混的被动语态加上抽象名词，是政治语言用来掩盖责任、麻痹判断的标准工具。语态在这里不是文采问题，是诚实问题——主动语态会逼出那个被刻意省略的主语。
场景 · BigCat
① 工程：故障复盘里，"数据库被打挂了"和"重试逻辑反复猛击数据库导致它过载"是两份不同的文档——前者没有可修的对象，后者直接指向施事者、也就指向了修复点。被动语态的复盘读起来像没人有错，于是也没人会改。② 育儿：示范诚实的能动性。"花瓶碎了"把孩子（和你自己）的手从句子里抹掉了；如实说出"谁、做了、什么"，才是在教责任。主动语态是写作层面的"风险共担"——把名字留在句子里。
English Summary
Active Voice — active voice is "actor → action → object," matching the causal structure of reality and how the mind models events. Passive voice hides the agent ("mistakes were made") — which is exactly why it's a tell. Voice is not just style but an ontological choice: active forces you to name who does what, surfacing hidden assumptions and missing actors. In a design doc, passive is where bugs hide — "the data is processed" omits which service, when, with what guarantee (a dangling edge in the causality graph). Active is usually shorter and faster to parse. But nuance: passive is right when the object is the topic or the actor is unknown/irrelevant ("the sample was heated to 80°C"). The rule is "default active, deliberate passive," not "ban passive."
AI Prompts
中文提示词
这是我写的文字（如设计文档/复盘/通知）：[贴上文字]。
请帮我做"语态与责任"审计：
① 找出所有被动句，逐句标出被隐去的施事者是谁；
② 对每一处判断：藏掉这个"谁"是有意的（受事才是话题）还是偷懒？偷懒的改成主动；
③ 特别标出那些"让没人需要负责"的句子——把它们改写成指名道姓、能指向修复点的主动句。
English Prompt
Here is something I wrote (e.g., a design doc / postmortem / announcement): [paste text].
Run a "voice and accountability" audit:
1. Find every passive sentence and name the agent it hides.
2. For each, judge whether hiding the actor is deliberate (the object is genuinely the topic) or just lazy — rewrite the lazy ones as active.
3. Flag sentences that make "no one responsible," and rewrite them as active sentences that name the actor and point to a fixable target.
非平凡点：① 减比加难，有两个原因。其一，减是隐形劳动——加东西有新成果可展示，删东西看不出你干了活。其二，损失厌恶——你在杀自己亲手生出来的字（"kill your darlings"，忍痛删掉最得意的句子）。② 最深的删除不是删字，是删想法：最勇敢的一刀往往砍掉你最自豪、却不服务读者主线的那一整段。判准永远是读者的路径，不是你的不舍。③ 减法会复利：删掉最弱一环，既抬高平均水准，又移除一个失效点——这正是否定之道（via negativa）、YAGNI、奥卡姆剃刀共享的逻辑：通过移除而变强，比添加更稳。④ AI 时代这条模型权重在上升：生成已廉价到无限，稀缺的能力从"写得出"转向"敢删、会删"——把模型吐出的洋洋洒洒砍到只剩骨头。
① 工程：最好的 PR 常常是净删行的——删掉一个功能、一个配置项，等于一次性移除它带来的维护成本和整片 bug 暴露面。"今天我让代码库变短了"往往比"我加了个功能"更有价值。② AI 工作流：你让模型起草，真正决定质量的是你之后那把删减的刀——把三页删到半页，留下的密度才是你的判断力。③ 沟通：少说，未说的部分自带分量。对文字、对代码、对话语，减法都是被严重低估的那半门技艺。
English Summary
Subtraction Beats Addition — our default instinct when improving anything is to add; research shows people overwhelmingly add rather than subtract. Yet writing improves mostly by cutting. Every word spends the reader's attention budget, so a word that doesn't earn its place is a net tax on every reader ("omit needless words"). Subtraction is harder for two reasons: it's invisible labor (nothing new to show), and it's loss-averse (you're killing words you birthed — "kill your darlings"). The deepest cut isn't words but ideas: deleting the paragraph you're proudest of because it doesn't serve the reader's path. Subtraction compounds — removing the weakest link both raises the average and removes a failure mode, the shared logic of via negativa, YAGNI, and Occam. In the AI era, when generation is free, the scarce skill shifts from producing to cutting. Practice: write long, cut 30%, edit in a dedicated deletion pass.
AI Prompts
中文提示词
这是我写的文字：[贴上文字]。请戴上"减法的眼睛"帮我删，目标是砍掉约 30% 而不损失任何读者真正需要的信息：
① 逐句标出可删的冗词、套话、空过渡，给出删后版本；
② 找出 1-2 处"我可能最得意、但其实不服务主线"的句子或段落，建议忍痛删掉，并说明删了为何更强；
③ 给出精简后的全文，并告诉我字数减少了多少。
English Prompt
Here is something I wrote: [paste text]. Put on a "deletion mindset" and cut it by ~30% without losing anything the reader genuinely needs:
1. Mark deletable filler, clichés, and empty transitions sentence by sentence, and show the trimmed version.
2. Identify 1–2 passages I'm probably proudest of that don't serve the main line — recommend cutting them and explain why the piece is stronger without them.
3. Return the tightened full text and tell me how many words it dropped.