# Beyond the Prompt: Claude Code # 提示词之外：Claude Code 进阶指南 **May 26, 2026 · 23 min · Arpan Patel** **2026年5月26日 · 预计阅读时间 23 分钟 · 作者：Arpan Patel** --- Claude Code is one of those tools where the difference between a casual user and someone who has internalized it is enormous. The casual user types prompts, accepts suggestions, and treats it like a fancier autocomplete. The daily driver uses it like a programmable agent with memory, custom commands, parallel sessions, and a project setup that compounds over time. This guide is for the second kind of person, assuming you already know what `claude` does when you type it in a terminal. Claude Code 就是这样一款工具：普通用户与将其内化为习惯的高级用户之间存在着巨大的鸿沟。普通用户输入提示词、接受代码建议，仅仅把它当作一个更高级的自动补全工具。而将其作为日常驱动的主力用户，则把它看作是一个拥有记忆、支持自定义命令、能够并行处理会话，并且项目配置能够随时间产生复利效应的“可编程智能体（Agent）”。本指南专为后一种用户编写，前提是你已经了解在终端中输入 `claude` 命令的基本作用。 --- ## 🚀 1. Claude Code Beyond the Basics ## 🚀 1. 基础之外的 Claude Code Once you stop thinking of Claude Code as a prompt-and-wait chatbot and start treating it as an autonomous agent that needs guardrails, your workflow shifts. The single most important principle from Boris Cherny and the Anthropic team: **give Claude a way to verify its own work**. Without that, you are the only feedback loop. With it, Claude iterates until things actually work, and Boris says this alone gives a 2-3x quality improvement. 一旦你不再将 Claude Code 视为一个“输入提示然后等待”的聊天机器人，而是将其看作一个需要设置护栏的自主智能体，你的工作流就会发生质的转变。来自其创作者 Boris Cherny 和 Anthropic 团队最重要的一条原则是：**给 Claude 一个验证自身工作成果的方法**。如果没有这一点，你就是唯一的反馈循环。有了这一点，Claude 会不断迭代直到代码真正能够运行。Boris 表示，单凭这一点就能带来 2 到 3 倍的质量提升。 A few patterns that change how you operate day to day: 以下几个范式将改变你的日常操作模式： **Explore, then plan, then code.** Plan mode (`Shift+Tab` twice) puts Claude into read-only exploration. Read files, trace flows, understand the data model. Then get a plan. Then execute. Skip planning for small fixes; use it for anything touching more than one file. **先探索，后计划，再编码。** 计划模式（连按两次 `Shift+Tab`）会让 Claude 进入只读的探索状态。阅读文件、追踪流程、理解数据模型。然后制定一份计划，接着再执行。对于小修小补可以跳过计划阶段；但对于任何涉及多个文件的更改，请务必使用该模式。 **Use plan mode like a design document.** Have one Claude write the plan, then spin up a second Claude in a fresh session to review it as a staff engineer, with no context bias, so it actually catches gaps. If implementation goes sideways, go back to plan mode and re-plan with verification steps included. **像对待设计文档一样使用计划模式。** 让一个 Claude 会话编写计划，然后在一个全新的会话中启动第二个 Claude，让它作为主任工程师（Staff Engineer）来审查计划。因为没有历史上下文的偏见，它能真正发现漏洞。如果实现过程中出现偏差，回到计划模式，重新制定包含验证步骤的计划。 **Reference, do not describe.** Instead of “look at the auth module”, type `@src/auth/login.py`. Instead of pasting an error, pipe it: `cat error.log | claude`. Exact context beats approximate description every time. **精确引用，而非模糊描述。** 与其说“看看 auth（认证）模块”，不如直接输入 `@src/auth/login.py`。与其粘贴一段报错，不如通过管道传递它：`cat error.log | claude`。精准的上下文永远胜过模糊的描述。 **Delegate, do not pair-program.** Cat Wu (Claude Code team): “The model performs best if you treat it like an engineer you’re delegating to, not a pair programmer you’re guiding line by line.” Write a crisp brief upfront, then let it run. **是委派，而非结对编程。** Cat Wu（Claude Code 团队成员）表示：“如果你把它当作一个你要委派任务的工程师，而不是一个你需要逐行指导的结对程序员，模型的表现会达到最佳。” 提前写好一份清晰的任务简报，然后让它自己运行。 > 💡 **Tip:** Press `Ctrl+G` to open Claude’s plan in your editor and tweak it before Claude proceeds. The plan is just text, so shape it before it becomes code. > 💡 **提示：** 按 `Ctrl+G` 可以在编辑器中打开 Claude 的计划，在 Claude 继续执行前对其进行微调。计划本质上只是文本，因此在它变成代码之前，请先对它进行良好的塑造。 > 💡 **Tip:** When Claude makes a mistake, end your prompt with “Update CLAUDE.md so you do not repeat this.” Boris calls Claude “eerily good at writing rules for itself” from its own failures. This habit compounds more than any other in this guide. > 💡 **提示：** 当 Claude 犯错时，在你的提示词结尾加上“更新 CLAUDE.md，这样你就不会再犯同样的错误。” Boris 称 Claude 在从自身失败中“为自己编写规则”方面表现得“出奇的好”。这个习惯带来的复利效应超过了本指南中的任何其他技巧。 --- ## 📁 2. The .claude Directory, Properly Understood ## 📁 2. 正确理解 .claude 目录 Most people open `.claude/` once, see `CLAUDE.md`, and never look further. It is actually a layered configuration system. 大多数人只打开过一次 `.claude/` 目录，看了一眼 `CLAUDE.md`，然后就再也没管过。它实际上是一个分层的配置系统。 Two scopes: **Project scope** lives in `.claude/` inside your repo, committed to git so your team shares it. **Global scope** lives in `~/.claude/` and applies across every project on your machine. 它包含两个作用域：**项目作用域 (Project scope)** 位于代码仓库内的 `.claude/` 目录中，会提交到 git 供整个团队共享。**全局作用域 (Global scope)** 位于 `~/.claude/`，适用于你机器上的每一个项目。 Mental model: project files describe the project, global files describe you. 心智模型：项目级文件描述项目的规则，全局级文件描述你个人的习惯。 | File (文件) | Scope (作用域) | Commit (提交至 Git) | What it does (作用) | |---|---|---|---| | `CLAUDE.md` | Project and global (项目与全局) | Yes (是) | Instructions loaded every session (每次会话加载的指令) | | `CLAUDE.local.md` | Project only (仅项目) | No, gitignore it (否，加入忽略名单) | Your private project notes (你的私有项目笔记) | | `settings.json` | Project and global (项目与全局) | Yes (是) | Permissions, hooks, env vars, model defaults (权限、钩子、环境变量、默认模型) | | `settings.local.json` | Project only (仅项目) | No (否) | Personal overrides, auto-gitignored (个人重写配置，自动忽略) | | `.mcp.json` | Project only (仅项目) | Yes (是) | Team-shared MCP servers (团队共享的 MCP 服务器) | | `skills//SKILL.md` | Project and global (项目与全局) | Yes (是) | Reusable prompts invoked with `/name` (通过 `/name` 调用的可复用提示词) | | `commands/*.md` | Project and global (项目与全局) | Yes (是) | Single-file slash commands (单文件斜杠命令) | | `agents/*.md` | Project and global (项目与全局) | Yes (是) | Subagent definitions (子智能体定义) | | `rules/*.md` | Project and global (项目与全局) | Yes (是) | Topic-scoped instructions, optionally path-gated (特定主题的指令，可选择是否按路径限制) | A typical layout: 典型的目录结构： ```text my-repo/ ├── .claude/ │ ├── settings.json │ ├── agents/ │ │ ├── pr-review.md │ │ └── test-writer.md │ ├── skills/ │ │ └── api-conventions/SKILL.md │ └── rules/ │ ├── frontend.md # path-gated to src/frontend/ │ └── migrations.md # path-gated to db/migrations/ ├── CLAUDE.md # checked in, team-shared ├── CLAUDE.local.md # gitignored, personal └── .mcp.json # team-shared MCP servers ``` A few things easy to miss: 几个容易忽略的细节： `CLAUDE.md` **files cascade.** In a monorepo, both `root/CLAUDE.md` and `root/services/billing/CLAUDE.md` load when you work in the billing service. Powerful for codebases with different conventions per folder. `CLAUDE.md` **文件具有级联效应。** 在 Monorepo（单体仓库）中，当你在 billing 服务中工作时，`root/CLAUDE.md` 和 `root/services/billing/CLAUDE.md` 都会被加载。这对于不同文件夹有不同约定的代码库来说非常强大。 `rules/*.md` **is path-gated.** Guidance specific to your migrations folder does not belong in `CLAUDE.md` bloating every session; it belongs in `.claude/rules/migrations.md` with a glob. `rules/*.md` **受路径限制。** 针对你数据库迁移 (migrations) 文件夹的具体指导不应该放在 `CLAUDE.md` 中，那会让每个会话变得臃肿；它属于带有 glob 匹配规则的 `.claude/rules/migrations.md`。 **Skills over commands.** `.claude/commands/*.md` and `.claude/skills//SKILL.md` both create slash commands, but skills support supporting files, `disable-model-invocation`, allowed tools, and agent overrides. New work should go in `skills/`. **Skills (技能) 优于 commands (命令)。** `.claude/commands/*.md` 和 `.claude/skills//SKILL.md` 都可以创建斜杠命令，但 skills 能够支持附加文件、`disable-model-invocation` (禁用模型主动调用)、允许的工具配置以及智能体覆盖。新的自定义命令应首选放在 `skills/` 中。 > 💡 **Tip:** Run `claude project purge ~/path/to/repo --dry-run` to see exactly what local state Claude holds for a project, handy before handing off a laptop. > 💡 **提示：** 运行 `claude project purge ~/path/to/repo --dry-run` 可以确切看到 Claude 为某个项目持有哪些本地状态，这在交接电脑前非常实用。 --- ## 📄 3. CLAUDE.md, The Way Boris Writes It ## 📄 3. 像 Boris 那样编写 CLAUDE.md `CLAUDE.md` is loaded at the start of every session. Get it wrong and Claude repeats the same mistakes. Get it right and the same prompt produces dramatically better output. `CLAUDE.md` 会在每次会话开始时加载。写得不好，Claude 会重复同样的错误。写得好，同样的提示词能产生戏剧性提升的高质量输出。 Boris is direct about two things that matter more than the rest: Boris 直言不讳地指出了比其他任何事都重要的两点： **Keep it short.** Long files bury important rules. For every line, ask: “Would removing this cause Claude to make a mistake?” If not, cut it. **保持简短。** 冗长的文件会掩盖重要的规则。对于每一行，都要问自己：“删掉这行会导致 Claude 犯错吗？” 如果不会，就删掉它。 **Let Claude write rules for itself.** Any time Claude does something wrong, tell it: “Update CLAUDE.md so you do not repeat this.” Claude is surprisingly good at distilling its own mistakes into precise rules. Do this for a few weeks and the file becomes a curated list of every gotcha your project has. **让 Claude 为自己写规则。** 任何时候 Claude 做错了事，告诉它：“更新 CLAUDE.md，这样你就不会重蹈覆辙。” Claude 在从自己的错误中提炼精确规则方面出奇地在行。坚持几周，这个文件就会变成你项目中每一个“坑点”的精选避坑指南。 ### 3.1 The Real CLAUDE.md From the Claude Code Team ### 3.1 Claude Code 团队真实的 CLAUDE.md Boris has shared the actual `CLAUDE.md` the Claude Code team checks into their own repo. The whole team contributes multiple times a week: Boris 分享了 Claude Code 团队提交到他们自己代码仓库中的真实 `CLAUDE.md` 文件。整个团队每周都会对其进行多次贡献修改： ```markdown # Development Workflow **Always use `bun`, not `npm`.** # 1. Make changes # 2. Typecheck (fast) bun run typecheck # 3. Run tests bun run test -- -t "test name" # Single suite bun run test:file -- "glob" # Specific files # 4. Lint before committing bun run lint:file -- "file1.ts" bun run lint # 5. Before creating PR bun run lint:claude && bun run test ``` That is the entire file. Build commands Claude cannot guess, the exact order to run things, single-test invocations, the pre-PR ritual. No style preferences. No codebase tours. No platitudes. 这就是整个文件。包含了 Claude 猜不到的构建命令、执行任务的确切顺序、单测调用方式，以及 PR 提交前的必备流程。没有任何代码风格偏好，没有代码库的冗长导览，也没有废话套话。 Boris also uses `@claude` in PR comments to have Claude commit a rule directly: Boris 还在 PR 评论中使用 `@claude`，让 Claude 直接提交一条规则： ```text nit: use a string literal, not a ts enum @claude add to CLAUDE.md to never use enums, always prefer literal unions ``` *(中文释义：小建议：使用字符串字面量，不要用 ts 枚举。@claude 添加到 CLAUDE.md 中：永远不要使用枚举，总是优先使用字面量联合类型)* He calls this “Compounding Engineering,” where every PR review becomes a CLAUDE.md improvement. 他称之为“复利工程学（Compounding Engineering）”，在这种模式下，每一次 PR 审查都会转化为对 CLAUDE.md 的一次改进。 A fleshed-out template following the same philosophy: 遵循同样理念的一个更充实的模板示例： ```markdown # Code style - Use ES modules (import/export), not CommonJS (require) # Workflow - Always use `bun`, not `npm` - Run `bun run typecheck` before claiming done - Never push to main directly. Always open a PR. # Architecture - All API routes go through src/api/middleware/auth.ts - New database queries go in src/db/queries/. No inline raw SQL. # Gotchas - `User` and `UserRecord` are distinct types. UserRecord is the DB row, User is the runtime object. - `formatCurrency` assumes USD. For international use `formatCurrencyByLocale`. ``` The “Gotchas” section is the magic. Every entry is a mistake Claude made, captured the moment it happened. **“Gotchas”（坑点/注意事项）部分是真正的魔法所在。** 每一条记录都是 Claude 犯过的错，在发生的那一刻被即时捕捉并记录下来。 **What does not belong in `CLAUDE.md`:** standard language conventions, file-by-file codebase descriptions, long tutorials, API docs, anything that changes frequently. **哪些内容不属于 `CLAUDE.md`：** 标准语言约定、逐个文件的代码库描述、冗长的教程、API 文档，以及任何频繁变动的内容。 > 💡 **Tip:** Words like `IMPORTANT` or `YOU MUST` improve adherence. Use them sparingly so they carry weight. > 💡 **提示：** 像 `IMPORTANT` (重要) 或 `YOU MUST` (你必须) 这样的词语能提高遵守度。要谨慎使用，这样它们才会显得有分量。 You can import other files using `@path` syntax to keep `CLAUDE.md` short while pulling in details: 你可以使用 `@path` 语法引入其他文件，从而在引入细节的同时保持 `CLAUDE.md` 的简短： ```markdown See @README.md for project overview and @package.json for scripts. @~/.claude/my-preferences.md ``` ### 3.2 Popular CLAUDE.md Files Worth Studying ### 3.2 值得学习的流行 CLAUDE.md 文件 * **[mattpocock/skills CLAUDE.md](https://github.com/mattpocock/skills/blob/main/CLAUDE.md)**: conventions for how skills should be written and tested (关于 skills 应如何编写和测试的约定) * **[anthropics/claude-code-action](https://github.com/anthropics/claude-code-action)**: Anthropic’s own repo, treated the same as internal tools (Anthropic 自己的仓库，像对待内部工具一样对待) * **[awesome-claude-code](https://github.com/hesreallyhim/awesome-claude-code)**: links to dozens of public `CLAUDE.md` files across language ecosystems (链接了跨多个语言生态系统的几十个公开的 `CLAUDE.md` 文件) * **[claudelog.com](https://claudelog.com/)**: community-curated examples organized by stack (由社区精选、按技术栈分类的示例) --- ## 🔒 4. CLAUDE.local.md as a Daily Driver ## 🔒 4. 将 CLAUDE.local.md 作为日常驱动 `CLAUDE.local.md` lives alongside `CLAUDE.md`, gets loaded the same way, but never leaves your machine. Add it to `.gitignore`. `CLAUDE.local.md` 与 `CLAUDE.md` 处于同级目录，加载方式也相同，但它永远不会离开你的本地机器。请将其加入 `.gitignore` 中。 The way I use it: after every PR I open, reviewers leave comments. Instead of trying to remember them, I dump them into `CLAUDE.local.md` the moment I see them. Over time it becomes a personalized rule file for exactly the feedback I get most often. 我是这样使用它的：在我提交的每个 PR 之后，审查者都会留下评论。与其试着去记住它们，不如在看到评论的那一刻就将其记录到 `CLAUDE.local.md` 中。久而久之，它就变成了一个高度个性化的规则文件，专门针对我最常收到的反馈。 ```markdown # Personal review notes (private) # From PR feedback - New SQS consumers need a DLQ and alarms in the same PR - Use `Optional` over null returns - Tests for new endpoints must include the auth-failure case - Prefer named tuples over plain dicts for return types with 3+ fields # My own quirks to correct - Stop using `console.log`; use the project logger instead - Always update the OpenAPI spec when adding endpoints ``` Loaded every session, Claude already knows to include auth-failure tests and update the OpenAPI spec without me mentioning it. Nitpick comments on my PRs dropped noticeably within a couple of weeks. 由于每次会话都会加载该文件，Claude 已经知道在无需我提醒的情况下，自动包含认证失败的测试用例并更新 OpenAPI 规范。仅仅几周后，我的 PR 中那些吹毛求疵的审查评论就显著减少了。 > 💡 **Tip:** Keep two sections clearly separated: project-specific feedback and personal habits to correct. Mixing them makes the file harder to prune later. > 💡 **提示：** 保持两部分内容界限分明：特定项目的反馈和需要改正的个人习惯。将它们混在一起会使以后清理文件变得困难。 > 💡 **Tip:** Prune after a few weeks. Things that have become muscle memory can go. The file should capture what is still learning, not what you already do automatically. > 💡 **提示：** 几周后进行清理修剪。那些已经成为肌肉记忆的规则可以删掉了。这个文件应该记录你仍在学习的东西，而不是你已经自然而然会做的事情。 --- ## ⭐ 5. Skills, In Depth ## ⭐ 5. 深入解析 Skills (技能) Skills let Claude Code go from “an agent that can do anything” to “an agent that does specific things really well for your project.” They are the unit of reusable expertise. Skills (技能) 让 Claude Code 从“一个什么都能干的智能体”蜕变成“一个在你的项目中把特定事情做得极好的智能体”。它们是可复用专业知识的基本单元。 ### 5.1 What Skills Actually Are ### 5.1 Skills 究竟是什么 A skill is a folder under `.claude/skills//` (project) or `~/.claude/skills//` (global) containing a `SKILL.md` with frontmatter and instructions. The folder name becomes the slash command. 一个 skill 就是 `.claude/skills//` (项目级) 或 `~/.claude/skills//` (全局级) 下的一个文件夹，其中包含一个带有元数据头 (frontmatter) 和指令的 `SKILL.md` 文件。文件夹的名称将直接成为斜杠命令的名称。 The simplest possible skill: 最简单的 skill 示例： ```markdown --- description: Summarizes uncommitted changes and flags anything risky. Use when the user asks what changed, wants a commit message, or asks to review their diff. --- ## Current changes !`git diff HEAD` ## Instructions Summarize the changes in two or three bullet points, then list any risks: missing error handling, hardcoded values, tests that need updating. ``` Save to `~/.claude/skills/summarize-changes/SKILL.md` and `/summarize-changes` is available in every session. 将其保存到 `~/.claude/skills/summarize-changes/SKILL.md`，然后 `/summarize-changes` 命令就能在每个会话中使用了。 **Three things that make Skills powerful:** **使 Skills 变得强大的三个特性：** * **Progressive disclosure.** Claude loads only frontmatter descriptions at session start (~100 tokens each). Full `SKILL.md` and helper files load only when the skill is actually needed. **渐进式披露。** Claude 在会话开始时仅加载元数据头的描述信息（每个约 100 tokens）。完整的 `SKILL.md` 及辅助文件只有在该技能实际被需要时才会加载。 * **Skills are folders, not files.** Bundle templates, reference docs, scripts, config. `SKILL.md` is just the entry point. **Skills 是文件夹，而非单一文件。** 你可以将模板、参考文档、脚本和配置打包在一起。`SKILL.md` 只是一个入口。 * **Inline shell.** Lines starting with `!` run a command and inject the output at invocation time. **内联 Shell 执行。** 以 `!` 开头的行将运行命令，并在调用时将输出内容注入到上下文中。 Frontmatter supports useful extras: 元数据头还支持更多有用的配置： ```yaml --- name: my-skill description: When to use this skill disable-model-invocation: true # only runs when user explicitly types /my-skill allowed-tools: Read, Grep, Bash agent: read-only --- ``` > 💡 **Tip:** Use `disable-model-invocation: true` for skills with side effects. You want `/ship` to deploy only when explicitly typed, not when Claude decides it is relevant. > 💡 **提示：** 对于带有副作用的技能，请使用 `disable-model-invocation: true` (禁用模型自动调用)。你希望 `/ship` (发布) 仅在你明确输入时才执行部署，而不是由 Claude 自作主张认为相关时去执行。 ### 5.2 Writing a Real Skill: Go API Conventions ### 5.2 编写一个真实的 Skill：Go API 规范 A complete skill for a Go service team, covering conventions, gotchas, and scaffolding for a new HTTP handler: 这是一个供 Go 服务团队使用的完整技能，涵盖了约定、注意事项以及新 HTTP handler (处理函数) 的脚手架： ```text .claude/skills/go-handler/ ├── SKILL.md ├── templates/ │ └── handler.go.tmpl └── examples/ └── healthz.go ``` ```markdown --- description: Scaffolds a new HTTP handler in our Go service following team conventions for routing, validation, error handling, and tests. Use when the user asks to add a new endpoint, a new handler, or extend an existing route group. --- # Go HTTP Handler Skill ## Stack - Go 1.22 with chi router - sqlc for typed queries, never write raw SQL strings in handlers - zap for structured logging, never fmt.Println - testify for assertions, table-driven tests preferred ## Gotchas - `chi.URLParam` returns `""` for missing params, not an error. Always check. - Our `httperr.Wrap` does not log. Log separately with `h.log.Error` before returning. - Auth middleware injects via `context.Value(authkey.User)`. Type-assert to `*models.User`. - sqlc nullable strings use `pgtype.Text`. Check `.Valid` before calling `.String`. - Tests must use `httptest.NewRecorder` and `httptest.NewRequest`. No real server. ``` A skill like this lets a new developer add a fully conventional endpoint without reading the entire codebase first. 拥有这样一个技能，新入职的开发者在完全不需要通读整个代码库的情况下，就能添加一个完全符合规范的新接口。 ### 5.3 Popular Skills Worth Installing ### 5.3 值得安装的流行 Skills **[mattpocock/skills](https://github.com/mattpocock/skills)**, the most popular skills repo (~100k stars). Standouts: **[mattpocock/skills](https://github.com/mattpocock/skills)**，最受欢迎的技能库（约10万星）。亮点包括： * `/grill-me`: interviews you about a plan before any code gets written (在写代码前对你的计划进行“盘问/面试”) * `/tdd`: enforces red-green-refactor strictly (严格执行“红-绿-重构”测试驱动开发) * `/diagnose`: disciplined debugging, reproduce, minimize, hypothesize, fix, regression test (有纪律的调试：复现、最小化、假设、修复、回归测试) Install: `npx skills@latest add mattpocock/skills` (安装命令：`npx skills@latest add mattpocock/skills`) **[Jeffallan/claude-skills](https://github.com/Jeffallan/claude-skills)** ships 66 language-specific profiles: `go-pro`, `python-pro`, `java-architect`, `typescript-pro`, `rust-engineer`, `sql-pro`, and more. Compose them; a Next.js task pulls in `nextjs-developer` and `typescript-pro` together. **[Jeffallan/claude-skills](https://github.com/Jeffallan/claude-skills)** 提供了 66 个针对特定语言的预设配置，例如 `go-pro`、`python-pro`、`java-architect` 等。你可以组合使用它们；比如一个 Next.js 任务可以同时引入 `nextjs-developer` 和 `typescript-pro`。 **Anthropic’s official skills:** **Anthropic 官方技能：** * `/code-review`: four parallel agents audit the diff, confidence-scored findings only (四个并行智能体审查代码差异，仅输出带置信度评分的发现) * `/simplify`: reviews recent code for reuse and efficiency (审查最近的代码以提高复用性和效率) * `/batch`: fans out a migration to dozens of parallel agents, each in its own worktree (将迁移工作分配给几十个并行智能体，每个都在自己的工作树中运行) * `/webapp-testing`: gives Claude Playwright control to test your local web app (赋予 Claude 控制 Playwright 的权限来测试你本地的 Web 应用) > 💡 **Tip:** If you do something more than once a day, turn it into a skill. Anything you repeat is a skill waiting to be written. > 💡 **提示：** 如果你每天做某件事超过一次，把它变成一个 skill。任何你重复做的事情，都是一个亟待编写的 skill。 > 💡 **Tip:** Check skills into git. They become institutional knowledge, and new engineers clone the repo and get the team’s accumulated practices for free. > 💡 **提示：** 将 skills 提交到 git 仓库中。它们将成为团队的机构知识，新工程师克隆仓库后，就能免费获得团队累积的最佳实践。 --- ## 🤖 6. Building Custom Subagents ## 🤖 6. 构建自定义子智能体 (Subagents) A subagent runs in its own context window with its own tool permissions and reports back a summary. It can read fifty files without filling up your main session. That is the entire value proposition. 子智能体 (Subagent) 在其自己的上下文窗口中运行，拥有独立的工具权限，并向你汇报摘要。它可以阅读五十个文件，而不会塞满你的主会话上下文。这正是它的核心价值主张。 A subagent is a markdown file under `.claude/agents/` (project) or `~/.claude/agents/` (global) with a frontmatter block declaring name, description, tools, and model. 子智能体是 `.claude/agents/` (项目级) 或 `~/.claude/agents/` (全局级) 下的一个 Markdown 文件，包含一个声明了名称、描述、可用工具和所用模型的元数据头 (frontmatter)。 ### 6.1 Walking Through a /pr-review Agent ### 6.1 详解 /pr-review 代理 ```markdown --- name: pr-review description: Reviews the current branch diff against main, looking for bugs, security issues, missed edge cases, and project-convention violations. Use proactively before opening a PR. tools: Read, Grep, Glob, Bash model: opus --- You are a senior staff engineer reviewing a pull request. Thorough, direct, goal is to catch issues before human reviewers do. ## Process 1. Run `git diff main...HEAD` 2. Run `git log main..HEAD --oneline` 3. Read full files, not just diff context 4. Cross-check against CLAUDE.md, CLAUDE.local.md, and .claude/rules/ ## Flag - Correctness bugs: off-by-one, null handling, error paths, race conditions - Security: injection risks, missing auth checks, secrets in code - Missing tests for new logic - N+1 queries - Convention violations from CLAUDE.md or rules/ ## Do NOT flag - Style preferences not in project rules - Refactoring suggestions for working code - Anything outside this diff ## Output Group by severity (Critical / High / Medium / Low). File + line + issue + suggested fix. End with a verdict: **SHIP**, **FIX FIRST**, or **REWORK**. ``` Run it by saying `Have the pr-review agent look at my current branch.` The subagent handles everything in its own context, your main session stays clean. 通过告诉主会话 `Have the pr-review agent look at my current branch.` (让 pr-review 代理看看我当前的分支) 来运行它。子智能体会处理自己上下文中的一切，你的主会话将保持干净。 Key design choices: `tools` is read-only, because a reviewer that modifies code gets biased toward defending its own edits. `model: opus` for high-stakes review. The “Do NOT flag” section keeps signal-to-noise high. 关键设计选择：`tools` 是只读的，因为如果审查者去修改代码，就会产生维护自己修改的偏见。`model: opus` 用于高风险、高质量要求的代码审查。“不要标记 (Do NOT flag)”部分能够保持高信噪比。 ### 6.2 Popular Subagents Worth Stealing ### 6.2 值得“偷师”的流行子智能体 The Claude Code team checks in: `build-validator`, `code-architect`, `code-simplifier`, `oncall-guide`, `verify-app`. Claude Code 团队在自己的代码中提交了这些子智能体：`build-validator` (构建验证器)、`code-architect` (代码架构师)、`code-simplifier` (代码精简器)、`oncall-guide` (值班指南)、`verify-app` (应用验证器)。 Community patterns worth adopting: 值得采用的社区优秀模式： | Agent (智能体名称) | What it does (作用) | |---|---| | `security-reviewer` | injection, auth, secrets, insecure deserialization (检查注入、认证、密钥泄露、不安全的反序列化) | | `test-writer` | generates tests, pairs with code-reviewer in a loop (生成测试，并在循环中与代码审查智能体配合工作) | | `debugger` | traces failing tests to root causes (将失败的测试追踪溯源至根本原因) | | `performance-auditor` | profiles flows and queries (对程序流和数据库查询进行性能分析) | | `migration-writer` | generates DB migrations matching project conventions (生成符合项目规范的数据库迁移脚本) | | `release-notes-writer` | changelogs from commit history (根据提交历史生成变更日志) | Curated repos: [VoltAgent/awesome-claude-code-subagents](https://github.com/VoltAgent/awesome-claude-code-subagents) (100+ agents) and [hesreallyhim/a-list-of-claude-code-agents](https://github.com/hesreallyhim/a-list-of-claude-code-agents). 精选仓库推荐：(见上方英文链接) > 💡 **Tip:** Chain agents: Session A implements, then call `Use the code-reviewer subagent to check the work.` The reviewer evaluates in a fresh context with no implementation bias. > 💡 **提示：** 链式调用智能体：会话 A 负责实现，然后输入 `Use the code-reviewer subagent to check the work.` (使用代码审查子智能体检查这项工作)。审查者将在一个没有“实现偏见”的全新上下文中进行评估。 > 💡 **Tip:** Add `isolation: worktree` to frontmatter to run the subagent in its own git worktree, especially powerful when fanning out a migration across dozens of parallel agents. > 💡 **提示：** 在元数据头中添加 `isolation: worktree`，让子智能体在其专属的 git 工作树中运行，这在跨几十个并行智能体铺开迁移工作时尤为强大。 --- ## 🧩 7. Plugins and the Marketplace ## 🧩 7. 插件与市场 (Plugins and Marketplace) Plugins bundle skills, hooks, subagents, and MCP servers into a single installable unit. Run `/plugin` to open the marketplace browser. Add community marketplaces with `/plugin marketplace add owner/repo`. 插件将 skills、钩子 (hooks)、子智能体 (subagents) 和 MCP 服务器打包成一个可安装的单元。运行 `/plugin` 可以打开市场浏览器。使用 `/plugin marketplace add owner/repo` 来添加社区市场。 **Day-one installs:** **首日必装插件推荐：** **`/code-review`** runs four parallel agents: two audit for CLAUDE.md compliance, one scans for bugs, one analyzes git blame for context. Confidence-scored, high signal-to-noise. **`/code-review`** 运行四个并行智能体：两个审计 CLAUDE.md 的合规性，一个扫描漏洞，一个分析 git blame 获取上下文。具备置信度评分，信噪比极高。 **`/feature-dev`** is the most popular skill on the official marketplace. Turns a feature brief into working code through seven phases: requirements → exploration → architecture → implementation → testing → review → docs. **`/feature-dev`** 是官方市场上最受欢迎的技能。它通过七个阶段将功能简报转化为可用的代码：需求 → 探索 → 架构 → 实现 → 测试 → 审查 → 文档。 **Language server plugin** provides precise symbol navigation and automatic diagnostics after every edit. The team consistently calls this the single highest-impact plugin you can install. **Language server (语言服务器) 插件** 提供精确的符号导航，并在每次编辑后自动进行诊断。团队一致认为这是你能安装的影响最大的单一插件。 **`/security-guidance`** is Anthropic’s official security skill, surfacing concerns before they ship. **`/security-guidance`** 是 Anthropic 官方的安全技能，在代码发布前暴露安全隐患。 Plugin categories worth knowing (1,000+ plugins across 75+ marketplaces as of mid-2026): 值得了解的插件类别（截至2026年中期，分布在75个以上市场中的1000多个插件）： * Git workflow, code intelligence (LSP), documentation generators, testing, browser automation (Playwright), design system (Figma), observability (Sentry, Datadog) (Git 工作流、代码智能 (LSP)、文档生成器、测试工具、浏览器自动化 (Playwright)、设计系统 (Figma)、可观测性 (Sentry, Datadog)) > 💡 **Tip:** A team-shared `.mcp.json` plus a few well-chosen plugins gets a new engineer productive within minutes of cloning the repo. Treat plugin choices as part of your onboarding story. > 💡 **提示：** 一个团队共享的 `.mcp.json` 加上几个精心挑选的插件，能让新入职的工程师在克隆代码仓库后的几分钟内就具备生产力。将插件选择作为团队新人入职流程的一部分。 --- ## ⌨️ 8. Underused Claude Code Commands ## ⌨️ 8. 被低估的 Claude Code 命令 Most users learn `/clear`, `/compact`, and `/init` and stop. A handful of the rest quietly do more for productivity than anything else. 大多数用户学会了 `/clear`、`/compact` 和 `/init` 之后就止步了。然而，剩下的少数命令在默默地发挥着极高生产力的作用。 | Command (命令) | What it does (作用) | |---|---| | `/insights` | Analyzes your usage patterns; run once a month (分析你的使用模式；建议每月运行一次) | | `/compact ` | Compresses session; hint controls what survives (压缩会话；提示词控制哪些内容保留下来) | | `/copy` | Copies last response; interactive picker for code blocks (复制最后一次回答；包含代码块的交互式选择器) | | `/rewind` | Undo for your whole session, restoring code, conversation, or both (整个会话的“撤销”键，可恢复代码、对话或两者) | | `/btw` | Side question that never enters conversation history (附带的旁侧问题，永远不会进入对话历史记录) | | `/context` | Visualizes context usage (可视化上下文使用情况) | | `/export ` | Dumps conversation to file (将会话导出到文件) | | `/branch` | Forks your session to try something risky (派生出你的会话分支，去尝试有风险的操作) | | `/batch` | Fans work out to parallel agents across worktrees (将工作扇出/分发给跨工作树的并行智能体) | | `/loop ` | Schedules Claude to run on repeat, up to 3 days (计划 Claude 重复运行，最长可达3天) | | `/schedule` | Cloud version of `/loop`, works even when your laptop is closed (`/loop` 的云端版本，即使你合上笔记本也能运行) | | `/teleport` | Moves a session between terminal and web (在终端和 Web 界面之间转移会话) | | `/focus` | Hides intermediate tool calls, shows only final result (隐藏中间的工具调用过程，只显示最终结果) | | `/voice` | Voice input; Boris says he codes mostly by speaking (语音输入；Boris 说他现在主要靠“说”来编程) | | `--bare` | Up to 10x faster startup for non-interactive `claude -p` usage (对于非交互式的 `claude -p` 使用场景，启动速度提升高达 10 倍) | **`/compact` vs `/clear`:** genuinely new task = `/clear` with a fresh hand-written brief. Related task where you still need context = `/compact` with a hint. `/compact` is a lossy LLM summary; `/clear` is your brief. That distinction matters. **`/compact` (压缩) 对比 `/clear` (清空)：** 真正全新的任务 = 使用 `/clear` 清空，并手动写一份新的任务简报。仍需要上下文的相关任务 = 使用 `/compact` 加提示词。`/compact` 是有损的 LLM 总结；`/clear` 则是你干净的简报。这种区分非常重要。 **`/rewind`** creates a checkpoint for every prompt, and those persist across sessions. When Claude goes down a wrong path, do not type “that did not work, try X,” as that pollutes context. Rewind and re-prompt with what you learned. **`/rewind` (倒带)** 会为每次提示创建一个检查点，并且这些检查点会跨会话持久保存。当 Claude 走错方向时，不要输入“那个不行，试试 X”，因为那会污染上下文。直接倒带，并根据你刚刚学到的情况重新提示。 > 💡 **Tip:** Use `!` as a shell escape. `!git status` or `!npm test` runs immediately with output landing in context. > 💡 **提示：** 使用 `!` 作为 shell 逃逸符。`!git status` 或 `!npm test` 会立即执行，并且输出结果会直接落入当前对话上下文中。 > 💡 **Tip:** Set `CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000`. Context rot kicks in around 300-400k tokens on the 1M model, so force earlier compaction to stay sharp. > 💡 **提示：** 设置 `CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000`。在 100万 Token 的模型中，上下文腐烂通常在 30万-40万 Tokens 左右开始，因此强制提前压缩能保持模型敏锐度。 **Fan-out pattern:** generate a task list, then loop: **扇出模式（并行分发）：** 生成任务列表，然后循环： ```bash for file in $(cat files.txt); do claude -p "Migrate $file from React to Vue. Return OK or FAIL." \ --allowedTools "Edit,Bash(git commit *)" \ --bare done ``` Test on three files. Fix the prompt. Then run on two thousand. (先用三个文件进行测试。修正你的提示词。然后放手让它跑两千个文件。) ### 8.1 /goal, the Ralph Loop Built In ### 8.1 /goal：内置的 Ralph 循环 `/goal` sets a completion condition. Claude keeps working until the condition is true. Every time it tries to stop, it checks the condition against the transcript. `/goal` 设置一个完成条件。Claude 会一直工作直到条件满足为止。每当它试图停止时，它都会对照上下文记录检查该条件。 ```bash /goal all tests in test/auth pass and the lint step is clean ``` *(中文释义：目标是 test/auth 中的所有测试均通过，且代码风格检查无报错)* Real examples: (真实的例子：) ```bash /goal all integration tests in tests/api pass without flaking 3 runs in a row /goal the OpenAPI spec validates and matches the actual response shapes /goal docker compose up runs cleanly and the healthcheck endpoint returns 200 /goal coverage on src/billing/ is above 80% and all new tests are not placeholders ``` Pick a verifiable, deterministic condition, tied to a test command, CLI exit code, or file state. Vague conditions like “the code is good” do not work. 选择一个可验证的、确定性的条件，并将其与测试命令、CLI 退出码或文件状态绑定。像“代码写得好”这种模糊的条件是行不通的。 Companions that pair well: 能与该功能绝佳搭配的组合： * `/loop`: repeat at an interval, burn down a backlog (按时间间隔重复，烧掉积压任务) * `/schedule`: run on a cadence in the cloud (在云端按节奏运行) * A `Stop` hook: gate on your own test suite or CI endpoint (一个 `Stop` 钩子：用你自己的测试套件或 CI 接口作为门控) * **Auto mode (自动模式)**: removes permission prompts so long goals do not stall (移除权限弹窗，避免长期目标因等待确认而停滞) > 💡 **Tip:** Combine `/goal` + auto mode + `/focus`. Write a crisp brief, set the goal, walk away. Come back to a finished PR. This is the workflow Boris and Cat Wu push for Opus 4.7. > 💡 **提示：** 结合 `/goal` + 自动模式 + `/focus`。写一份清晰的简报，设定目标，然后离开去喝杯咖啡。回来时迎接一个已完成的 PR。这就是 Boris 和 Cat Wu 针对 Opus 4.7 所推崇的工作流。 --- ## 🔌 9. MCPs as Power Tools ## 🔌 9. 作为强力工具的 MCP MCP (Model Context Protocol) turns Claude Code from a coding agent into a system-aware coding agent. An MCP server exposes external tools like a database, a design tool, your error tracker, or your notes to Claude in a standardized way. MCP (模型上下文协议) 将 Claude Code 从一个仅仅是写代码的智能体，变成了一个具有系统感知能力的智能体。MCP 服务器以标准化的方式向 Claude 暴露外部工具，比如数据库、设计工具、错误追踪器或你的笔记。 Without MCP, Claude reads files and runs commands. With MCP, Claude reads your Linear tickets, queries your Postgres, pulls up a Figma component, fetches live Sentry stack traces, or reads your Obsidian vault, all without leaving the terminal. 如果没有 MCP，Claude 只能读取文件和运行命令。有了 MCP，Claude 可以读取你的 Linear 工单、查询你的 Postgres 数据库、拉取 Figma 组件、获取实时的 Sentry 堆栈追踪、或者读取你的 Obsidian 笔记库，所有这一切都无需离开终端。 **The go-to MCPs for engineering work:** **工程工作常用的 MCP：** | MCP | What it unlocks (带来的能力解锁) | |---|---| | GitHub | Repo management, PRs, issues, code search (仓库管理、PR、Issue、代码搜索) | | Context7 | Live, up-to-date library docs; append `use context7` to any prompt (实时最新的依赖库文档；在任何提示词后加上 `use context7`) | | Sentry | Real error context, stack traces, breadcrumbs (真实的错误上下文、堆栈追踪、面包屑导航) | | Linear | Read/create tickets, update status (读取/创建工单、更新状态) | | Playwright | Browser automation via accessibility snapshots (通过可访问性快照实现浏览器自动化) | | Figma | Live design tree: auto-layout, spacing tokens, component refs (实时设计树：自动布局、间距 token、组件引用) | | Postgres / Supabase | Query your dev DB directly (直接查询你的开发数据库) | | Slack | Read threads, summarize discussions, draft responses (阅读讨论串、总结讨论内容、起草回复) | Local servers use stdio, vendor-hosted use HTTP with OAuth: 本地服务器使用标准输入输出 (stdio)，供应商托管的服务使用带有 OAuth 的 HTTP： ```bash claude mcp add --transport http sentry https://mcp.sentry.dev/mcp ``` Team-shared MCPs go in `.mcp.json` at the project root. Personal MCPs go in `~/.claude.json`. 团队共享的 MCP 放在项目根目录的 `.mcp.json` 中。个人的 MCP 则放在 `~/.claude.json` 中。 ### 9.1 A Real Obsidian Workflow ### 9.1 真实的 Obsidian 工作流 The Obsidian + Claude Code pairing becomes genuinely powerful when you use it as a three-tier memory architecture, not just “Claude can read my vault.” 当你将 Obsidian + Claude Code 的组合作为“三层记忆架构”使用，而不仅仅是“Claude 可以读取我的笔记库”时，它们将变得真正强大。 **Setup:** Install `obsidian-claude-code-mcp` in Obsidian (exposes the vault on a local WebSocket, port 22360). Claude Code auto-discovers it. Add a `CLAUDE.md` to your vault explaining the folder structure. **设置：** 在 Obsidian 中安装 `obsidian-claude-code-mcp`（在本地 WebSocket 的 22360 端口上暴露笔记库）。Claude Code 会自动发现它。在你的笔记库中添加一个 `CLAUDE.md` 来解释文件夹结构。 **Folder structure:** **文件夹结构：** ```text vault/ ├── 00-Inbox/ # raw capture (原始捕获) ├── 10-Daily/ # one note per day (每天一篇笔记) ├── 20-Projects/ # active project notes (活跃项目笔记) │ └── billing-v2/ │ ├── README.md # goal, status, open questions (目标、状态、待解决问题) │ ├── decisions/ # ADRs (架构决策记录) │ └── sessions/ # one log per Claude session (每个 Claude 会话一份日志) ├── 30-Decisions/ # cross-project ADRs (跨项目的架构决策记录) ├── 40-Atoms/ # reusable knowledge, linked (可复用的知识，互相关联) └── 90-Archive/ ``` **The three tiers:** **三层架构：** **Hot storage: daily session log.** Every Claude session writes a timestamped log to `10-Daily/.md`. A `Stop` hook can do this automatically: when the agent finishes, it appends a structured summary. **热存储：日常会话日志。** 每个 Claude 会话都会将带有时间戳的日志写入 `10-Daily/.md`。一个 `Stop` 钩子可以自动完成这件事：当智能体完成任务时，追加一份结构化总结。 **Warm storage: project notes.** Each project has a folder under `20-Projects/`. Before any new session, Claude reads the project README and last 2-3 session logs to rebuild context. Two weeks of context in 30 seconds. **温存储：项目笔记。** 每个项目在 `20-Projects/` 下都有一个文件夹。在任何新会话开始之前，Claude 都会读取项目 README 和最近 2-3 次会话日志来重建上下文。30秒内找回过去两周的上下文。 **Cold storage: decisions and atoms.** Architectural decisions get promoted into `30-Decisions/` as ADRs. Reusable knowledge gets distilled into `40-Atoms/` and linked via wikilinks across all projects. **冷存储：决策与原子知识。** 架构决策会被提升到 `30-Decisions/` 作为 ADRs (架构决策记录)。可复用的知识被提炼到 `40-Atoms/` 中，并通过 Wiki 链接在所有项目间相互关联。 **Daily workflows:** **日常工作流命令示例：** * `What is in my inbox? Summarize and suggest where each item belongs.` (我的收件箱里有什么？总结一下并建议它们各自应归档到哪里。) * `Check 30-Decisions/ for anything related to retry policies.` (检查 30-Decisions/ 目录，看看有没有关于重试策略的决策。) * `Read the last 3 session logs for billing-v2. Tell me where I left off.` (读取 billing-v2 项目最近的 3 份会话日志。告诉我上次进行到哪里了。) > 💡 **Tip:** Resist installing every MCP. Each one expands the tool list Claude reasons over, and bloated tool lists hurt decision quality. Starter set: GitHub, Context7, plus one or two domain-specific. > 💡 **提示：** 忍住不要安装所有的 MCP。每一个都会扩大 Claude 需要推理的工具列表，臃肿的工具列表会损害决策质量。初始推荐集：GitHub、Context7，外加一两个特定领域的即可。 > 💡 **Tip:** Run `/mcp` inside Claude Code to list every active server and its connection status. First place to check when something is not working. > 💡 **提示：** 在 Claude Code 内部运行 `/mcp`，列出每一个活跃的服务器及其连接状态。当出故障时，这应该是你首先检查的地方。 --- ## 📅 10. Optimizing Your Daily Workflow ## 📅 10. 优化你的日常工作流 **Morning.** Open Claude Code in the project. Skim what subagents and scheduled jobs did overnight. Run `/insights` once a week. **早晨。** 在项目中打开 Claude Code。快速浏览子智能体和计划任务在夜间做了什么。每周运行一次 `/insights`。 **New feature.** Plan mode → edit plan with `Ctrl+G` → implement → invoke `/pr-review` subagent or spin up a fresh Claude session to review. **开发新功能。** 计划模式 → 使用 `Ctrl+G` 编辑计划 → 代码实现 → 调用 `/pr-review` 子智能体，或启动一个新的 Claude 会话进行审查。 **Bug.** Reproduce first. Pipe the error: `cat error.log | claude`. Ask Claude to write a failing test that reproduces it. Only then ask it to fix. The test prevents the fix from being a guess. **修 Bug。** 首先复现。将错误通过管道传递：`cat error.log | claude`。让 Claude 写一个会报错的测试用例来复现问题。在此之后才让它修复。测试用例能防止它的修复只是靠瞎猜。 **Migrations or mass changes.** Use `/batch`, which interviews you about the change, then fans out to parallel agents, each in its own worktree, each testing and creating a PR. **代码迁移或大规模变更。** 使用 `/batch` 命令，它会针对这些变更对你进行询问/面试，然后将任务分发给并行的智能体，每一个都在自己的工作树中独立运行、测试并创建 PR。 **Unfamiliar code.** Use a subagent: “Use a subagent to investigate how our auth handles token refresh.” It reads dozens of files in its own context and reports back a summary. Your main session stays clean. **面对不熟悉的代码。** 使用子智能体：“使用一个子智能体来调查我们的认证系统是如何处理 Token 刷新的。” 它会在自己的上下文中阅读数十个文件并向你汇报总结，而你的主会话仍保持干净。 **Parallel sessions.** Boris and the team call this the single biggest productivity unlock: three to five git worktrees, each running its own Claude session. Use the agent view (`claude agents`) as a control plane. **并行会话。** Boris 团队称这是释放生产力最大的一招：开启三到五个 git 工作树 (worktrees)，每一个都运行着自己的 Claude 会话。使用智能体视图 (`claude agents`) 作为你的控制平面。 **Writer/Reviewer pattern.** Session A implements. Session B reviews in a fresh context. Copy the review back, fix, repeat. **编写者/审查者模式。** 会话 A 负责实现。会话 B 在一个全新的上下文中进行审查。将审查意见复制回 A 会话，修复代码，然后重复。 **Compact at milestones.** After finishing a logical chunk: `/compact Preserve the decisions made, files changed, and test commands.` **在里程碑处进行压缩。** 在完成一块逻辑任务后，执行：`/compact 保留已做的决策、更改的文件以及测试命令。` > 💡 **Tip:** Never let Claude claim success without evidence, whether that is tests, screenshots, or real command output. The trust-then-verify gap is the single biggest source of bad output. > 💡 **提示：** 永远不要让 Claude 在没有证据的情况下声称自己成功了，无论是测试通过记录、截图还是真实的命令输出。盲目信任缺乏验证是产生糟糕代码的最大源头。 --- ## 💡 11. Tips From the Anthropic Team ## 💡 11. 来自 Anthropic 团队的提示 Collected from Boris, Cat Wu, Thariq, and the broader team. These are the patterns that actually change how they work: 这些技巧收集自 Boris、Cat Wu、Thariq 以及更广泛的团队成员。这些正是真正改变他们工作方式的范式： * **“Give Claude a way to verify its output. Once you do that, Claude will iterate until the result is great.”** Boris’s single most-repeated tip. **“给 Claude 一个验证其输出的方法。一旦你做到这一点，Claude 会不断迭代，直到结果完美。”** 这是 Boris 重复最多的一条提示。 * **Use Opus with high or xhigh effort for almost everything.** The smaller model that needs more correction is often slower overall, which is Boris’s reasoning for defaulting to Opus. **几乎在所有事情上都使用 Opus 模型并设置为 high 或 xhigh effort。** 需要更多纠正的小模型往往在总体上更慢，这也是 Boris 默认推荐使用 Opus 的原因。 * **Run 3-5 sessions in parallel.** Worktrees over checkouts. Use `claude --worktree` or the Desktop app. The agent view ties them together. **并行运行 3-5 个会话。** 优先使用 Worktrees（工作树）而不是 checkouts。使用 `claude --worktree` 命令或桌面应用。Agent view (智能体视图) 会将它们整合在一起。 * **Maintain a notes directory per project, updated after every PR.** Tell Claude to keep notes in a directory, point `CLAUDE.md` at it. The codebase compounds in self-knowledge. **为每个项目维护一个笔记目录，在每次 PR 后更新。** 告诉 Claude 把笔记保存在某个目录中，并在 `CLAUDE.md` 里指向它。你的代码库会在自我认知上产生复利。 * **Build a `/techdebt` slash command.** Run it at the end of every session to find and kill duplicated code. **构建一个 `/techdebt` (技术债务) 斜杠命令。** 在每次会话结束时运行它，以发现并清除重复代码。 * **The team’s `CLAUDE.md` is shared and edited multiple times a week.** Anytime someone sees Claude do something incorrectly, they add a rule. Treat it as a living document. **团队的 `CLAUDE.md` 会被共享且每周会被编辑多次。** 每当有人看到 Claude 犯错时，他们就会添加一条规则。把它当作一份随时在进化的“活文档”来对待。 * **`Esc` twice opens rewind.** Combined with checkpoints: try risky things, find out they failed, rewind cleanly. **连按两下 `Esc` 打开倒带 (rewind) 面板。** 结合检查点使用：放手去尝试有风险的事情，发现失败了，就干脆利落地“倒带”回去。 * **For UI changes, set up Playwright MCP.** Boris uses the Chrome extension every time he works on web code, where Claude opens a browser, clicks around, and verifies. **对于 UI 变更，请设置 Playwright MCP。** Boris 每次处理 Web 代码时都会使用 Chrome 扩展，Claude 可以借此打开浏览器、四处点击并进行视觉验证。 * **Install a language server plugin.** Type errors and unused imports caught after every edit. Highest-impact plugin you can install. **安装 Language Server (语言服务器) 插件。** 每次编辑后都能捕捉到类型错误和未使用的导入包。这是你能安装的作用最大的插件。 * **Use `/voice` for prompting.** You speak 3x faster than you type, and prompts get way more detailed as a result. **使用 `/voice` (语音) 来提供提示词。** 你说话的速度比打字快 3 倍，而且提示词的内容会因此变得详细得多。 * **Auto mode + `/focus` + `/goal`.** Crisp brief, set the goal, walk away. Come back to a finished PR. **自动模式 + `/focus` + `/goal`。** 给定简明扼要的任务，设定好目标，转身离开。回来时迎接一个已经完成的 PR。 * **Use `Ctrl+G` to edit Claude’s plan in your editor before implementation.** Faster than typing corrections in the chat. **在 Claude 开始代码实现之前，使用 `Ctrl+G` 在编辑器中修改它的计划。** 这比在聊天框里打字纠正要快得多。 * **Ask Claude to draw ASCII diagrams of new protocols and codebases.** Boris’s tip for understanding unfamiliar code quickly. **让 Claude 用 ASCII 字符画出新协议和代码库的架构图。** 这是 Boris 快速理解陌生代码的绝招。 --- ## 📚 12. Resources ## 📚 12. 资源 **Official docs (官方文档)** * [Claude Code documentation](https://code.claude.com/docs/en/overview) * [Explore the .claude directory](https://code.claude.com/docs/en/claude-directory) * [Best practices for Claude Code](https://code.claude.com/docs/en/best-practices) * [Memory (CLAUDE.md, rules, auto memory)](https://code.claude.com/docs/en/memory) * [Skills](https://code.claude.com/docs/en/skills) · [Subagents](https://code.claude.com/docs/en/sub-agents) · [Plugins](https://code.claude.com/docs/en/plugins) · [MCP](https://code.claude.com/docs/en/mcp) · [Hooks](https://code.claude.com/docs/en/hooks) **Boris and the team (Boris 与团队访谈资料)** * [How Boris Uses Claude Code](https://howborisusesclaudecode.com/): 89+ tips from the creator, sourced from his X threads * [Anthropic blog: Best practices for Opus 4.7 with Claude Code](https://claude.com/blog/best-practices-for-using-claude-opus-4-7-with-claude-code) * [shanraisshan/claude-code-best-practice](https://github.com/shanraisshan/claude-code-best-practice) **Skills (技能仓库)** * [mattpocock/skills](https://github.com/mattpocock/skills): “Skills for Real Engineers” * [Jeffallan/claude-skills](https://github.com/Jeffallan/claude-skills): 66 language-specific skills * [addyosmani/web-quality-skills](https://github.com/addyosmani/web-quality-skills): web performance and quality * [Anthropic skills cookbook](https://platform.claude.com/cookbook/skills-notebooks-01-skills-introduction) **Subagents (子智能体仓库)** * [VoltAgent/awesome-claude-code-subagents](https://github.com/VoltAgent/awesome-claude-code-subagents): 100+ organized by category * [hesreallyhim/a-list-of-claude-code-agents](https://github.com/hesreallyhim/a-list-of-claude-code-agents) **Plugins and marketplaces (插件与市场)** * [Chat2AnyLLM/awesome-claude-plugins](https://github.com/Chat2AnyLLM/awesome-claude-plugins): 1000+ plugins across 75+ marketplaces * [claudemarketplaces.com](https://claudemarketplaces.com/) **MCPs (MCP 服务器)** * [Obsidian Claude Code MCP plugin](https://github.com/iansinnott/obsidian-claude-code-mcp) * [Official MCP servers list](https://github.com/modelcontextprotocol/servers) * [claude.com/partners/mcp](https://claude.com/partners/mcp) --- ## 📝 Closing Notes ## 📝 结语 Claude Code clicked for me only after I stopped treating it like ChatGPT in a terminal. You stop thinking “I need to write this code” and start thinking “I need to set up Claude to write this code well.” The setup is the work. The execution is mostly verification. 只有当我不再把 Claude Code 当成是一个放在终端里的 ChatGPT 时，我才真正领悟到了它的强大。你要停止去想“我需要写这行代码”，而是开始思考“我需要为 Claude 设置好环境，让它能写好这行代码”。配置这些环境才是真正的工作，而执行过程大部分时间只需你进行验证即可。 A few things that have genuinely changed how I work: 以下几点彻底改变了我的工作方式： **`CLAUDE.md` is compounding infrastructure.** Every mistake Claude makes is a rule waiting to be written. After a few weeks of “update CLAUDE.md so you do not repeat this”, the same prompts produce dramatically better output. **`CLAUDE.md` 是具备复利效应的基础设施。** Claude 犯的每一个错，都是一条等待被写入的规则。在持续几周对它说“更新 CLAUDE.md，这样你就不会再犯同样的错误”之后，相同的提示词将会产出极其出色的结果。 **`CLAUDE.local.md` captures PR feedback.** Your reviewers are giving you free training data. Convert recurring feedback into rules. Let Claude apply them next time. **`CLAUDE.local.md` 用于捕获 PR 反馈。** 代码审查者实际上在为你提供免费的训练数据。将反复出现的反馈转化为规则，让 Claude 下次直接替你应用它们。 **Skills are the unit of reusable expertise.** If you find yourself prompting the same instructions twice, that is a skill waiting to be written. **Skills 是可复用专业知识的基本单元。** 如果你发现自己在重复输入相同的提示词，那么这正是一个亟待编写的 skill (技能)。 **Subagents over kitchen-sink prompts.** Separate concerns, keep contexts clean, and the quality of every individual task goes up. **子智能体 (Subagents) 优于“大杂烩”式的提示词。** 分离关注点，保持上下文干净，每一个独立任务的完成质量都会直线上升。 **Parallel sessions are the unlock everyone underestimates.** Three Claudes in three worktrees is a different kind of leverage. Try it for a day. **并行会话是被大家严重低估的潜力解锁键。** 在三个工作树里同时运行三个 Claude，这是一种截然不同的工作杠杆。去试一天看看吧。 The real shift happens when you stop thinking of Claude Code as a tool you use to write code and start thinking of it as something you train, configure, and operate. Most people stop at the prompts. Going past that, into the directory structure, skills, agents, plugins, and MCPs, is where it stops feeling like a tool and starts feeling like a teammate. 当你不再将 Claude Code 仅仅视为一个用来写代码的工具，而是把它当成一个需要你去训练、配置和运营的实体时，真正的质变才会发生。大多数人只停留在了写提示词的阶段。跨过这个阶段，深入到它的目录结构、skills、智能体、插件以及 MCP 之中，它将不再像是一个工具，而会真正成为你的队友。