— 04 / 第四章

记忆 · 从零写一份 CLAUDE.md.

同一件事你跟它说了十遍。这一章从零写一份真的 CLAUDE.md,并且用一组前后对照证明它确实改变了行为 —— 不是「感觉好点了」,是同一句提示、同一个仓库,做出来的东西不一样。顺带把四层记忆的位置与优先级、@import、rules 目录的路径限定,和一条容易踩的时间差讲清:它在会话开始读进内存,你中途改了不生效。You have told it the same thing ten times. This chapter writes a real CLAUDE.md from zero and then proves with a before/after pair that it actually changed behavior — not 'it feels better,' but the same prompt on the same repo producing different work. Along the way: the four memory tiers and their precedence, @import, path-scoped rules files, and one timing trap worth knowing — it is read into memory at session start, so a mid-session edit does not apply.

9 分钟 · 初稿 2026.07 · v2 2026.08

MD 海报

用这个包管理器,不是那个。测试要跑 npm test 不是 npx jest。migrations 目录别碰。—— 这些话你跟它说过十遍,因为模型不跨会话记事。这一章从零写一份真的 CLAUDE.md,然后用一组前后对照证明它确实改变了行为:同一句提示、同一个仓库,做出来的东西不一样。不是「感觉好点了」,是看得见的差别。Use this package manager, not that one. Tests run with npm test, not npx jest. Don't touch the migrations directory. — You've said these ten times, because models don't remember across sessions. This chapter writes a real CLAUDE.md from zero, then proves with a before/after pair that it actually changed behavior: the same prompt on the same repo producing different work. Not 'it feels better' — a difference you can see.

— I

第一步:先拍一张「没有记忆」的底片.

对照要成立,得先有前片。挑一个你的项目有特殊做法的小任务 —— 加一个组件、加一条接口、加一个测试 —— 用一句最普通的提示交给它,然后什么都别纠正,把结果原样记下来。

在这个仓库里加一个 <最小的、你们有特殊做法的东西>,照你觉得合适的方式做。
做完告诉我三件事:
1. 你把文件放在了哪里、为什么选这个位置;
2. 你跑了什么命令来验证,输出是什么;
3. 你对这个项目的约定做了哪些假设。Add a <smallest thing your team does its own way> to this repository, however you think best.
When you're done, tell me three things:
1. where you put the files and why you chose that location;
2. what command you ran to verify, and its output;
3. which assumptions you made about this project's conventions.

在这个仓库里加一个 带单元测试的日期格式化工具函数,照你觉得合适的方式做。
做完告诉我三件事:
1. 你把文件放在了哪里、为什么选这个位置;
2. 你跑了什么命令来验证,输出是什么;
3. 你对这个项目的约定做了哪些假设。Add a date-formatting utility with a unit test to this repository, however you think best.
When you're done, tell me three things:
1. where you put the files and why you chose that location;
2. what command you ran to verify, and its output;
3. which assumptions you made about this project's conventions.

第 3 问是这一步的重点。它列出的每一条「假设」,都是一条它本来该知道、但没人告诉过它的事实 —— 那张假设清单,就是你 CLAUDE.md 的第一版目录。

— II

第二步:写那份文件,一行一个否决问题.

起点不用手写:跑一次 /init,它会分析代码库,把探出来的构建命令、测试方式、项目惯例生成一份初稿;已经有 CLAUDE.md 的话,它提改进而不是覆盖。然后你做减法。Don't write the first draft by hand: run /init once and it analyzes the codebase into a starter — build commands, test setup, conventions it discovered; if a CLAUDE.md exists, it proposes improvements instead of overwriting. Then you subtract. CLAUDE.md 的敌人不是太短,是太长:它整份进入每次会话的上下文,臃肿的文件会让真正要紧的规则被淹没 —— 官方直说,写得太满,Claude 反而开始忽略你的指令。2注 2Claude Code Docs · Best practices —— 写 CLAUDE.md 的取舍表:收录它猜不到的构建命令、和默认不同的风格规则、测试跑法、仓库礼仪、项目特有的架构决定与坑;排除读代码就能推断的、语言通识、频繁变动的信息、逐文件的代码库描述。每一行问一句「删掉它会不会导致出错」,不会就删;臃肿的 CLAUDE.md 会让真正的规则被忽略。截至 2026-08-05。Claude Code Docs · Best practices — the CLAUDE.md include/exclude table: include build commands it can't guess, style rules that differ from defaults, how to run tests, repository etiquette, project-specific architecture decisions and gotchas; exclude what reading the code reveals, standard language conventions, fast-changing information, file-by-file codebase descriptions. Ask of every line 'would removing this cause mistakes' — if not, cut it; a bloated CLAUDE.md gets your real rules ignored. As of 2026-08-05.纪律是对每一行问同一个问题:删掉这行,它会不会因此犯错?不会,就删。2注 2Claude Code Docs · Best practices —— 写 CLAUDE.md 的取舍表:收录它猜不到的构建命令、和默认不同的风格规则、测试跑法、仓库礼仪、项目特有的架构决定与坑;排除读代码就能推断的、语言通识、频繁变动的信息、逐文件的代码库描述。每一行问一句「删掉它会不会导致出错」,不会就删;臃肿的 CLAUDE.md 会让真正的规则被忽略。截至 2026-08-05。Claude Code Docs · Best practices — the CLAUDE.md include/exclude table: include build commands it can't guess, style rules that differ from defaults, how to run tests, repository etiquette, project-specific architecture decisions and gotchas; exclude what reading the code reveals, standard language conventions, fast-changing information, file-by-file codebase descriptions. Ask of every line 'would removing this cause mistakes' — if not, cut it; a bloated CLAUDE.md gets your real rules ignored. As of 2026-08-05.CLAUDE.md's enemy isn't brevity, it's length: the whole file enters every session's context, and a bloated one drowns the rules that matter — the docs say it plainly: overstuffed files get your instructions ignored.2注 2Claude Code Docs · Best practices —— 写 CLAUDE.md 的取舍表:收录它猜不到的构建命令、和默认不同的风格规则、测试跑法、仓库礼仪、项目特有的架构决定与坑;排除读代码就能推断的、语言通识、频繁变动的信息、逐文件的代码库描述。每一行问一句「删掉它会不会导致出错」,不会就删;臃肿的 CLAUDE.md 会让真正的规则被忽略。截至 2026-08-05。Claude Code Docs · Best practices — the CLAUDE.md include/exclude table: include build commands it can't guess, style rules that differ from defaults, how to run tests, repository etiquette, project-specific architecture decisions and gotchas; exclude what reading the code reveals, standard language conventions, fast-changing information, file-by-file codebase descriptions. Ask of every line 'would removing this cause mistakes' — if not, cut it; a bloated CLAUDE.md gets your real rules ignored. As of 2026-08-05. The discipline is one question per line: would removing this cause a mistake? If not, cut it.2注 2Claude Code Docs · Best practices —— 写 CLAUDE.md 的取舍表:收录它猜不到的构建命令、和默认不同的风格规则、测试跑法、仓库礼仪、项目特有的架构决定与坑;排除读代码就能推断的、语言通识、频繁变动的信息、逐文件的代码库描述。每一行问一句「删掉它会不会导致出错」,不会就删;臃肿的 CLAUDE.md 会让真正的规则被忽略。截至 2026-08-05。Claude Code Docs · Best practices — the CLAUDE.md include/exclude table: include build commands it can't guess, style rules that differ from defaults, how to run tests, repository etiquette, project-specific architecture decisions and gotchas; exclude what reading the code reveals, standard language conventions, fast-changing information, file-by-file codebase descriptions. Ask of every line 'would removing this cause mistakes' — if not, cut it; a bloated CLAUDE.md gets your real rules ignored. As of 2026-08-05.

值得一行	不值得
它猜不到的构建 / 测试命令	读代码就能推断的结构
和默认不同的风格规则	语言的通行惯例
禁区目录、仓库礼仪、环境怪癖	经常变动的信息
踩过的坑和非直觉行为	逐文件的代码库导览

CLAUDE.md

# 这个仓库的规矩
 
## 命令
- 测试:`npm test`(不是 npx jest —— 我们的 jest 配置只在这个脚本里生效)
- 单文件测试:`npm test -- <路径>`
- 构建前必须先 `npm run codegen`,否则类型对不上
 
## 约定
- 组件放 `src/components/<名字>/`,一个目录含组件、测试、样式三个文件
- 日期一律用 `src/lib/date.ts` 里的工具,不要直接 new Date()
- 新增依赖前先问我 —— 这个项目对包体积有硬预算
 
## 禁区
- 不要改 `src/db/migrations/`,那是线上迁移历史
- 不要碰 `.env*`
- 不要在提交信息里加 co-author 尾注
 
## 坑
- `src/legacy/` 下的代码没有测试覆盖,改那里要手动验证并说明怎么验的

一份真实形状的项目记忆:每一行都通过了「删掉会不会出错」

什么时候写?触发律是第二次:同一个错误出现第二遍,就是写一行的时候。3注 3Claude Code Docs · Extend Claude Code —— 各扩展机制的触发时机:同一个约定或命令错第二次,写进 CLAUDE.md;反复粘贴的多步流程,做成 skill;必须每次发生的,写 hook。指令写在 CLAUDE.md 或 skill 里「是请求,不是保证」。截至 2026-08-05。什么时候搬走?一段记忆长成了多步流程,它就不该住在这里 —— 那是 skill 的形状(第 5 章)。When to write? The trigger is the second time: when the same mistake appears twice, it's time for a line.3注 3Claude Code Docs · Extend Claude Code —— 各扩展机制的触发时机:同一个约定或命令错第二次,写进 CLAUDE.md;反复粘贴的多步流程,做成 skill;必须每次发生的,写 hook。指令写在 CLAUDE.md 或 skill 里「是请求,不是保证」。截至 2026-08-05。 When to move out? A memory that has grown into a multi-step procedure no longer belongs here — that's the shape of a skill (chapter 5).

— III

第三步:重跑同一句,看差别.

关键的一步,也是最容易被跳过的一步:/clear 开一个新会话(记住 —— 中途改文件不生效,下一节讲),把第 I 步一字不改的那句提示再发一次。The crucial step, and the one most often skipped: /clear into a fresh session (remember — a mid-session edit doesn't apply; next section), and send the exact same sentence from step I, word for word unchanged.

没有 CLAUDE.md

文件落在 src/utils/(它的通用直觉),直接用了 new Date(),顺手 npx jest 跑测试 —— 报了个配置错,它又装了一个依赖去绕。交回来的东西能跑,但三条项目约定全踩了,而它列出的「假设」正好就是这三条。The file lands in src/utils/ (its generic instinct), calls new Date() directly, and runs npx jest — which errors on config, so it installs a dependency to route around it. What comes back runs, but breaks all three project conventions — and the 'assumptions' it listed are exactly those three.

有 CLAUDE.md

文件落在 src/components/ 的目录约定里,日期走 src/lib/date.ts,验证用的是 npm test 并贴了输出;想加依赖时它停下来问你。没有一次纠正,因为不需要纠正 —— 这就是那份文件替你说的十遍话。The file lands under the src/components/ directory convention, dates go through src/lib/date.ts, verification is npm test with the output pasted; and when it wants a dependency it stops and asks you. Not one correction, because none was needed — that file said your ten repetitions for you.

这组对照的意义不只是「它变听话了」。它把一件抽象的事变具体:一份写得好的 CLAUDE.md,对产出质量的改善常常大于换一个更强的模型 —— 因为多数「模型不够聪明」的错,其实是「没人告诉它」的错。也正因如此,这份文件值得像代码一样进版本库、被 review。

— IV

放哪一层,以及两个时间差.

四层记忆按从宽到窄的顺序加载,项目层出现在用户层之后。1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.The four memory layers load broadest to most specific, so a project instruction lands in context after a user instruction.1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.

层级	位置	放什么
组织	managed policy(各系统固定路径)	全公司的编码与合规规范
用户	`~/.claude/CLAUDE.md`	你个人的偏好,跟你走所有项目
项目	`./CLAUDE.md` 或 `./.claude/CLAUDE.md`	团队共享:命令、架构、惯例,进版本库
项目 · 个人	`./CLAUDE.local.md`	你的沙盒地址、测试数据,进 .gitignore

查找方式值得知道:它从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md —— 所以在 monorepo 的子包里开工,上层仓库的规矩照样带得到。反方向则是懒加载:当前目录之下的子目录 CLAUDE.md 不在启动时加载,等 Claude 真读到那些子目录里的文件才带进来。1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.这条设计很值钱:子包各自的规矩不用挤占你的开场上下文。The lookup is worth knowing: it walks up the tree from your working directory, checking every level for CLAUDE.md and CLAUDE.local.md — so starting inside a monorepo package still picks up the rules above it. Downward it's lazy: CLAUDE.md files in subdirectories below your working directory don't load at launch; they arrive when Claude actually reads files there.1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05. That design earns its keep: per-package rules cost you nothing in opening context. 文件长了有两个泄压阀。@path 把别的文件拼进来,最多递归四层,相对路径按「含这条导入的文件」解析而不是工作目录 —— 这一条踩错过很多人。1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.另一个是 .claude/rules/,按主题拆文件,还能用 paths 限定只在 Claude 碰到匹配文件时加载。1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.Two relief valves when the file grows. @path splices other files in, up to four hops, with relative paths resolved against the file containing the import rather than your working directory — a detail that has tripped many people.1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05. The other is .claude/rules/: topic files that can be scoped with paths so they load only when Claude touches matching files.1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.

— 实践提示

已经在维护一份 AGENTS.md(给别家 agent 用的)?Claude Code 读的是 CLAUDE.md,不读 AGENTS.md。官方给的桥接写法很省事:CLAUDE.md 里只写一行 @AGENTS.md,或者做个符号链接 —— 一份内容两边都认。1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.(想在 CLAUDE.md 里提一个路径又不想把它导进来,用反引号包住:导入解析会跳过行内代码与围栏代码块。1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.)Already maintaining an AGENTS.md for other agents? Claude Code reads CLAUDE.md and does not read AGENTS.md. The documented bridge is cheap: a CLAUDE.md containing the single line @AGENTS.md, or a symlink — one body of content serving both.1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05. (To mention a path in CLAUDE.md without importing it, wrap it in backticks: import parsing skips code spans and fenced blocks.1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.)

两个时间差别踩。其一:CLAUDE.md 在会话开始读入内存,中途改了不生效 —— 新内容要等下次 /clear、/compact 或重启才加载。4注 4Claude Code Docs · How Claude Code uses prompt caching —— 项目根与用户级 CLAUDE.md 在会话开始读入内存,中途编辑不废缓存、但也不生效,下次 /clear、/compact 或重启才加载新内容。截至 2026-08-05。改完发现它还按旧规矩走,不是它叛逆,是你还没翻页。其二:除了你写的记忆,它也在自己记 —— auto memory 把「下次会用上」的东西写进项目记忆目录,每次会话加载 MEMORY.md 的前 200 行或 25KB。1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05./memory 能看到两套记忆的全部文件:定期翻一翻,删掉过时的 —— 记忆也会腐烂,烂记忆比没记忆更误事。Two timing traps. First: CLAUDE.md is read into memory at session start, and a mid-session edit does not apply — new content waits for the next /clear, /compact, or restart.4注 4Claude Code Docs · How Claude Code uses prompt caching —— 项目根与用户级 CLAUDE.md 在会话开始读入内存,中途编辑不废缓存、但也不生效,下次 /clear、/compact 或重启才加载新内容。截至 2026-08-05。 If it keeps following the old rule after your edit, it isn't rebelling; you haven't turned the page. Second: beyond what you write, it takes notes of its own — auto memory files what it expects to reuse into the project's memory directory, loading the first 200 lines or 25KB of MEMORY.md each session.1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05. /memory shows every file of both memories: browse it periodically and delete what's stale — memory rots too, and rotten memory misleads worse than none.

— 注意

把预期校准准确:文档对两套记忆的定性是「上下文,不是强制配置」—— 它读了、大概率照做,但没有任何机制保证。1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.所以「提交前必须跑 lint」写在这里是叮嘱,写成 hook 才是闸门(第 8 章);「绝不许碰 .env」想要硬拦,官方指名用 PreToolUse hook,不是加粗字体。1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05.记忆负责让它懂你的项目,hooks 负责让它越不过线 —— 两件事,别用一个工具硬扛。Calibrate expectations precisely: the docs classify both memory systems as 'context, not enforced configuration' — it reads them, very likely follows them, and nothing guarantees it.1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05. So 'run lint before committing' written here is a reminder; written as a hook it's a gate (chapter 8). To truly stop it touching .env, the docs name a PreToolUse hook, not bold text.1注 1Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05. Memory's job is making it understand your project; hooks' job is making the line uncrossable — two jobs, one tool won't carry both.

本章验收 · 这份文件立起来了吗:

拍过底片:同一句提示在没有 CLAUDE.md 时跑过一次,它列出的「假设」你留着。
写好了文件:跑过 /init 打底,并用「删掉会不会出错」的尺子至少删掉三行。
做过对照:/clear 后一字不改地重跑同一句,亲眼看见差别 —— 这是你说服自己和队友的证据。
分清了层:个人偏好在 ~/.claude/,团队规矩进版本库,私人地址进 CLAUDE.local.md。

— 引用与参考

参考.

01
Claude Code Docs · How Claude remembers your project —— 四层 CLAUDE.md 按加载顺序从宽到窄:managed policy(macOS /Library/Application Support/ClaudeCode/CLAUDE.md、Linux 与 WSL /etc/claude-code/CLAUDE.md、Windows C:\Program Files\ClaudeCode\CLAUDE.md)、用户 ~/.claude/CLAUDE.md、项目 ./CLAUDE.md 或 ./.claude/CLAUDE.md、本地 ./CLAUDE.local.md;项目层出现在用户层之后。查找方式是从当前工作目录向上逐级走,沿途每一层都查 CLAUDE.md 与 CLAUDE.local.md;当前目录之下的子目录 CLAUDE.md 不在启动时加载,而是 Claude 读到那些子目录里的文件时才带进来。@path 导入相对路径按「含该导入的文件」解析而非工作目录,最多递归四层;导入解析会跳过 markdown 行内代码与围栏代码块,想提路径又不想导入就用反引号包住。项目文件里的导入有信任对话框保护(防别人往共享仓库里塞),用户层导入是你自己写的,不弹。Claude Code 读 CLAUDE.md 而不是 AGENTS.md,官方给的桥接写法是在 CLAUDE.md 里写 @AGENTS.md 或做符号链接。两套记忆都在每次会话开始加载,并且「Claude treats them as context, not enforced configuration」—— 要不论模型怎么决定都拦住某个动作,用 PreToolUse hook。auto memory 每次会话加载 MEMORY.md 的前 200 行或 25KB。截至 2026-08-05。Claude Code Docs · How Claude remembers your project — four CLAUDE.md scopes in load order, broadest to most specific: managed policy (macOS /Library/Application Support/ClaudeCode/CLAUDE.md, Linux and WSL /etc/claude-code/CLAUDE.md, Windows C:\Program Files\ClaudeCode\CLAUDE.md), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and local ./CLAUDE.local.md — a project instruction appears in context after a user instruction. Lookup walks up the directory tree from the working directory, checking every level for CLAUDE.md and CLAUDE.local.md; CLAUDE.md files in subdirectories below the working directory are not loaded at launch but are included when Claude reads files there. In @path imports, relative paths resolve against the file containing the import rather than the working directory, with a maximum depth of four hops; import parsing skips Markdown code spans and fenced code blocks, so backtick-wrapping a path mentions it without importing. Imports in project files are gated by a trust dialog that 'protects you from files other people commit to a shared project'; user-scope imports are files you wrote yourself and load without it. Claude Code reads CLAUDE.md, not AGENTS.md — the documented bridge is a CLAUDE.md containing @AGENTS.md, or a symlink. Both memory systems load at the start of every conversation, and 'Claude treats them as context, not enforced configuration' — to block an action regardless of what Claude decides, use a PreToolUse hook. Auto memory loads the first 200 lines or 25KB of MEMORY.md each session. As of 2026-08-05. Claude Code Docs · Memory
↩
02
Claude Code Docs · Best practices —— 写 CLAUDE.md 的取舍表:收录它猜不到的构建命令、和默认不同的风格规则、测试跑法、仓库礼仪、项目特有的架构决定与坑;排除读代码就能推断的、语言通识、频繁变动的信息、逐文件的代码库描述。每一行问一句「删掉它会不会导致出错」,不会就删;臃肿的 CLAUDE.md 会让真正的规则被忽略。截至 2026-08-05。Claude Code Docs · Best practices — the CLAUDE.md include/exclude table: include build commands it can't guess, style rules that differ from defaults, how to run tests, repository etiquette, project-specific architecture decisions and gotchas; exclude what reading the code reveals, standard language conventions, fast-changing information, file-by-file codebase descriptions. Ask of every line 'would removing this cause mistakes' — if not, cut it; a bloated CLAUDE.md gets your real rules ignored. As of 2026-08-05. Claude Code Docs · Best practices
↩
03
Claude Code Docs · Extend Claude Code —— 各扩展机制的触发时机:同一个约定或命令错第二次,写进 CLAUDE.md;反复粘贴的多步流程,做成 skill;必须每次发生的,写 hook。指令写在 CLAUDE.md 或 skill 里「是请求,不是保证」。截至 2026-08-05。 Claude Code Docs · Extend Claude Code
↩
04
Claude Code Docs · How Claude Code uses prompt caching —— 项目根与用户级 CLAUDE.md 在会话开始读入内存,中途编辑不废缓存、但也不生效,下次 /clear、/compact 或重启才加载新内容。截至 2026-08-05。 Claude Code Docs · Prompt caching
↩

同一句话说第三遍之前, 把它变成一行记忆.

Before you say it a third time, make it one line of memory.

Aklman Library

— 讨论

讨论.

评论区初始化中…

记忆 · 从零写一份 CLAUDE.mdMemory · Write a Real CLAUDE.md from Zero.

第一步:先拍一张「没有记忆」的底片Step One: Photograph the Repo Without Memory.

第二步:写那份文件,一行一个否决问题Step Two: Write the File, One Veto Question per Line.

第三步:重跑同一句,看差别Step Three: Rerun the Same Sentence, See the Difference.

放哪一层,以及两个时间差Which Layer — and Two Timing Traps.

讨论Discussion.

记忆 · 从零写一份 CLAUDE.md.

第一步:先拍一张「没有记忆」的底片.

第二步:写那份文件,一行一个否决问题.

第三步:重跑同一句,看差别.

放哪一层,以及两个时间差.

讨论.