每开一个新对话，你都把同一段背景再贴一遍。这一章你亲手建一个 Project：挑三份文件进知识库、写一段自定义指令、然后去 Settings > Memory 把它归纳的你读一遍、改错的那几条。钉的和攒的，各归各位。Every new chat, you paste the same background again. Here you build a Project by hand: three files into the knowledge base, one custom instruction, then open Settings > Memory and read back the version of you it has formed — fixing what's wrong. What you pin and what it accrues, each in its place.
每开一个新对话，你都把同一段背景再贴一遍 —— 我是谁、这个项目怎么回事、希望你用什么口吻。上一章那段账号级指令解决了「我是谁」，这一章解决剩下的两样。Every new chat, you paste the same background again — who I am, what this project is, the tone I want. The account-level block from the last chapter settled 'who I am'; this chapter settles the other two.
— I
四层记忆，各管各的Four Layers, Each Minding Its Own.
混着用，就会出现「我明明说过」和「你怎么还记得那件事」同时发生。Mix them up and you get 'but I told you that' and 'why do you still remember that' happening at once.普通对话是默认那层，聊完就散。真正会留下来的是下面这四样 —— 认清哪句话该写在哪一层，是这一章唯一需要你理解的事：A plain chat is the default layer and disperses when you're done. What actually persists are the four below — and knowing which sentence belongs on which layer is the only thing this chapter asks you to understand:
这一层Layer
装的是What it holds
谁来管、何时用Who manages it, when
账号级指令（第 2 章那段）Account instructions (Chapter 2)
跨所有对话的你：称呼、语气底线、固定格式The you across every chat: how to address you, baseline tone, fixed format
设置里手写，全局生效Hand-written in settings, applies everywhere
Memory
它从对话里学到的习惯与偏好Habits and preferences it learns from chats
自动攒，可查、可改、可关Accrued on its own; viewable, editable, switchable
你亲手钉，只在这个工作区生效Pinned by hand, active only in the workspace
Skill
这类事的做法（下一章）The method for a kind of thing (next chapter)
你固化成文件夹，触发时才上场Pinned into a folder, appears when triggered
指令层还有第五层要知道：Team / Enterprise 的管理员可以设组织级指令，全组织每次对话都带上；和你个人那条直接顶上时，以组织级为准。4注 4Note 4Claude 帮助中心 · 理解 Claude 的个性化功能 —— 账号级 profile instructions 作用于全部对话；project instructions 只在该项目内生效，付费档才有；组织级指令与个人指令冲突时以组织级为准。截至 2026-08。Claude Help Center · Understanding Claude's personalization features — account-level profile instructions apply to every conversation; project instructions apply only inside that project and are a paid-plan feature; where organization instructions conflict with individual ones, the organization's take precedence. As of 2026-08.One more layer on the instruction side: Admins on Team / Enterprise can set organization instructions that ride on every conversation org-wide; where yours directly contradicts theirs, theirs wins.4注 4Note 4Claude 帮助中心 · 理解 Claude 的个性化功能 —— 账号级 profile instructions 作用于全部对话；project instructions 只在该项目内生效，付费档才有；组织级指令与个人指令冲突时以组织级为准。截至 2026-08。Claude Help Center · Understanding Claude's personalization features — account-level profile instructions apply to every conversation; project instructions apply only inside that project and are a paid-plan feature; where organization instructions conflict with individual ones, the organization's take precedence. As of 2026-08.
— II
动手：建一个真 ProjectDo It: Build a Real Project.
Project 钉的是你不想再说第二遍的话。A Project pins what you don't want to say twice.一个 Project 把两样东西钉在一个工作区里：一组固定文件（知识库），和一段自定义指令（口吻、角色、规则），作用于该项目内的每一次对话；它还自带独立的对话历史。1注 1Note 1Claude 帮助中心 · 什么是 Projects —— 项目是自带对话历史与知识库的独立工作区；可上传文件、写项目指令；Free 档最多 5 个项目；付费档在知识接近上下文上限时自动启用检索（RAG），把容量扩到约 10 倍；项目共享仅 Team / Enterprise 有。截至 2026-08。Claude Help Center · What are projects — a project is a self-contained workspace with its own chat history and knowledge base; you can upload files and write project instructions; Free is capped at 5 projects; paid plans switch to retrieval (RAG) as knowledge nears the context limit, expanding capacity ~10x; project sharing is Team / Enterprise only. As of 2026-08.付费档在知识接近上下文上限时会自动转成检索（RAG），把可用容量扩到约 10 倍 —— 也就是说，它不再每次重读全部，而是只取相关的几页。1注 1Note 1Claude 帮助中心 · 什么是 Projects —— 项目是自带对话历史与知识库的独立工作区；可上传文件、写项目指令；Free 档最多 5 个项目；付费档在知识接近上下文上限时自动启用检索（RAG），把容量扩到约 10 倍；项目共享仅 Team / Enterprise 有。截至 2026-08。Claude Help Center · What are projects — a project is a self-contained workspace with its own chat history and knowledge base; you can upload files and write project instructions; Free is capped at 5 projects; paid plans switch to retrieval (RAG) as knowledge nears the context limit, expanding capacity ~10x; project sharing is Team / Enterprise only. As of 2026-08.Free 档最多建 5 个，这反而是个好约束：逼你只为真正长期的事建。A Project pins two things into one workspace: a fixed set of files (the knowledge base) and a custom instruction (tone, role, rules) that applies to every chat inside it — plus its own chat history.1注 1Note 1Claude 帮助中心 · 什么是 Projects —— 项目是自带对话历史与知识库的独立工作区；可上传文件、写项目指令；Free 档最多 5 个项目；付费档在知识接近上下文上限时自动启用检索（RAG），把容量扩到约 10 倍；项目共享仅 Team / Enterprise 有。截至 2026-08。Claude Help Center · What are projects — a project is a self-contained workspace with its own chat history and knowledge base; you can upload files and write project instructions; Free is capped at 5 projects; paid plans switch to retrieval (RAG) as knowledge nears the context limit, expanding capacity ~10x; project sharing is Team / Enterprise only. As of 2026-08. On paid plans it shifts to retrieval (RAG) as knowledge nears the context limit, expanding usable capacity roughly tenfold1注 1Note 1Claude 帮助中心 · 什么是 Projects —— 项目是自带对话历史与知识库的独立工作区；可上传文件、写项目指令；Free 档最多 5 个项目；付费档在知识接近上下文上限时自动启用检索（RAG），把容量扩到约 10 倍；项目共享仅 Team / Enterprise 有。截至 2026-08。Claude Help Center · What are projects — a project is a self-contained workspace with its own chat history and knowledge base; you can upload files and write project instructions; Free is capped at 5 projects; paid plans switch to retrieval (RAG) as knowledge nears the context limit, expanding capacity ~10x; project sharing is Team / Enterprise only. As of 2026-08. — meaning it stops re-reading everything and pulls only the relevant pages. Free is capped at five, which is a useful constraint: it forces you to create one only for things that are genuinely long-term.
01
挑一件你今年还会一直做的事Pick something you'll still be doing at year's end
不是「这周的活」，是「这条线」：一份持续在写的周刊、一个长期客户、一门在学的课、一个在维护的产品。判断标准：半年后这件事还在，且它的背景材料大半没变。Not 'this week's task' but 'this thread': a newsletter you keep writing, a long-term client, a course you're taking, a product you maintain. The test: it still exists in six months, and most of its background material hasn't changed.
02
往知识库里放三份文件，不多不少Put exactly three files in the knowledge base
第一份是「长这样」的样本（往期成品、一封满意的旧邮件、一段风格范文）；第二份是「以此为准」的规范（术语表、译名标准、写作规则、接口文档）；第三份是「这件事的背景」（客户资料、项目现状、课程大纲）。三份是刻意的下限 —— 先跑起来，缺什么再加。One 'looks like this' sample (a past deliverable, an old email you liked, a style reference); one 'defer to this' spec (a glossary, a naming standard, writing rules, an API doc); one 'here's the background' (client profile, project status, syllabus). Three is a deliberate floor — get it running, then add what's missing.
03
写一段自定义指令，用下面的模板Write the custom instruction from the template below
一句角色 + 一句背景 + 指向知识库某份文件的一句「以它为准」+ 三条「每次都要」。别堆十条规则 —— 规则一多，它开始在不该应用的地方应用。One role, one line of context, one 'defer to this file' pointing into the knowledge base, and three 'every time' rules. Don't pile up ten rules — past a certain count it starts applying them where they don't belong.
04
立刻验一次：新开对话，只说任务Verify immediately: new chat, task only
在这个 Project 里开一个新对话，只说任务本身，一句背景都别交代。它要是还问你「这是给谁看的」，说明背景那句没写清 —— 回去补，这一步才是产出。Open a new chat inside the Project and state only the task, with no background at all. If it asks 'who is this for,' your context line wasn't clear enough — go back and fix it. That correction is the actual deliverable.
你是我的<一句话角色>。
背景：<这个项目是什么、为谁做、现在到哪一步>。
材料：知识库里的<某文件>是<权威规范 / 风格样本 / 术语表>，以它为准；和它冲突的以它为准。
每次都要：<先给结论 / 用要点 / 保留所有数字 / 不替我做决定>。
不确定就问，别猜着往下做。You are my <one-line role>.
Context: <what this project is, who it's for, where it stands now>.
Material: <a file> in the knowledge base is the <authoritative spec / style sample / glossary>; defer to it, and where anything conflicts with it, it wins.
Every time: <conclusion first / use bullets / keep every number / don't decide for me>.
When unsure, ask — don't guess and run with it.
你是我的中文科技 newsletter 编辑助手。
背景：这份周刊发给已付费的开发者读者，每期 5 条，口吻克制、不卖课。
材料：知识库里的「往期样刊.md」是风格样本，「术语表.md」是译名标准，以它们为准；和它们冲突的以它们为准。
每次都要：先给 3 个标题候选；正文每条不超过 80 字、动词开头、带一个具体数字；不替我定选题。
不确定就问，别猜着往下做。You are my editor for a Chinese tech newsletter.
Context: a weekly to paying developer readers, 5 items per issue, restrained tone, no course-selling.
Material: 'past-issues.md' in the knowledge base is the style sample and 'glossary.md' is the term standard; defer to them, and where anything conflicts with them, they win.
Every time: give 3 headline options first; each item under 80 characters, verb first, with one concrete number; don't pick the topics for me.
When unsure, ask — don't guess and run with it.
— III
动手：去读一遍它记的你Do It: Go Read the You It Formed.
Memory 不用你钉，它自己学 —— 所以更该有人去看看它学成了什么样。Memory isn't pinned; it learns on its own — which is exactly why someone should go look at what it learned.Project 要你主动放文件、写指令；Memory 不用。它对 Free / Pro / Max 开放，铺开后默认启用；而且和早期那个「每 24 小时归纳一次」的版本不同，现在是实时读写更新的 —— 你这句话说完，它可能已经记下了。它还按项目分隔：每个 Project 有自己独立的记忆空间和项目摘要，不和别的项目或普通对话混。2注 2Note 2Claude 帮助中心 · 用对话搜索与记忆延续上下文 —— 新的记忆体验对 Free / Pro / Max 开放，铺开后默认启用；与旧版每日归纳不同，它在你聊天时实时读写与更新；每个项目有独立的记忆空间与项目摘要；在「设置 > 记忆」按类别查看与编辑，也能在对话里直接告诉它要记什么。截至 2026-08。Claude Help Center · Use chat search and memory to build on previous context — the new memory experience is available on Free / Pro / Max and is enabled by default once rolled out; unlike the legacy daily synthesis, it reads, writes, and updates entries in real time as you chat; each project has its own separate memory space and project summary; view and edit it by category under Settings > Memory, or just tell Claude in the conversation what to remember. As of 2026-08.A Project asks you to add files and write instructions; Memory doesn't. It's on Free / Pro / Max and enabled by default once rolled out — and unlike the earlier version that synthesized once every 24 hours, it now reads, writes, and updates in real time: it may have recorded the sentence you just finished. It's also partitioned by project — each Project has its own memory space and project summary, kept apart from other projects and from ordinary chats.2注 2Note 2Claude 帮助中心 · 用对话搜索与记忆延续上下文 —— 新的记忆体验对 Free / Pro / Max 开放，铺开后默认启用；与旧版每日归纳不同，它在你聊天时实时读写与更新；每个项目有独立的记忆空间与项目摘要；在「设置 > 记忆」按类别查看与编辑，也能在对话里直接告诉它要记什么。截至 2026-08。Claude Help Center · Use chat search and memory to build on previous context — the new memory experience is available on Free / Pro / Max and is enabled by default once rolled out; unlike the legacy daily synthesis, it reads, writes, and updates entries in real time as you chat; each project has its own separate memory space and project summary; view and edit it by category under Settings > Memory, or just tell Claude in the conversation what to remember. As of 2026-08.它不是黑箱。去设置 > 记忆，你能读到它归纳出的你，按类别分组；错的当场改、不该记的当场删。也能在对话里直接说「记住我以后都要 X」，不用离开当前窗口。2注 2Note 2Claude 帮助中心 · 用对话搜索与记忆延续上下文 —— 新的记忆体验对 Free / Pro / Max 开放，铺开后默认启用；与旧版每日归纳不同，它在你聊天时实时读写与更新；每个项目有独立的记忆空间与项目摘要；在「设置 > 记忆」按类别查看与编辑，也能在对话里直接告诉它要记什么。截至 2026-08。Claude Help Center · Use chat search and memory to build on previous context — the new memory experience is available on Free / Pro / Max and is enabled by default once rolled out; unlike the legacy daily synthesis, it reads, writes, and updates entries in real time as you chat; each project has its own separate memory space and project summary; view and edit it by category under Settings > Memory, or just tell Claude in the conversation what to remember. As of 2026-08.这件事比你想的重要：它替你记住的东西，会悄悄影响之后每一次回答 —— 而你从没审过它。It isn't a black box. Go to Settings > Memory and you can read the version of you it has formed, grouped by category; fix what's wrong and delete what shouldn't be there, on the spot. You can also just say in a chat, 'remember that from now on I want X,' without leaving the window.2注 2Note 2Claude 帮助中心 · 用对话搜索与记忆延续上下文 —— 新的记忆体验对 Free / Pro / Max 开放，铺开后默认启用；与旧版每日归纳不同，它在你聊天时实时读写与更新；每个项目有独立的记忆空间与项目摘要；在「设置 > 记忆」按类别查看与编辑，也能在对话里直接告诉它要记什么。截至 2026-08。Claude Help Center · Use chat search and memory to build on previous context — the new memory experience is available on Free / Pro / Max and is enabled by default once rolled out; unlike the legacy daily synthesis, it reads, writes, and updates entries in real time as you chat; each project has its own separate memory space and project summary; view and edit it by category under Settings > Memory, or just tell Claude in the conversation what to remember. As of 2026-08. This matters more than it sounds: what it remembers quietly shapes every answer that follows — and you have never reviewed it.
01
打开设置 > 记忆，从头读一遍Open Settings > Memory and read it top to bottom
第一次读大概率会有点不适 —— 它对你的归纳比你以为的具体。这正是要读的原因。The first read is usually a little uncomfortable — its summary of you is more specific than you expected. That's the point of reading it.
02
找三类东西：记错的、过期的、不该记的Look for three kinds of entry: wrong, expired, shouldn't-be-there
记错的直接改；过期的（三个月前那个临时项目）删掉；不该记的（某次一次性提到的敏感信息）删掉。它最隐蔽的成本不是记错，是它还记着三个月前的你，而你已经变了。Fix what's wrong; delete what's expired (that temporary project from three months ago); delete what shouldn't be there (sensitive detail you mentioned once). Its most hidden cost isn't getting you wrong — it's still remembering the you of three months ago, after you've changed.
03
补一条它没学到、但你希望它记住的Add one thing it never learned that you want it to keep
用下面那段直接在对话里说，比在设置里敲更顺手。补完再回设置里确认它真的写进去了 —— 这一步是验收。Say it in the conversation with the block below — smoother than typing in settings. Then go back to settings and confirm it landed. That confirmation is the sign-off.
提示词Prompt教 Memory 记什么、忘什么Teach Memory what to keep
记住这些（稳定的偏好）：<我的角色 / 长期约束 / 固定的格式与语气>。
别记这些（会变的近况）：<这周的临时项目 / 一次性的数字 / 还没定的探索>。
这条只用这一次、别写进记忆：<一次性的敏感信息>。
说完之后告诉我你实际记下了哪几条，用你自己的话复述一遍 —— 我要核对。Remember these (stable preferences): <my role / long-term constraints / fixed format and tone>.
Don't remember these (changing state): <this week's temporary project / one-off numbers / unsettled exploration>.
Use this once, don't write it to memory: <one-off sensitive info>.
Then tell me which entries you actually recorded, in your own words — I want to check them.
记住这些：我是独立开发者，做 iOS 工具类 app；回答一律中文、先结论后细节；代码用 Swift，别用 force-unwrap。
别记这些：我这周在赶的「报税 app」是临时外包，下个月就结束，别当成我的长期方向。
这条只用这一次、别写进记忆：这是某客户的真实姓名和合同金额。
说完之后告诉我你实际记下了哪几条，用你自己的话复述一遍 —— 我要核对。Remember these: I'm a solo developer building iOS utility apps; answer in Chinese, conclusion before detail; code in Swift, no force-unwraps.
Don't remember these: the 'tax-filing app' I'm rushing this week is temporary contract work that ends next month — don't treat it as my long-term direction.
Use this once, don't write it to memory: this is a client's real name and contract amount.
Then tell me which entries you actually recorded, in your own words — I want to check them.
— IV
换个域，Project 里放什么What Goes in the Project, by Field.
样本：一份你自己写过、觉得好用的笔记。规范：课程大纲或考纲（它决定它该讲到什么深度就停）。背景：你现在的水平、已经学过哪几章、卡在哪。指令里加一句「不要直接给答案，先反问一句检查我理解到哪」—— 这一句是学习类 Project 和其它类最大的差别。Sample: a note you wrote yourself and found useful. Spec: the syllabus or exam outline (it decides where the depth should stop). Context: your current level, which chapters you've covered, where you're stuck. Add one line to the instruction: 'don't hand me the answer — ask one question back first to check where my understanding is.' That single line is the biggest difference between a study Project and any other kind.
样本：一份你满意的文献综述。规范：引用格式与你所在期刊的方法论要求。背景：这个课题的共识与争议在哪。指令里必须有一句「知识库里没有的，明说没有，不要从常识补」—— 研究类 Project 最大的风险不是它不知道，是它用通识把空缺填得看不出来。Sample: a literature review you were happy with. Spec: the citation format and your target journal's methodology requirements. Context: where the consensus and disputes sit in this area. The instruction must contain 'if it isn't in the knowledge base, say so — do not fill the gap from general knowledge.' The biggest risk in a research Project isn't that it doesn't know; it's that it patches the gap with generalities so smoothly you can't see the seam.
样本：三封你真的发出去过、效果好的邮件或帖子。规范：品牌词表（哪些词必须用、哪些绝不能用）。背景：这个客户或这条产品线的现状与禁忌。这一类最适合建成「一个客户一个 Project」—— 因为背景那份文件是客户特有的，而样本和规范可以在几个 Project 之间复制。Sample: three emails or posts you actually sent that worked. Spec: the brand word list (what must be used, what must never be). Context: the current state and the taboos for this client or product line. This field suits 'one Project per client' — the context file is client-specific, while the sample and spec can be copied across several Projects.
样本：一份上季度的周报或复盘（它决定颗粒度）。规范：口径定义 —— 什么叫「完成」、什么算「阻塞」、指标怎么算。背景：团队构成、汇报对象、这个季度的目标。指令里加一句硬约束：「所有数字必须来自我给的材料，缺的写待确认，不要推算」。管理类 Project 出事都出在推算上。Sample: last quarter's weekly update or retro (it sets the granularity). Spec: definitions — what counts as 'done,' what counts as 'blocked,' how each metric is computed. Context: team composition, who you report to, this quarter's goals. Add one hard constraint: 'every number must come from the material I gave you; write TBC where it's missing — do not extrapolate.' Management Projects go wrong at exactly the extrapolation step.
四种「三份文件」的填法。共同点：第一份永远是「长这样」的成品样本 —— 那是最省口舌的一份，比任何形容词都准。Four ways to fill the 'three files.' What they share: the first is always a finished sample of 'looks like this' — the cheapest of the three to supply and more precise than any adjective.
— V
边界：钉的与攒的，冲突了听谁的Boundaries: Pinned vs Accrued, and Who Wins.
冲突时以 Project 指令为准The Project instruction wins
你钉的是显性约定，你说了算；它攒的是隐性习惯，它替你归纳。两者打架，听你钉的那句。What you pin is an explicit convention — you decide. What it accrues is an implicit habit — it inferred. When they fight, the pinned one wins.
Memory 记习惯，不记近况Memory holds habits, not status
值得让它记的是稳定偏好（格式、语气、长期约束）；不值得的是会变的状态（你这周在忙的临时项目）—— 记下来下个月反而误导它。Worth remembering: stable preference (format, tone, long-term constraints). Not worth it: changing state (the temporary project you're busy with this week), which only misleads it next month.
别把临时对话倒进 ProjectDon't pour throwaway chats into a Project
边界在「这个对话还属不属于这件长期的事」。属于，就在 Project 里开；不属于，开个普通对话。The line is whether this chat still belongs to that long-term thing. If it does, open it inside the Project; if it doesn't, use a plain chat.
Cowork 那边不通 MemoryMemory doesn't reach Cowork
一条容易踩的边界：记忆不在普通对话和 Cowork 之间流通，只有 Project 能跨这两者带上下文。5注 5Note 5Claude 帮助中心 · 上手 Claude Cowork —— 已知边界：记忆不在普通对话与 Cowork 之间流通，只有项目（Projects）能跨这两者带上下文。截至 2026-08。Claude Help Center · Get started with Claude Cowork — a known boundary: memory does not carry between chat and Cowork; only projects carry context across the two. As of 2026-08.所以第 8 章那件要交办的活，如果依赖你在 chat 里聊过的背景，那段背景必须落在 Project 里，不能指望 Memory 带过去。An easy one to trip on: memory does not carry between ordinary chat and Cowork — only a Project carries context across the two.5注 5Note 5Claude 帮助中心 · 上手 Claude Cowork —— 已知边界：记忆不在普通对话与 Cowork 之间流通，只有项目（Projects）能跨这两者带上下文。截至 2026-08。Claude Help Center · Get started with Claude Cowork — a known boundary: memory does not carry between chat and Cowork; only projects carry context across the two. As of 2026-08. So if the job you hand off in Chapter 8 depends on background you discussed in chat, that background has to live in the Project; Memory won't ferry it over.
— VI
收束 · 本章验收Sign-off.
建了一个 Project，对应一件半年后还在的事。Created one Project for something that will still exist in six months.
知识库里有三份文件：一份样本、一份规范、一份背景。The knowledge base holds three files: a sample, a spec, a background.
写了自定义指令，含一句「以知识库某文件为准」。Wrote the custom instruction, including one 'defer to this file in the knowledge base' line.
新开对话只说任务、不说背景，它没有反问你背景。Opened a new chat, stated only the task, and it didn't ask you for background.
去「设置 > 记忆」读完了一遍，至少改掉或删掉一条。Read Settings > Memory end to end and corrected or deleted at least one entry.
补了一条你希望它长期记住的偏好，并回设置里确认它真的写进去了。Added one preference you want kept long-term, and confirmed in settings that it actually landed.
你钉的是约定，
它攒的是习惯.
What you pin is conventions,
what it accumulates is habits.
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
