这本书不讲 Claude 有哪些功能，讲你怎么把它装配起来 —— 每章结束，你的账号里多一件持久的东西。开工只做两件事：给一件真活留下「重写前」的样子，和列一张十二格的装配清单。This book isn't about what Claude can do — it's about assembling it. Every chapter leaves one durable thing behind in your account. Getting started takes two steps: capture what one real task looks like before any of this, and draw up a twelve-slot build list.
你大概是这样用 Claude 的：打开网页，在那个输入框里打字，读回答，再打字。一年下来几千条对话，几乎全在同一个框里。Here's probably how you use Claude: open the web page, type into that one input box, read the reply, type again. Thousands of conversations over a year, nearly all in the same box.这本书不打算再给你一份功能清单。清单你在官网见过了 —— 见过之后你什么也没做。这本书要你动手：十二章，每章结束，你的账号里多一件持久的东西。读完的标志不是「我知道了」，是你手上真的有了那十二件。This book isn't going to hand you another feature list. You've seen the feature list on the vendor site — and afterward you did nothing. This book asks you to build: twelve chapters, and after each one your account holds one more durable thing. The mark of having finished isn't 'I know this now,' it's that you actually have those twelve things.
— I
广度不是福利，是税Breadth Isn't a Perk, It's a Tax.
你买的是一整套产品，用的是其中一个聊天框。You bought a whole set of products and use one chat box.截至 2026-08，一份 Pro 是每月 $20（年付折合 $17）。1注 1Note 1Anthropic · Claude 套餐与价格页 —— Pro 月付 $20、年付折合 $17，Max 从 $100 起；Free 档现含 memory、连接器、web search、文件生成、桌面扩展与 Skills。截至 2026-08。Anthropic · Claude Plans & Pricing — Pro at $20/mo (or $17/mo billed annually), Max from $100/mo; the Free tier now includes memory, connectors, web search, file creation, desktop extensions, and Skills. As of 2026-08.这 $20 不只买了网页聊天 —— 它同时打开了把文件和指令钉住的 Projects、自动记住你习惯的 Memory、把重复提示沉淀成方法的 Skills、让 Claude 读你 Gmail 和 Drive 的连接器、能直接交出真 Excel 的内置技能、在桌面与云端替你跑完整件活的 Cowork，还有终端里的 Claude Code。3注 3Note 3Claude 帮助中心 · 上手 Claude Cowork —— 描述结果、走开、回来拿成品；对 Pro / Max / Team / Enterprise 开放，覆盖桌面端、网页与移动端，会话跑在云端（合上笔记本也会继续）。截至 2026-08。Claude Help Center · Get started with Claude Cowork — describe an outcome, step away, come back to finished work; available to Pro / Max / Team / Enterprise across desktop, web, and mobile, with sessions running in the cloud (they continue if you close your laptop). As of 2026-08.背后是同一组模型：Claude Opus 5、Sonnet 5、Haiku 4.5。2注 2Note 2Claude API Docs · Models overview —— 当前阵容 Claude Opus 5（$5/$25 每百万 token，1M 上下文）、Claude Sonnet 5（$3/$15，1M）、Claude Haiku 4.5（$1/$5，200K），以及面向长时程 agent 的 Claude Fable 5（$10/$50，2026-06-09 起 GA）。Opus 4.8 及更早已列入 legacy。截至 2026-08。Claude API Docs · Models overview — the current lineup is Claude Opus 5 ($5/$25 per Mtok, 1M context), Claude Sonnet 5 ($3/$15, 1M), Claude Haiku 4.5 ($1/$5, 200K), plus Claude Fable 5 for long-horizon agent work ($10/$50, GA since 2026-06-09). Opus 4.8 and earlier are now listed as legacy. As of 2026-08.As of 2026-08, a Pro plan is $20 a month ($17 billed annually).1注 1Note 1Anthropic · Claude 套餐与价格页 —— Pro 月付 $20、年付折合 $17，Max 从 $100 起；Free 档现含 memory、连接器、web search、文件生成、桌面扩展与 Skills。截至 2026-08。Anthropic · Claude Plans & Pricing — Pro at $20/mo (or $17/mo billed annually), Max from $100/mo; the Free tier now includes memory, connectors, web search, file creation, desktop extensions, and Skills. As of 2026-08. That $20 doesn't only buy web chat — it also opens Projects (which pin your files and instructions), Memory (which learns your habits on its own), Skills (which fold repeated prompts into method), connectors (which let Claude read your Gmail and Drive), the built-in skills that hand you a real spreadsheet, Cowork (which runs whole tasks for you on the desktop and in the cloud),3注 3Note 3Claude 帮助中心 · 上手 Claude Cowork —— 描述结果、走开、回来拿成品；对 Pro / Max / Team / Enterprise 开放，覆盖桌面端、网页与移动端，会话跑在云端（合上笔记本也会继续）。截至 2026-08。Claude Help Center · Get started with Claude Cowork — describe an outcome, step away, come back to finished work; available to Pro / Max / Team / Enterprise across desktop, web, and mobile, with sessions running in the cloud (they continue if you close your laptop). As of 2026-08. and Claude Code in the terminal. Behind all of it is the same set of models: Claude Opus 5, Sonnet 5, Haiku 4.5.2注 2Note 2Claude API Docs · Models overview —— 当前阵容 Claude Opus 5（$5/$25 每百万 token，1M 上下文）、Claude Sonnet 5（$3/$15，1M）、Claude Haiku 4.5（$1/$5，200K），以及面向长时程 agent 的 Claude Fable 5（$10/$50，2026-06-09 起 GA）。Opus 4.8 及更早已列入 legacy。截至 2026-08。Claude API Docs · Models overview — the current lineup is Claude Opus 5 ($5/$25 per Mtok, 1M context), Claude Sonnet 5 ($3/$15, 1M), Claude Haiku 4.5 ($1/$5, 200K), plus Claude Fable 5 for long-horizon agent work ($10/$50, GA since 2026-06-09). Opus 4.8 and earlier are now listed as legacy. As of 2026-08.问题不是你懒。问题是每多一个面，就多一次「要不要花十分钟学」的决定 —— 而你手头有正事，于是默认全部跳过，退回那个最熟的框。所以广度不是会员的福利，是会员的税。The problem isn't that you're lazy. It's that every extra surface adds one more 'is this worth ten minutes' decision — and you have real work to do, so you skip all of them and retreat to the box you know best. Breadth isn't the membership's perk; it's its tax.减税的办法只有一个：不要「了解」它们，要「装配」它们。了解是零成本零收益的，装配是一次成本、长期收益的。There's only one way to cut that tax: don't learn about them — assemble them. Learning about costs nothing and returns nothing; assembling costs once and pays out for as long as you keep the account.
— II
十二章，十二件东西Twelve Chapters, Twelve Things.
每一章的验收标准，都是一件留在你账号里的东西。Every chapter's acceptance criterion is one thing left behind in your account.下面这张表就是这本书的全部承诺。它同时是目录、是进度条，也是你在第 12 章要逐项打勾的验收单。The list below is this book's entire promise. It doubles as the table of contents, the progress bar, and the sign-off sheet you'll tick item by item in Chapter 12.
01一条基线记录One baseline record装配之前的你You, before any of this
02一份账号级提问模板An account-level ask template全局生效，后面每一面都吃它Live everywhere; every later surface feeds on it
03一张三行模型路由表A three-row model routing table哪类活走哪个模型Which work goes to which model
04一个真 Project + 一段校过的 MemoryA real Project + proofread Memory不用再重贴背景No more re-pasting background
05一个你亲手写的 SkillA Skill you wrote yourself附一份内置技能产出的真 ExcelPlus a real spreadsheet from a built-in skill
06一个接好的连接器One connected connector上下文从源头直取Context pulled from the source
07一个发布过的 ArtifactA published Artifact有链接，别人打开就能用A link others can open and use
08一件交办并验收过的 Cowork 任务A Cowork job handed off and signed off外加一个定时任务Plus one scheduled run
09一份 Research 报告 + 一张核对表A Research report + a citation check十条引用里几条站得住How many of ten citations hold
10一份 CLAUDE.md + 一次跑通的循环A CLAUDE.md + one closed loop不碰代码可跳过Skippable if you never touch code
11一本墙的账 + 一个档位决定A ledger of walls + a tier decision为撞的墙付钱，不为「更高级」Pay for the wall, not for prestige
12一张验收单 + 一条维护习惯A sign-off sheet + a maintenance habit回到第 1 章那条基线Back to the baseline from Chapter 1
十二件资产。它们有依赖关系：第 2 章的模板会被第 4 章的 Project 指令吃掉，第 4 章的 Project 是第 5 章 Skill 的触发场，第 5 章的 Skill 会被第 8 章的 Cowork 调用。按顺序做最省力。Twelve assets, and they depend on each other: Chapter 2's template gets absorbed into Chapter 4's Project instruction, Chapter 4's Project is where Chapter 5's Skill gets triggered, and Chapter 5's Skill is what Chapter 8's Cowork reaches for. In order is the cheapest path.
砍掉的是什么，也说清楚。Agent SDK、Managed Agents、在 Bedrock / Vertex / Foundry 上的部署、SCIM 与合规 API —— 那些是给开发者和企业 IT 的。Claude Design（研究预览，做设计稿与原型）4注 4Note 4Anthropic · Claude Design（Anthropic Labs）—— 2026-04-17 起研究预览，对 Pro / Max / Team / Enterprise 开放，做设计稿、原型与 slides，可导出 Canva / PDF / PPTX / HTML。Anthropic · Claude Design (Anthropic Labs) — a research preview since 2026-04-17 for Pro / Max / Team / Enterprise, producing mockups, prototypes, and slides, exportable to Canva / PDF / PPTX / HTML.在第 7 章只占一节，因为它还没稳定到值得你现在投入十分钟。Claude Science（面向科研的工作台，预置基因组学、蛋白组学等）5注 5Note 5Anthropic · Claude Science —— 面向科研的工作台，beta，macOS 与 Linux，对 Pro / Max / Team / Enterprise 开放；预置基因组学、单细胞、蛋白组学与化学信息学，接 60+ 科学数据库。截至 2026-08。Anthropic · Claude Science — an AI workbench for scientists, in beta on macOS and Linux for Pro / Max / Team / Enterprise; pre-configured for genomics, single-cell, proteomics, and cheminformatics, backed by 60+ scientific databases. As of 2026-08.整本不讲 —— 它很好，只是它的读者不是「被产品广度淹没的普通付费用户」。它们都不在你这十分钟里。What gets cut is named too. The Agent SDK, Managed Agents, deploying on Bedrock / Vertex / Foundry, SCIM and the compliance API — those are for developers and enterprise IT. Claude Design (a research preview for mockups and prototypes)4注 4Note 4Anthropic · Claude Design（Anthropic Labs）—— 2026-04-17 起研究预览，对 Pro / Max / Team / Enterprise 开放，做设计稿、原型与 slides，可导出 Canva / PDF / PPTX / HTML。Anthropic · Claude Design (Anthropic Labs) — a research preview since 2026-04-17 for Pro / Max / Team / Enterprise, producing mockups, prototypes, and slides, exportable to Canva / PDF / PPTX / HTML. gets one section in Chapter 7 and no more, because it isn't settled enough to deserve ten minutes of yours today. Claude Science (a research workbench pre-configured for genomics, proteomics, and the like)5注 5Note 5Anthropic · Claude Science —— 面向科研的工作台，beta，macOS 与 Linux，对 Pro / Max / Team / Enterprise 开放；预置基因组学、单细胞、蛋白组学与化学信息学，接 60+ 科学数据库。截至 2026-08。Anthropic · Claude Science — an AI workbench for scientists, in beta on macOS and Linux for Pro / Max / Team / Enterprise; pre-configured for genomics, single-cell, proteomics, and cheminformatics, backed by 60+ scientific databases. As of 2026-08. isn't covered at all — it's good, its reader just isn't 'an ordinary paying user drowning in product breadth.' None of them are in your ten minutes.
— III
动手：把「之前」留下来Do It: Capture the Before.
没有基线的改进，都是感觉。Improvement without a baseline is just a feeling.第 12 章要你把同一件活再跑一遍，和今天这一次比。所以今天这一次必须留档 —— 原样的、没优化过的、带着你平时那些毛病的一次。别现在就把提示词写好，那会毁掉整本书唯一的对照组。Chapter 12 will ask you to run the same task again and compare it against today's run. So today's run has to be on the record — as-is, unoptimized, carrying all your usual bad habits. Don't tidy the prompt now: that would destroy the only control group this book has.
01
挑一件你每周都做、且做完有个成品的活Pick a task you do weekly that ends in a deliverable
要求就两条：每周至少一次，和「做完了」有客观标准（一封发出去的邮件、一张表、一段能跑的代码、一份交给别人的报告）。纯思考类的活不适合当基线 —— 你没法证明它变好了。Two requirements only: at least weekly, and 'done' has an objective test (an email you send, a table, code that runs, a report someone else receives). Pure thinking tasks make bad baselines — you can't prove they improved.
02
用你平时的说法跑一次，别修Run it your usual way — don't fix it
打开一个普通对话，用你平时会打的那句话问。它答得不好也别追问、别补充 —— 追问一次就不是基线了。就要这个「不好」的样子。Open an ordinary chat and ask it the way you normally would. If the answer is bad, don't follow up and don't add context — one follow-up and it's no longer a baseline. The bad version is the point.
03
把这四样存进一个文件Save these four things into one file
原始提示词、它的回答（截短也行）、你还得自己补几轮才拿到能用的结果、以及从开口到拿到成品花了多少分钟。存成一个 md 文件或一条便签，第 12 章要用。The original prompt, its answer (truncated is fine), how many extra rounds you needed before the result was usable, and how many minutes passed from asking to having the deliverable. Save it as one markdown file or a single note — Chapter 12 needs it.
第三步嫌麻烦的话，让它替你记：把下面这段贴在那次对话的最后，它会把基线整理成一段可存档的结构。If step three feels like a chore, have it record the baseline for you: paste this at the end of that conversation and it will format the whole thing into one archivable block.
提示词Prompt基线存档Baseline archive
把我们刚才这一轮整理成一段基线记录，用于日后对照，别做任何优化建议。
照这个结构给我，纯文本，我要直接存档：
- 日期：<今天>
- 任务：<一句话说清这件活>
- 原始提示词：<原样复制我第一句话，一个字别改>
- 首答问题：<你自己看，这次回答缺了什么、跑偏在哪，三条以内>
- 追加轮次：<我一共又问了几次才够用>
- 耗时：<我说的分钟数>
最后一行留空，写「重跑结果：」，我以后填。Turn the exchange we just had into a baseline record for later comparison. Do not suggest any improvements.
Use this structure, plain text — I'm archiving it verbatim:
- Date: <today>
- Task: <one line describing this piece of work>
- Original prompt: <copy my first message exactly, change nothing>
- Problems in the first answer: <your own read — what was missing or off, three items max>
- Follow-up rounds: <how many more times I had to ask before it was usable>
- Minutes elapsed: <the number I give you>
Leave a final empty line reading 'Re-run result:' for me to fill in later.
把我们刚才这一轮整理成一段基线记录，用于日后对照，别做任何优化建议。
照这个结构给我，纯文本，我要直接存档：
- 日期：2026-08-05
- 任务：把本周三个项目的进展写成一封给客户的周报邮件
- 原始提示词：原样复制我第一句话（「帮我写个周报」），一个字别改
- 首答问题：你自己看，这次回答缺了什么、跑偏在哪，三条以内
- 追加轮次：我一共又问了 4 次才够用
- 耗时：22 分钟
最后一行留空，写「重跑结果：」，我以后填。Turn the exchange we just had into a baseline record for later comparison. Do not suggest any improvements.
Use this structure, plain text — I'm archiving it verbatim:
- Date: 2026-08-05
- Task: turn this week's progress on three projects into a client-facing weekly email
- Original prompt: copy my first message exactly ('write me a weekly update'), change nothing
- Problems in the first answer: your own read — what was missing or off, three items max
- Follow-up rounds: 4 more times before it was usable
- Minutes elapsed: 22
Leave a final empty line reading 'Re-run result:' for me to fill in later.
— IV
挑哪件活当基线Which Task to Baseline.
挑错了活，后面十一章都会觉得跟自己无关。Pick the wrong task and the next eleven chapters will feel like they're about someone else.按你的域挑，下面四个各给一件够典型、也够能验收的：Pick by your field — here are four that are typical enough and testable enough:
把一周读的三五篇论文或长文，整理成一份带出处的笔记。成品是那份笔记；「做对了」的标准是每条结论都能点回原文。这件活会在第 4 章（Project 知识库）、第 9 章（Research 与引用核对）各被打一次。Turn the three to five papers or long reads from this week into one sourced note. The deliverable is the note; 'done right' means every conclusion clicks back to its source. This task gets hit twice later — in Chapter 4 (a Project knowledge base) and Chapter 9 (Research and citation checking).
把一段口语化草稿改成能发出去的产品文案。成品是定稿；「做对了」的标准是不用你再改语气。这件活在第 2 章（说清）和第 5 章（写成 Skill）会有最大落差 —— 它是最适合做成 Skill 的一类。Turn a chatty draft into product copy you'd actually publish. The deliverable is the final text; 'done right' means you don't have to fix the tone yourself. This one shows the widest gap in Chapter 2 (phrasing) and Chapter 5 (folding it into a Skill) — it's the single best candidate for a Skill.
把某个客户这个季度散在邮件和会议纪要里的承诺，汇成一张跟进表。成品是那张表；「做对了」的标准是每一行都能指回一封邮件。第 6 章（连接器）会让这件活的耗时掉得最狠。Pull one client's commitments this quarter out of scattered emails and meeting notes into a single follow-up table. The deliverable is the table; 'done right' means every row points back to an email. Chapter 6 (connectors) is where the time on this task drops hardest.
把一个文件夹里的零散材料（发票、报名表、反馈）整理成一份带合计的表格或一摞排好版的文档。成品是那份文件；「做对了」的标准是数字对得上。这件活是第 8 章（Cowork）的教科书案例。Turn a folder of scattered material (invoices, sign-up forms, feedback) into one spreadsheet with totals or a stack of formatted documents. The deliverable is the file; 'done right' means the numbers reconcile. This is the textbook case for Chapter 8 (Cowork).
共同点：都有一个别人会看到的成品，都每周至少一次，都能用「返工几轮」来度量。What they share: each ends in something another person sees, each happens at least weekly, and each can be measured in rounds of rework.
— V
边界：什么不该当基线Boundaries: What Makes a Bad Baseline.
有几类活看着很典型，做基线却什么也证明不了。A few kinds of task look typical and prove nothing as a baseline.
没有客观成品的No objective deliverable
「陪我想想产品方向」这种。它可能真的帮了你，但你没法在第 12 章证明它变好了 —— 换成有成品的活。Things like 'help me think through product direction.' It may genuinely have helped, and you'll have no way to show it improved in Chapter 12. Swap in something that ends in an artifact.
一次性的One-off
这个月才做一次的活，装配它是负收益。这本书的每一件资产都靠重复摊薄成本，一次性的活不值得装配。Something you'll do once this month is a net loss to assemble. Every asset in this book amortizes over repetition; a one-off never earns it back.
含敏感数据的Carrying sensitive data
真实客户姓名、合同金额、身份信息 —— 基线要跑很多遍、还要在第 12 章重跑，脱敏一次比事后清理省事得多。Real client names, contract amounts, identity data — a baseline gets run repeatedly and re-run in Chapter 12; scrubbing it once now is far cheaper than cleaning up later.
你已经做得很好的Ones you already do well
如果你现在就一遍过、零返工，装配它的收益接近零。挑那件让你烦躁的 —— 落差最大的地方，收益也最大。If you already get it in one pass with zero rework, assembling it returns almost nothing. Pick the one that annoys you — the widest gap is where the payoff is.
— VI
收束 · 本章验收Sign-off.
挑定了一件每周至少做一次、且有客观成品的活。Picked one task you do at least weekly that ends in an objective deliverable.
用平时的说法原样跑了一次，没有优化提示词、没有追问。Ran it once your usual way — no polished prompt, no follow-ups.
存下了四样：原始提示词、它的回答、追加轮次、总耗时。Saved four things: the original prompt, its answer, the number of follow-up rounds, and total minutes.
确认这段基线里没有真实敏感信息。Confirmed the baseline contains no real sensitive information.
知道这份存档要留到第 12 章 —— 别删。Know this archive has to survive until Chapter 12 — don't delete it.
下一章是唯一一章不分面、建议所有人都读的：把话说清楚。它是其余每一件资产的乘数 —— 模板写不好，后面的 Project 指令、Skill 说明、Cowork brief 全都跟着差一档。The next chapter is the one chapter that isn't about a surface and that everyone should read: saying it clearly. It's the multiplier on every other asset — get the template wrong and the Project instruction, the Skill description, and the Cowork brief all come out a notch worse.
你买的是一整套，
用的是一个输入框.
You bought the whole set,
you use one input box.
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
