收编 · Skills 与输出风格Codify · Skills and Output Styles.
你有一套反复输入的提示：同样的审查清单、同样的发布步骤、同样的「用这个口吻写」。每次重打一遍，既慢又不稳。Claude Code 给了两个收编的去处：重复的流程连脚本带资源打包成 Skill，一个 /名字 唤起、用到才加载，不挤占上下文 —— 从前的自定义 /命令 也已并入这里；要改的若是它整场对话的角色和口吻，那是输出风格的事，在设置里切换。这一章讲两者怎么分工：哪些活沉成 skill、哪些交给风格、哪些根本不值得固化。You have prompts you retype constantly: the same review checklist, the same release steps, the same 'write it in this voice.' Retyping each time is slow and inconsistent. Claude Code gives them two places to settle: a repeated procedure packs into a Skill — scripts and resources included, summoned by one /name, loaded only when used so it doesn't crowd the context; custom /commands have been folded in here. If what you're changing is the role and voice of the whole conversation, that's an output style, switched in settings. This chapter splits the work: what settles into a skill, what goes to a style, and what isn't worth codifying at all.
9 分钟 · 初稿 2026.079 Min · Drafted 2026.07
你有几段每周都要重打的提示:发版前那套检查、审 PR 的固定清单、「用我们团队的口吻写 changelog」。每次凭记忆重打,慢,而且每次都略有不同 —— 于是结果也略有不同。这一章讲收编:把你的重复劳动变成两类资产,流程沉成 Skill,说话方式定成输出风格。收编过的东西不再依赖你的记性,也不再吃你的手速。You have prompts you retype weekly: the pre-release checklist, the fixed PR-review list, 'write the changelog in our team's voice.' Retyped from memory, each is slow — and slightly different, so the results differ slightly too. This chapter is about codifying: turning that repetition into two kinds of assets — procedures settle into Skills, manner of speaking into an output style. What's codified no longer depends on your memory, or your typing speed.
— I
Skills:把流程存成一个词Skills: A Procedure Stored as One Word.
一个 skill 就是一个目录加一份 SKILL.md:frontmatter 里一句 description 说清「什么时候该用我」,正文写做法,旁边还能放脚本、模板、示例文件。存好之后有两条入口:你打 /名字 直接唤起,或者 Claude 自己判断相关就加载。1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.关键的经济学在加载方式:描述常驻、正文用到才进上下文 —— 一份几百行的发版手册,平时几乎不占地方。1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.这也是它和 CLAUDE.md 的分工:每次都要的一句事实住第 3 章,偶尔才用的一套流程住这里。A skill is a directory with a SKILL.md: one description line in the frontmatter says when to use it, the body says how, and scripts, templates, and examples can sit alongside. Once saved, two entrances: type /name to summon it, or Claude loads it when it judges the task relevant.1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17. The key economics is in the loading: the description stays resident, the body enters context only when used — a several-hundred-line release manual costs almost nothing on idle days.1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17. That's also the split with CLAUDE.md: a fact needed every session lives in chapter 3; a procedure needed occasionally lives here.如果你记得「自定义 slash 命令」:它们已经并入 skills —— .claude/commands/deploy.md 和 .claude/skills/deploy/SKILL.md 生成同一个 /deploy,旧文件继续可用,skills 只是多了支撑文件、调用控制这些新能力。1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.一份真实形状的 skill 长这样:If you remember 'custom slash commands': they've been merged into skills — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md create the same /deploy, old files keep working, and skills simply add supporting files and invocation control on top.1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17. A real-shaped skill looks like this:
三个 frontmatter 字段值得先认识。disable-model-invocation: true:只许你手动触发 —— 发版、发消息这类有副作用的流程,别让它「觉得代码差不多了」就自己跑。1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.allowed-tools:这一轮预批准的工具,让 /release 跑 git 命令不用逐条弹窗。1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.context: fork:整个 skill 丢进 subagent 的独立上下文里执行,重活不弄脏主对话(第 5 章的机制)。1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.正文里那行 !`git status` 是动态注入:调用瞬间先执行命令、把输出内联进内容 —— 你的 skill 拿到的是活数据,不是你写死的假设。1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.Three frontmatter fields worth meeting first. disable-model-invocation: true: only you can trigger it — release and send-message flows with side effects shouldn't run because it 'felt the code was ready.'1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.allowed-tools: tools pre-approved for that turn, so /release runs git without a prompt per command.1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.context: fork: the whole skill executes in a subagent's separate context, keeping heavy work out of your main conversation (chapter 5's machinery).1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17. And that !`git status` line is dynamic injection: the command runs at invocation and its output is inlined — your skill reads live data, not the assumption you hard-coded.1注 1Note 1Claude Code Docs · Skills —— 一个 skill 是一个目录加一份 SKILL.md:frontmatter 写清何时该用,正文是做法;描述常驻上下文,正文只在被调用或被判定相关时加载;/名字 直接唤起。「自定义命令已并入 skills」—— .claude/commands/deploy.md 与 .claude/skills/deploy/SKILL.md 同样生成 /deploy,旧命令文件继续可用。常用 frontmatter:disable-model-invocation(只许你手动触发)、user-invocable: false(只许它自己用)、allowed-tools(该轮免批工具)、context: fork(丢进 subagent 跑)。正文里还能用 $ARGUMENTS 参数替换、!`命令` 动态注入(执行后把输出内联进内容)。位置:企业 / 个人 ~/.claude/skills/ / 项目 .claude/skills/ / 插件;SKILL.md 建议 500 行内,细料放支撑文件。截至 2026-07-17。Claude Code Docs · Skills — a skill is a directory with a SKILL.md: frontmatter says when to use it, the body says how; the description stays in context while the body loads only when invoked or judged relevant; /name summons it directly. 'Custom commands have been merged into skills' — .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working. Useful frontmatter: disable-model-invocation (only you can trigger it), user-invocable: false (only Claude can), allowed-tools (pre-approved for the invoking turn), context: fork (run in a subagent). In the body, $ARGUMENTS substitution and !`command` dynamic injection (runs the command and inlines its output) are also available. Locations: enterprise / personal ~/.claude/skills/ / project .claude/skills/ / plugin; keep SKILL.md under ~500 lines with detail in supporting files. As of 2026-07-17.
— II
输出风格:换的不是知识,是口吻Output Styles: Not What It Knows — How It Speaks.
第二根柱子管的是另一类重复:「解释详细点」「别铺垫直接干」「像导师一样带我」—— 这些不是流程,是整场对话的角色设定。输出风格直接改 system prompt 里的角色、口吻与输出格式,一次设定,每一轮生效。2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17.文档的界定句很准:改变它怎么回应,不改变它知道什么。2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17.The second pillar handles a different repetition: 'explain more,' 'skip the preamble, just do it,' 'walk me through it like a mentor' — not procedures, but the whole conversation's role setting. An output style modifies the system prompt's role, tone, and output format: set once, applied every turn.2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17. The docs' boundary line is exact: it changes how it responds, not what it knows.2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17.内置四款够覆盖多数场景:Default 是原生工程师;Proactive 更敢直接动手、少停下来问;Explanatory 边干活边给你讲原理;Learning 最有意思 —— 它会在代码里留 TODO(human),把关键的一小段留给你亲手写,拿它学新代码库比纯看快得多。2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17.切换走 /config 里的 Output style 选项,或直接在设置文件里写 outputStyle(独立的 /output-style 命令已经移除)。2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17.自定义也简单:markdown 文件放进 .claude/output-styles/,用 keep-coding-instructions 决定保不保留原生编码指令 —— 保留,它是个换了口吻的工程师;去掉,它可以彻底变成写作助手或数据分析员。2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17.两条边界记住:风格只作用于主对话,派出去的 subagent 不受影响;会话中途换风格不生效,下个会话才加载。2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17.The four built-ins cover most needs: Default is the stock engineer; Proactive acts first and pauses less; Explanatory teaches while it works; Learning is the most interesting — it leaves TODO(human) markers in the code so you write the key passage yourself, which beats pure reading for learning a new codebase.2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17. Switch via the Output style entry in /config, or set outputStyle in a settings file (the standalone /output-style command is gone).2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17. Custom styles are simple too: a markdown file in .claude/output-styles/, with keep-coding-instructions deciding whether the stock coding instructions stay — keep them and it's an engineer with a new voice; drop them and it becomes a writing assistant or data analyst outright.2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17. Two boundaries to remember: styles apply to the main conversation only — subagents are unaffected; and a mid-session switch doesn't take effect until the next session.2注 2Note 2Claude Code Docs · Output styles —— 「输出风格改变 Claude 怎么回应,不改变它知道什么」:修改 system prompt 里的角色、口吻与输出格式,作用于整场对话。内置四款:Default、Proactive(先动手少停顿)、Explanatory(边做边讲解)、Learning(留 TODO(human) 让你亲手写一段);自定义放 .claude/output-styles/,keep-coding-instructions 决定保不保留原生编码指令。经 /config 选择或直接改 outputStyle 设置(独立的 /output-style 命令已移除);只作用于主对话,subagent 不受影响。截至 2026-07-17。Claude Code Docs · Output styles — 'output styles change how Claude responds, not what Claude knows': they modify the system prompt's role, tone, and output format for the whole conversation. Four built-ins: Default, Proactive (acts first, pauses less), Explanatory (teaches while working), Learning (leaves TODO(human) markers for you to write a piece yourself); custom styles live in .claude/output-styles/ with keep-coding-instructions deciding whether the stock coding instructions stay. Set via /config or the outputStyle setting (the standalone /output-style command has been removed); applies to the main conversation only — subagents are unaffected. As of 2026-07-17.
— III
什么收、收到哪、什么不收What to Codify, Where — and What Not To.
到这里,固化的去处一共四个。判断题每次只有一道:这段重复,是流程、是事实、是口吻,还是铁律?By now there are four places for codification. The question each time is single: is this repetition a procedure, a fact, a voice, or an iron rule?
这段重复是The repetition is
收到Codify into
性质Nature
多步流程、带脚本的做法A multi-step procedure, with scripts
Skill(本章)A skill (this chapter)
按需加载,模型执行Loads on demand, model executes
每次都要的一句事实A fact needed every session
CLAUDE.md(第 3 章)CLAUDE.md (chapter 3)
常驻上下文Always in context
整场对话的角色与口吻The conversation's role and voice
输出风格(本章)An output style (this chapter)
改 system promptModifies the system prompt
必须每次都发生的动作An action that must fire every time
Hook(第 7 章)A hook (chapter 7)
harness 执行,不经模型Harness-run, no model in the loop
头两行的分界线,官方给的触发律是「第三次」:同一段提示重打到第三次、同一套做法第三次贴进对话,就存成 skill。3注 3Note 3Claude Code Docs · Extend Claude Code —— 收编的触发时机:同一段提示重打到第三次、或同一套多步流程第三次贴进对话,存成 skill;hook 与 skill 的分界:hook 在事件上必然执行,skill 由模型理解并执行、结果可能有出入。截至 2026-07-17。Claude Code Docs · Extend Claude Code — codification triggers: a prompt retyped a third time, or a multi-step playbook pasted a third time, becomes a skill; the hook/skill line: a hook always fires on its event, while a skill is interpreted and executed by the model, so outcomes can vary. As of 2026-07-17.末行的分界线更硬:skill 是模型理解后执行的,结果可能有出入;hook 在事件上必然执行 —— 「希望它跑 lint」是 skill 的事,「必须跑过 lint 才许提交」是 hook 的事。3注 3Note 3Claude Code Docs · Extend Claude Code —— 收编的触发时机:同一段提示重打到第三次、或同一套多步流程第三次贴进对话,存成 skill;hook 与 skill 的分界:hook 在事件上必然执行,skill 由模型理解并执行、结果可能有出入。截至 2026-07-17。Claude Code Docs · Extend Claude Code — codification triggers: a prompt retyped a third time, or a multi-step playbook pasted a third time, becomes a skill; the hook/skill line: a hook always fires on its event, while a skill is interpreted and executed by the model, so outcomes can vary. As of 2026-07-17.反过来,三类东西不值得收:只做一次的活,写清楚 prompt 就够;还在每周变的流程,固化了就是提前腐烂;以及模型本来就会做对的事 —— 给它立规矩,和给它写说明书一样,都有维护成本。For the first two rows, the official trigger is 'the third time': a prompt retyped a third time, a playbook pasted a third time — make it a skill.3注 3Note 3Claude Code Docs · Extend Claude Code —— 收编的触发时机:同一段提示重打到第三次、或同一套多步流程第三次贴进对话,存成 skill;hook 与 skill 的分界:hook 在事件上必然执行,skill 由模型理解并执行、结果可能有出入。截至 2026-07-17。Claude Code Docs · Extend Claude Code — codification triggers: a prompt retyped a third time, or a multi-step playbook pasted a third time, becomes a skill; the hook/skill line: a hook always fires on its event, while a skill is interpreted and executed by the model, so outcomes can vary. As of 2026-07-17. The last row's line is harder: a skill is interpreted by the model, so outcomes can vary; a hook always fires on its event — 'hoping it runs lint' is skill territory, 'no commit until lint passes' is hook territory.3注 3Note 3Claude Code Docs · Extend Claude Code —— 收编的触发时机:同一段提示重打到第三次、或同一套多步流程第三次贴进对话,存成 skill;hook 与 skill 的分界:hook 在事件上必然执行,skill 由模型理解并执行、结果可能有出入。截至 2026-07-17。Claude Code Docs · Extend Claude Code — codification triggers: a prompt retyped a third time, or a multi-step playbook pasted a third time, becomes a skill; the hook/skill line: a hook always fires on its event, while a skill is interpreted and executed by the model, so outcomes can vary. As of 2026-07-17. And three things aren't worth codifying: one-off work, where a clear prompt suffices; procedures still changing weekly, which codified is just rot on a schedule; and things the model already gets right — rules, like manuals, carry maintenance costs.动手 · 本周就收编三样:Hands-on · codify three things this week:
01
把重打过两次的 prompt 存成第一个 skillSave your twice-retyped prompt as a first skill
找出你本周重打过的那段提示,原样放进 .claude/skills/<名字>/SKILL.md 的正文,description 写清触发时机。下次用 /名字 唤起。Take the prompt you retyped this week, put it verbatim into the body of .claude/skills/<name>/SKILL.md, and write the trigger into the description. Summon it with /name next time.
02
给一个有副作用的流程上锁Lock one procedure with side effects
发版、发通知、清数据 —— 挑一个,加 disable-model-invocation: true,让触发权只在你手里。Release, notify, purge — pick one, add disable-model-invocation: true, and keep the trigger in your hands only.
03
试一周 Learning 或 Explanatory 风格Run a week on Learning or Explanatory
在你最近接手的陌生代码库里开一周,再决定回不回 Default —— 风格是可逆的,试错成本几乎为零。Turn it on for a week in the unfamiliar codebase you just picked up, then decide whether to return to Default — styles are reversible, and the trial costs nearly nothing.
重打第三遍的提示,
是一个还没存档的 skill.
A prompt retyped a third time
is a skill not yet saved.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return