同一个问题，换种说法，答案差很远。这一章你会写出五槽模板，跑一次 A/B，再把它固化成账号级「给 Claude 的指令」—— 全书第一件持久资产，写完全局生效，后面每一面都吃它。Same question, different phrasing, very different answer. In this chapter you write a five-slot template, run one A/B, then pin it as an account-level instruction for Claude — the book's first durable asset, live everywhere, and every surface ahead feeds on it.
同一个问题，换种说法，答案差很远。最便宜的一次升级，不是换更强的模型、也不是切到新的一面，是把话说清楚。Same question, different phrasing, very different answer. The cheapest upgrade isn't a stronger model or a new surface — it's saying it clearly.但「把话说清楚」只讲道理是没用的 —— 你读完点头，明天照样打那句「帮我写个周报」。所以这一章不停在道理上：你会写出一份五槽模板，拿它跟你昨天的基线跑一次 A/B，然后把它固化成账号级指令，从此每一次对话自动带上。But 'say it clearly' as advice is useless — you'll nod along and tomorrow type 'write me a weekly update' again. So this chapter doesn't stop at advice: you'll write a five-slot template, A/B it against yesterday's baseline, and then pin it as an account-level instruction that rides along on every conversation from then on.
— I
把它当聪明但缺背景的新员工Treat It as a Brilliant New Hire Who Lacks Context.
它不缺脑子，缺的是你脑子里没说出口的背景。It isn't short on brains — it's short on the context you never said out loud.Anthropic 自己的说法：把 Claude 想成一个绝顶聪明、但对你的规矩和流程一无所知的新同事。1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08.它不会读心。有个一句话的检验：把你写的那段话，拿给一个不了解这件事的同事看 —— 他要回你几个问题，Claude 就会在那几处含糊。2注 2Note 2Claude API Docs · Prompt engineering overview —— 「把你的提示拿给一个不了解任务的同事看，他困惑，Claude 多半也困惑」即出自这套指南。Claude API Docs · Prompt engineering overview — the 'show your prompt to a colleague who lacks context; if they're confused, Claude likely is too' rule of thumb comes from this guide.Anthropic's own framing: think of Claude as a brilliant new colleague who knows nothing about your norms and workflows.1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08. It can't read your mind. There's a one-line test: show what you wrote to a colleague who doesn't know the task — wherever they'd ask you a question, Claude will be just as unsure.2注 2Note 2Claude API Docs · Prompt engineering overview —— 「把你的提示拿给一个不了解任务的同事看，他困惑，Claude 多半也困惑」即出自这套指南。Claude API Docs · Prompt engineering overview — the 'show your prompt to a colleague who lacks context; if they're confused, Claude likely is too' rule of thumb comes from this guide.
缺背景No context
帮我润色一下这段。Polish this for me.
说清背景、约束、格式Context, constraints, format
把这段改简洁，读者是没技术背景的客户，保留所有数字和专有名词，控制在 200 字内。Tighten this for a non-technical client, keep every number and proper noun, under 200 words.
多花的是十几个字，省下的是三轮来回。The cost is a dozen extra words; the saving is three rounds of back-and-forth.
— II
给理由，别只给禁令Give the Reason, Not Just the Ban.
说该做什么、并说为什么，比堆一串「不要」更管用。Saying what to do — and why — beats stacking a list of 'don'ts.'两条合成一条。其一，给它你的理由，它能举一反三到你没列到的情况。官方那个例子很传神：与其说「不要用省略号」，不如说「这段会被语音朗读，省略号读不出来」—— 它就懂了，凡是朗读读不出的标点都避开，而不只是省略号。1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08.其二，说该做什么，别说不该做什么：「用要点列出」比「不要写成大段」更容易照做。1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08.Two rules in one. First, give it your reason and it generalizes to cases you didn't list. The official example lands it: instead of 'no ellipses,' say 'this will be read aloud by text-to-speech, which can't pronounce ellipses' — now it avoids every punctuation that won't read aloud, not just ellipses.1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08. Second, say what to do, not what not to do: 'use bullet points' is easier to follow than 'don't write big paragraphs.'1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08.
只给禁令Bans only
不要用省略号。不要太正式。No ellipses. Not too formal.
给理由 + 正向Reason + positive
这段会被 TTS 朗读，省略号读不出来 —— 用句号断句。读者是老朋友，写得像聊天、用「你」。This is read aloud by TTS, which can't say ellipses — break with periods. The reader's an old friend, so write like a chat, using 'you.'
还有三个最省力的杠杆：角色（一句话给它一个身份，语气和侧重立刻收窄）、例子（贴 1–3 个「长这样」的样例，比任何形容词都稳）、格式（明说散文还是要点、要不要小标题、多长）。1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08.当一条消息里既有指令又有要处理的材料时，用标签把两者切开（<instructions> 和 <context>），它就不会把材料误当成指令；长材料放最上面、你的问题放最后。1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08.Three more low-effort levers: a role (one sentence of identity snaps tone and focus into place), examples (paste 1–3 'looks like this' samples — steadier than any adjective), and a format (spell out prose vs bullets, headers or not, how long).1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08. When one message holds both instructions and material to work on, split them with tags (<instructions> and <context>) so it won't mistake the material for instructions; put long material at the top and your question last.1注 1Note 1Claude API Docs · Prompting best practices —— 当前模型的提示工程总参考：清晰直接、给动机、用示例、XML 标签分区、给角色、明确输出格式。截至 2026-08。Claude API Docs · Prompting best practices — the reference for prompting current models: be clear and direct, give motivation, use examples, separate with XML tags, give a role, specify output format. As of 2026-08.
— III
动手：五槽模板与一次 A/BDo It: Five Slots and One A/B.
把上面那些合成一件可复制的东西。Fold all of that into one thing you can copy.
提示词Prompt五槽模板 · 全书的地基The five-slot template
角色：你是<一句话身份>。
背景：<它不知道、但影响答案的关键背景>。
要做：<具体动作>，因为<理由>。
格式：<散文 / 要点>，<长度上限>，先结论后展开。
像这样：<贴 1-3 个「长这样」的例子>。Role: you are <a one-line identity>.
Context: <key background it can't know but that changes the answer>.
Do: <specific action>, because <reason>.
Format: <prose / bullets>, <length cap>, conclusion first.
Like this: <paste 1-3 'looks like this' examples>.
角色：你是一位挑剔的科技产品文案编辑。
背景：这是一封发给已付费 Pro 用户的续费提醒邮件，读者反感营销腔。
要做：把下面这段草稿改得更短更直接，因为现在读着像群发广告。
格式：要点，不超过 5 条，先给改后版本再说改了什么。
像这样：把「我们非常高兴地通知您」改成「你的 Pro 下周到期」。Role: you are a picky copy editor for tech products.
Context: this is a renewal-reminder email to paying Pro users who hate marketing-speak.
Do: make the draft below shorter and more direct, because it currently reads like a mass ad.
Format: bullets, no more than 5, give the rewrite first then say what changed.
Like this: turn 'We are thrilled to inform you' into 'Your Pro plan ends next week.'
现在拿它去打第 1 章那条基线。这一步别跳 —— 没有这个数字，你后面十章都在凭感觉。Now take it to the baseline from Chapter 1. Don't skip this step — without the number, the next ten chapters run on vibes.
01
翻出基线里那句原始提示词Dig out the original prompt from your baseline
就是你昨天那句没优化过的话。开一个新的空白对话，原样再发一次 —— 这是 A 组。记下它这次答成什么样。The unpolished sentence from yesterday. Open a fresh blank chat and send it verbatim again — that's arm A. Note how it lands this time.
02
按五槽把它重写一遍，再开一个空白对话发Rewrite it into the five slots, then send it in another blank chat
五个槽全填满，尤其是「背景」和「因为」—— 那两格通常就是你心里默认、却从没打出来的东西。这是 B 组。Fill all five slots, especially Context and the 'because' — those two are usually what you assume and never type. That's arm B.
03
数追问次数，记下差额Count the follow-ups and record the gap
两边各自还要追加几句才拿到能用的结果？差额就是「说清」这件事对你的现金价值。基线里 4 轮变成 1 轮，就是每周省三轮。How many extra messages does each arm need before the result is usable? The gap is what 'saying it clearly' is worth to you in cash. Four rounds down to one is three rounds saved every week.
04
把跨任务都成立的那几条，写进账号级指令Put the parts that hold across tasks into your account instructions
五槽里，「角色 / 背景 / 理由」是每件活不同的，别写进去；「格式」和你的固定偏好（说中文、先结论、别用某类词、代码用哪个语言）是每件活都一样的 —— 那几条搬进设置里那段全局的「给 Claude 的指令」，从此不用再打。Of the five slots, Role / Context / Reason differ per task — leave them out. Format and your standing preferences (answer in Chinese, conclusion first, avoid a certain register, code in a specific language) are the same every time — move those into the global instructions-for-Claude block in settings, and stop typing them.
提示词Prompt账号级指令 · 本章要存下的那件东西Account-level instructions · the asset to save
我是<一句话职业与领域>，主要用 Claude 做<两三类活>。
默认用<语言>回答；先给结论，再展开。
长度按题目难度自己拿捏：简单问题两三句，别铺垫。
要点用要点，别把要点写成段落；<你的固定格式偏好>。
代码默认用<语言 / 框架>，遵守<一条你团队的硬规矩>。
不确定就问我，别猜着往下做；拿不准的地方标出来，别写得像确定的。I am <one line: role and field>, and I mostly use Claude for <two or three kinds of work>.
Answer in <language> by default; conclusion first, detail after.
Calibrate length to difficulty: two or three sentences on simple questions, no wind-up.
Use bullets when the content is a list, and don't turn a list into prose; <your standing format preference>.
Code in <language / framework> by default, following <one hard rule your team has>.
When unsure, ask me rather than guessing — and mark the uncertain parts instead of writing them as settled.
我是一名做 B 端 SaaS 的独立产品经理，主要用 Claude 做需求文档、客户邮件和竞品调研。
默认用中文回答；先给结论，再展开。
长度按题目难度自己拿捏：简单问题两三句，别铺垫。
要点用要点，别把要点写成段落；不要用「赋能 / 抓手 / 闭环」这类词。
代码默认用 TypeScript，遵守「不引入新依赖除非我点头」。
不确定就问我，别猜着往下做；拿不准的地方标出来，别写得像确定的。I am a solo product manager working on B2B SaaS, and I mostly use Claude for specs, customer emails, and competitive research.
Answer in English by default; conclusion first, detail after.
Calibrate length to difficulty: two or three sentences on simple questions, no wind-up.
Use bullets when the content is a list, and don't turn a list into prose; avoid corporate filler like 'leverage', 'synergy', 'circle back'.
Code in TypeScript by default, following 'no new dependencies unless I approve them'.
When unsure, ask me rather than guessing — and mark the uncertain parts instead of writing them as settled.
— IV
换个域，同一套骨架Same Skeleton, Different Field.
五个槽不变，填法按域走。挑你自己那一栏抄。The five slots don't change; what goes in them does. Copy the tab that's yours.
角色写「一位会把难点拆成台阶的助教」；背景写你现在的水平和已经会了什么（这一格决定它从哪儿讲起）；理由写「因为我要在周五给同学讲一遍」—— 它就会自动往能复述的方向组织，而不是给你一篇正确但记不住的综述。格式写「先给一句话的直觉，再给严谨版本」。Role: 'a teaching assistant who breaks hard things into steps.' Context: your current level and what you already know — that slot decides where it starts. Reason: 'because I have to explain this to classmates on Friday' — and it will organize toward something retellable instead of a correct, unmemorable survey. Format: 'give the one-sentence intuition first, then the rigorous version.'
角色写「一位苛刻的同行评审」；背景写这个领域的共识和争议在哪；理由写「因为我要投的期刊会挑方法论的刺」。格式那格最关键：明说「证据冲突就并列，别替我调和」和「拿不准的单独列一节」—— 不写这两句，你拿到的会是一份读着顺、但把分歧抹平了的东西。Role: 'an exacting peer reviewer.' Context: where the consensus and the disputes sit in this field. Reason: 'because the journal I'm submitting to will attack the methodology.' The Format slot matters most here: say outright 'where evidence conflicts, lay both sides out — don't reconcile them for me' and 'put anything uncertain in its own section.' Without those two lines you get something that reads smoothly and has quietly flattened the disagreement.
角色写具体到人：「一位写惯了开发者邮件的文案」比「一位专业文案」有用十倍。背景写读者是谁、他们烦什么。理由是这一栏的胜负手 —— 「因为上一封的打开率只有 11%，读者反馈说太像群发」比「因为要写好一点」精确得多。例子那格贴一封你真的满意的旧邮件，胜过一百个形容词。Make the role a specific person: 'a copywriter who writes developer emails all day' beats 'a professional copywriter' tenfold. Context: who the reader is and what annoys them. The reason is the decisive slot here — 'because the last one got an 11% open rate and readers said it felt like a blast' is far sharper than 'because it should be better.' In the examples slot, paste one old email you were genuinely happy with; it beats a hundred adjectives.
角色写「一位替我起草、但不替我决策的幕僚」—— 这句能挡住它最常见的越界。背景写团队规模、汇报对象、这件事的政治敏感度。理由写「因为这份要在周会上念，念错一个数字比慢一天贵」。格式加一条硬约束：「所有数字必须来自我给的材料，你不确定的写『待确认』，不要推算」。Role: 'a chief of staff who drafts for me but doesn't decide for me' — that one line blocks its most common overreach. Context: team size, who you report to, how politically sensitive this is. Reason: 'because this gets read aloud at the weekly meeting, and one wrong number costs more than one lost day.' Add a hard format constraint: 'every number must come from the material I gave you; write TBC where you're unsure — do not extrapolate.'
共同点：每一栏的「因为」都不是客套，是真的会改变答案 —— 这是最容易被跳过、也最值钱的一格。What they share: the 'because' in every tab isn't politeness — it genuinely changes the answer. It's the slot people skip most and the one worth the most.
— V
边界：什么时候不该写长 promptBoundaries: When Not to Write a Long Prompt.
五个槽不是每次都要填满 —— 有几种情况，填满反而更差。You don't fill all five slots every time — in a few situations, filling them makes things worse.
你自己也没想全You haven't thought it through
任务一复杂，硬写一份面面俱到的指令，等于把你的糊涂精确地传达给它。这时反过来 —— 让它先问你。下面那个模板就是干这个的。On a complex task, forcing a comprehensive brief just transmits your own fog with precision. Flip it — make it ask you first. The template below does exactly that.
一次性的简单查询A one-off simple lookup
「这个词英文怎么说」不需要角色和格式。当前模型会按题目难度自校准回答长度，简单题它自己就短。<Footnote n={4} />给简单题套模板，是给自己加班。'How do you say this in English' needs no role and no format. Current models calibrate response length to task difficulty, so simple questions already come back short.<Footnote n={4} /> Templating a trivial ask is unpaid overtime for yourself.
它已经答对了It already got it right
别为了「用上模板」去改一句本来就管用的提示。这本书的每一件资产都只在有落差的地方才划算 —— 没落差就别装。Don't rewrite a prompt that already works just to use the template. Every asset in this book only pays off where there's a gap; where there isn't, don't build.
账号级指令写太长Account instructions written too long
这段是每次对话都要带上的。写成一篇规章，你会发现它开始在不相干的问题上照搬那些规矩。控制在六到八行，只放跨任务都成立的偏好。This block rides along on every single conversation. Write it as a rulebook and you'll start seeing those rules applied to unrelated questions. Keep it to six or eight lines, and only preferences that hold across tasks.
提示词Prompt让它先问、先计划、先列风险Make it ask, plan, flag — first
这是个略复杂的任务，别急着做。先按这三步来：
1. 把你不清楚的地方列成问题问我，一次问完。
2. 给我一个分几步做的计划，等我点头。
3. 顺手说一句最大的风险或最容易出错的地方。
我确认完，你再动手。
任务：<把你的任务贴这里>。This is a somewhat complex task — don't rush in. Do these three first:
1. List what's unclear to you as questions and ask me, all at once.
2. Give me a step-by-step plan and wait for my go-ahead.
3. Add one line on the biggest risk or the easiest thing to get wrong.
Once I've confirmed, then start.
Task: <paste your task here>.
这是个略复杂的任务，别急着做。先按这三步来：
1. 把你不清楚的地方列成问题问我，一次问完。
2. 给我一个分几步做的计划，等我点头。
3. 顺手说一句最大的风险或最容易出错的地方。
我确认完，你再动手。
任务：把我们这季度的 12 份用户访谈整理成一份带主题归类的洞察报告。This is a somewhat complex task — don't rush in. Do these three first:
1. List what's unclear to you as questions and ask me, all at once.
2. Give me a step-by-step plan and wait for my go-ahead.
3. Add one line on the biggest risk or the easiest thing to get wrong.
Once I've confirmed, then start.
Task: turn our 12 user interviews from this quarter into an insight report grouped by theme.
— VI
收束 · 本章验收Sign-off.
五槽模板已写成一段可复制的文本，五个槽都填过一次真内容。The five-slot template exists as copyable text, with all five slots filled once with real content.
跑完了 A/B：旧提示词一次、新模板一次，两边的追问次数都记下来了。Ran the A/B: the old prompt once, the new template once, with follow-up counts recorded for both.
新模板至少少追问一轮。没少的话，回去看「背景」和「因为」那两格是不是还空着。The new template needed at least one fewer follow-up. If it didn't, go back and check whether Context and 'because' are still empty.
跨任务都成立的偏好已写进账号级「给 Claude 的指令」，六到八行以内。The cross-task preferences are written into the account-level instructions for Claude, within six to eight lines.
新开一个空白对话随便问一句，确认语气和格式已经是你要的 —— 这是那段指令真的生效的证据。Opened a blank chat, asked anything, and confirmed the tone and format already match — that's your proof the instruction is live.
这段账号级指令是全书第一件持久资产，也是唯一一件在每一面都生效的。后面每一章的产出 —— Project 指令、Skill 说明、Cowork 的交办 brief —— 都是这套五槽在不同位置上的变体。That account-level block is the book's first durable asset and the only one that's live on every surface. Every later chapter's output — the Project instruction, the Skill description, the Cowork brief — is this same five-slot shape wearing a different hat.
把背景说出口，
它就不用猜.
Say the context out loud,
and it stops guessing.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return
