我在 Claude Code、Codex 和 Pi-Bolt 之间换着用，经常遇到这种情况：只想让 AI 帮忙写个页面，配置里先冒出一排名词，像来办事之前还得考个证。
拿“根据 Figma 设计稿实现页面”来说，MCP 可以负责接入设计稿，Skill 规定图片怎么处理，Workflow 安排确认和执行的顺序，Rules 限制哪些文件不能动，Memory 留下下次还需要的约定。
它们可以一起用，分工大致如下：
各家产品的命名还会重叠。一个 Skill 可以包含 Workflow，某个客户端也可能把规则文件放在“Memory”菜单里。判断它做什么，得看加载时机、维护方式和执行行为，光认菜单名容易认错亲戚。
后面提到的 Agent，就是能读写文件、调用工具的 AI 助手；宿主指运行它的客户端。
一、先从处理设计稿图片说起
我做小程序开发时，经常要从 Figma 导出素材，压缩 PNG，上传腾讯云 COS，复制地址，再放回代码。页面还没开始写，我已经先给图片当起了快递员。
遇上一个页面有好几张图，就得把这套操作重复好几遍。后来我把它做成了 image-auto-upload Skill，带两个脚本，一个压缩 PNG，一个上传 COS，再把素材处理接到页面开发里。
当时我用 Figma MCP 读取设计稿。图片怎样压缩、传到哪里、怎样返回链接，由 Skill 说明；先确认素材清单再上传，是这项任务的 Workflow。密钥不能出现在代码和日志里，则适合留在长期规则中。
实际调用压缩服务和 COS SDK 的是脚本。AI 判断当前任务需要走哪些步骤，通过宿主工具运行它们。
如果已有脚本能完成任务，Agent 也能运行命令，就可以先这么用。没必要为了把配置凑齐，再给每个脚本套一层 MCP。图片还没传上去，先多维护一个服务，这笔账不太划算。
旧文采用了上传成功后清理源图的做法。现在给新手设计类似流程，我会先保留源文件，核验远端结果后，再按约定清理。密钥即使放进不提交的环境文件，也要防着日志、截图和工具输出把它带出去。
二、Rules：把反复纠正的事写下来
如果刚开始用 AI 编程，我会先写几条规则。
“不要顺手把其他问题也改了”“工作区原来的修改，别算成你这轮做的”“测试没跑，就写没跑”，这些要求可能已经说过很多次。只想补个按钮时，我通常不期待收到一份顺便重构半个仓库的惊喜。
这类要求适合放进长期指令文件。Claude Code 常见的是 CLAUDE.md，Codex 和 Pi 常见的是 AGENTS.md。客户端决定文件怎么发现、在哪个范围加载。
项目里可以先写这么几条：
# 项目协作约定
- 修改前先检查工作区，保留与当前任务无关的改动。
- 沿用项目现有组件和依赖；新增依赖前说明原因。
- 接口失败必须保留错误信息，不用空数据冒充成功。
- 修改后运行相关检查，并记录通过、失败和未运行项。
- 没有明确授权，不暂存、提交、推送代码。
每条都能对应到一次具体行为。出了问题，也方便检查是哪条没被遵守。我更愿意写这些，至于要不要先授予 AI 一个“世界顶级工程师”的头衔，可以缓缓。
别让所有规范都跟着每个任务上班
个人所有项目都适用的偏好，放用户级；团队共享的构建命令和目录约定，放项目级；只管接口目录的规范，就尽量限定到那个范围。
例如，Claude Code 可以把路径规则放在 .claude/rules/api.md：
---
paths:
- "src/api/**/*.ts"
---
修改接口调用时，沿用现有请求封装。
不要在页面组件里重复拼接服务端地址。
这是 Claude Code 的具体语法，换到其他工具要另查支持方式。
拆成多个文件，也要检查它们怎样加载。主文件如果把所有规范都直接引入，模型每次照样可能读一大堆。想减少无关上下文，需要配合客户端支持的路径范围或按需读取机制，不能只给文件搬个家。
规则用久了也得清理。项目从 npm 换成 pnpm，旧命令就该更新；临时排障加的限制，结束后再看看还要不要留。
对于确实不能允许的动作，Markdown 里的“禁止”只是一条指令。权限、沙箱、命令检查和 CI 可以在相应环节拦截，具体能拦到哪儿，需要核实。文件里把“严禁”写三遍，也不会自动多出一道门禁。
三、Skill：这类活，下次照这份做
拿不准 Skill 怎么写时，我会先用 skill-designer 整理结构。它帮助梳理适用场景、入口说明和参考资料的分工。在 cc，pi、codex等agent 中，可以这样提要求：
/skill-designer
帮我把“排查页面问题”整理成一个 Skill。
先确认适用场景和输入，区分只读分析与修改代码。
列出连接失败时的处理方式，以及交付前的检查点。
说明哪些内容放入口，哪些拆成参考文档，哪些适合写成脚本。
这是我自己使用的 Skill，要先安装到客户端能发现的位置。不同工具的调用语法也不一样，后面有对照。
草稿出来以后，仍要拿真实任务试。看看会不会误触发、漏步骤，失败后能不能如实停下。目录看着整齐，和实际干活靠谱，还得分别检查。
哪些任务值得做成 Skill？
同一套要求开始反复粘贴，或者一次返工留下了可复用的方法，我就会考虑做成 Skill。下次再做，至少不用先去聊天记录里考古。
图片上传、代码审查、排查某类接口问题，都可以整理。先把任务入口和完成标准想清楚，再决定需要哪些文件。
常见目录是这样：
page-check/
├── SKILL.md
├── references/
│ └── checklist.md
└── scripts/
└── check-output.mjs
SKILL.md 说明什么时候用、按什么步骤做。详细检查项可以放进 references，稳定重复的计算和校验交给脚本。没有脚本也能做 Skill，不必为了让目录显得专业，先建一排空文件夹。
第一份 Skill，写清触发条件和交付物就很好
这是一个简化的页面排查示例，用来展示写法，尚不是验证过的通用排障工具：
---
name: page-check
description: 排查运行中的页面显示和交互问题。适用于用户要求复现页面故障；仅阅读代码时不使用。
---
1. 确认目标页面、运行环境和用户看到的异常。
2. 阅读项目的启动与验证说明，再打开实际页面。
3. 记录复现步骤，检查相关 DOM、控制台和网络请求。
4. 对照代码定位原因。没有修改授权时，只给分析结论。
5. 获得修改授权后修复，并按原步骤复测。
6. 交付原因、改动范围、验证结果和未验证项。
浏览器连接失败时，说明阻塞；不要把代码推断写成现场复现。
详细检查项见 references/checklist.md，进入页面检查时再读取。
它说明了该做什么，也说明了连接失败、没有修改授权时怎么处理。
描述里最好写上适用和不适用的情况。“优化开发体验”太宽，模型很难决定什么时候该用。范围一大，问个命令都可能被拖去做全套体检。
“渐进式披露”，就是用到哪儿读到哪儿
在我用的 Pi-Bolt 中，模型先看到 Skill 的名称、描述和路径，决定使用后再读取正文。正文可以指引它，到某一步再读相应的详细资料。
Codex、Claude Code 也采用按需加载的思路，具体字段、发现规则和上下文预算有所不同。
我之前做微缩动画 Skill，就把画风、参考图处理、运动检查、音频和导出分开。设计场景时不用读整份编码说明，到了正式出片再读。
在关于 Skill 与 Prompt、新街口微缩动画的那篇文章里，我写过一句：“拆文件还不够，入口得写清楚读取时机。”
至于“Skill 到底是不是 Prompt”，正文当然还是给模型看的指令。我保存的是这类任务的说明、资料、模板和脚本，以及何时读取它们。换个文件名，模型不会多长一个脑子。
四、MCP：需要外部接入时，再来配它
MCP 的全称是 Model Context Protocol，可以先理解成客户端连接工具和数据服务的一套协议。浏览器服务可以提供读取页面、检查请求的能力，文档服务提供搜索和阅读，设计工具服务则提供设计稿信息。
连接成功以后，排查方法、目标环境、登录状态和权限仍需确认。浏览器工具能打开页面，接下来该复现问题、查请求还是对照代码，还得有一套做法。
什么时候值得装
希望 AI 直接检查运行中的页面、读取设计稿、查询文档服务时，可以看看有没有成熟的 MCP 接入。
如果只是读本地文件、运行测试、处理几张图片，而客户端已经能读文件和运行命令，就先用现成能力。CLI、SDK、HTTP API 都可能满足需求。
我用 Tauri 开发的这个桌面应用，就有个容易踩的坑：通过 MCP 桥接检查实际界面时，开发版必须带 dev-mcp 特性启动。普通开发命令也能把应用打开，桥接能力却没编进去。看着窗口好好的，连接就是不通，挺适合让人怀疑人生。
这种启动要求适合写进项目 Skill，让后续验证先读说明。
先连一个只读服务，实际查一次
例如，Codex 可以在 ~/.codex/config.toml 中配置官方文档 MCP：
[mcp_servers.openaiDeveloperDocs]
url = "https://developers.openai.com/mcp"
配置后检查连接状态，再让它查一条文档，核对返回内容和来源。这个官方只读文档服务，不会因此得到你的数据库或云账户权限。
Claude Code、Pi 的格式和位置不同，要按各自文档配置，别把 TOML 原样粘进 JSON。
有些 MCP 通过远程网址连接，有些需要在本地启动服务程序，通过标准输入输出通信，常称为 stdio。后者可能依赖 Node、Python 等环境，还要处理服务端安装和文件访问授权。
列表里出现名字后，继续检查客户端是否读到了配置、服务是否连上、需要的工具能否调用，以及调用结果是不是来自正确的环境。能读公开页面，不代表能进登录后的后台；能查测试库，也不代表该给生产库写权限。
浏览器尤其要看清它连的是哪个窗口和会话。我习惯复用可见浏览器，方便确认环境。会不会另起浏览器、支持哪种连接方式，则取决于具体服务。
Chrome DevTools MCP：查网页时，别把我挤下线
我之前遇到过，新开一个 AI 会话，Chrome 就弹连接授权；让 AI 打开网页，它又切走我正在看的标签页。换成独立浏览器，登录态却分开了。原本想让它帮忙查个页面，结果我还得先给它安排座位。
当时的完整配置和脚本在：《chrome-devtools-mcp 总弹授权、抢标签页？我是这样一步步配好的》。
那次是 macOS、Codex CLI 和官方包 1.12.0 的实践，分别处理了这些问题：
--autoConnect 负责选择连接目标；自写的 quiet 启动器转发调用、复用连接，并约束后台操作。这里的 quiet 是额外适配代码，没有一个官方“quiet 开关”包办这几件事。
后台创建页面用 background: true，选择任务页面但不抢前台用 bringToFront: false。除了写规则提醒，我当时还在适配脚本里固定了这两个参数，减少漏传。
首次连接仍要用户允许。Chrome、连接或常驻进程重启后，也可能再次要求授权。客户端批准工具调用，和 Chrome 允许远程调试，是两层权限，前者不能代替后者。
配好后要另开一个 AI 会话测试：是否复用原进程，能否后台开页，原标签页是否仍选中。当前会话成功一次，还没验证到跨会话复用。日常也别每完成一个任务，就顺手把共用进程停掉。
在这套配置里，MCP 提供浏览器操作，配套 Skill 介绍用法，Rules 约束不要动用户原有标签页，适配代码处理连接复用和固定参数。启动器依赖特定包版本的内部模块，升级时需要一起核查，也不能直接套到所有系统。
服务装得多，还要考虑上下文成本。工具定义、调用结果和日志都可能占空间；部分客户端支持工具发现或延迟加载，实际开销不能只按 MCP 数量算。先留当前任务会用到的服务，连接出了问题也好找。
五、Workflow：把顺序和确认条件写明白
有的产品把可重复调用的 Markdown 步骤叫 Workflow，有的提供能保存状态、执行分支的工作流引擎。在 Skill 里组织一套步骤，也是在设计工作流。
所以，Skill 可以装下工作流：前者是能力包，后者描述执行过程。遇到具体产品，还得看它把哪些功能叫作 Workflow。
图中街景为示意，实际作品见前面引用的新街口动画回答。我制作时采用的顺序是：
阅读参考图，确认风格要求
生成同一动画的四张实际样帧
等待我确认当前版本
执行检查，再渲染全片
如果画面或剧情有改动，重新确认
“先让我看看”说得太笼统。看概念图还是实际帧？认可车辆造型，算不算批准整部片子？改了镜头以后，旧批准还有效吗？一句夸奖要是能附送整部片子的制作授权，这优惠力度未免太大。
后来我补了批准记录与版本核对，出片前由程序检查批准是否对应当前版本。Markdown 说明该遵循的顺序，代码检查在指定执行入口拒绝不符合条件的请求。
开发任务也可以这样组织：
复现问题并记录证据
确认原因和修改范围
修复并补相关测试
按原步骤复测
检查 diff，交付验证结果
确认环节放在会扩大范围、产生费用、覆盖数据或影响外部系统的步骤。改一句文案就停下来审批五次，我也受不了；能连续完成的部分，就让它连续做。
需要无人值守、固定分支或可靠重试时，可以把适合确定执行的部分交给脚本、CI 或工作流引擎，由 AI 处理需要判断的步骤。自然语言清单未必能独自承担这些要求。
六、Memory：下次别让我从头解释
隔几天换了会话，我仍希望 AI 能找到之前确认的约定：某个项目暂时不能升级依赖，某个环境只能用特定方式启动，或者先做分析，得到授权后才能修改代码。
这些信息可以保存为 Memory。稳定偏好、已确认决策和仍有效的环境限制，适合长期留着。今天读到哪个文件、某项测试还没跑，则更适合当天记录或任务进度。什么都永久收藏，最后只好在便签里翻便签。
保存以后，还得试试能不能找回来
文件叫 MEMORY.md，不意味着模型每次完整读取，更不能证明内容都正确。要看客户端加载哪些部分、有没有按需检索，新增记录多久生效，旧信息怎么修改或删除。
我在 Pi-Bolt 中用了 pi-memory 扩展，把长期记录、每天日志和待办分开保存，提供读写、检索和遗忘等工具。搜索还依赖可选的 qmd 检索工具；按意思找相近内容的语义检索，需要相应的索引和模型。这是我的扩展配置，安装 Pi 本身并不默认附带这一整套记忆。
Claude Code 有自己的 auto memory，和人工维护的 CLAUDE.md 在写入方式、用途上不同。换到 Codex，也要检查对应版本和配置支持什么。双方都能读 Markdown，离自动同步记忆还差不少事情。
在已经配置好记忆能力的客户端里，可以这样提要求：
请记住：在这个项目里，没有明确授权，不要暂存或提交代码。
请保存为项目长期约定，告诉我写到了哪里。
不要把测试账号、密钥或本轮临时日志存进长期记忆。
然后换个新会话，问它约定是什么、从哪里读到，再检查保存的文件。不能只听到一句“记住了”就放心，我自己的记性都没好到敢这么保证。
留下范围和原因，也记得清理
一条依赖限制可以这样记录：
范围：当前项目的支付依赖。
结论：暂时保持现有版本。
原因：新版本兼容性尚未验证。
重新评估：完成兼容性验证后，再确认是否升级。
这是示意，实际原因和批准人得来自真实讨论，不能让模型补个顺耳的理由。决策或环境变了，就更新、删除旧记录，否则同一个错误会跟着记忆跨会话复活。
规则和记忆也有交集。“未经授权不推送”既可以作为个人偏好被保存，也可以写进需要长期执行的规则。记忆强调保存与召回，规则强调希望它遵守什么。
操作生产环境时仍须重新核实。Memory 可以提示上次的情况，今天的环境却可能已经变了。
七、命令、脚本、Hooks，又分别管什么？
同样一个斜杠入口，换到不同工具，背后的机制也可能不同。
Pi 的 Prompt Template 可以把常用提示变成 /review。只想少粘贴一段审查提示，用模板就行；需要参考资料和完整处理方法，再考虑 Skill。
Claude Code 已把自定义 Commands 合并进 Skills 的使用体系，旧目录仍兼容。别看都长成 /名字，就直接认作同一套机制。
脚本执行固定逻辑。文件摘要、数据转换、格式校验，这类结果明确的工作，我愿意让 Agent 调用现成脚本，省得每次重新推演一遍。
Hooks 则在某个事件发生时执行配置的命令或检查，适合在指定环节提醒、拦截。某个 Hook 能覆盖哪些入口，需要结合宿主实现验证。在 Skill 里写“工具调用前检查”，不会自动生成一个 Hook；装了 Hook，也不能据此认定所有操作都被安全隔离。
插件可以打包 Skill、连接和其他资源，方便分发。Pi 扩展还能注册工具或响应事件，增加宿主能力。等手头任务需要这些能力时再用，不必提前把编辑器装修成工具展览馆。
八、换成 Pi、Codex、Claude Code，怎么落地？
常见本地入口如下。~ 是用户主目录；项目目录里的配置则对相应项目生效。加载行为仍受版本和设置影响。
我本机启动的是 pi-bolt，修改 Skill 后用 /reload 刷新。上游 Pi、Pi-Bolt、不同版本以及已安装扩展会有差异，我本机能用的功能，未必全是裸 Pi 的默认配置。
Claude Code 当前也有条件支持读取 AGENTS.md，受版本、项目是否已有 CLAUDE.md 和相关设置影响。多工具共用规范时，要核对实际加载来源；需要兼容时，可以在 CLAUDE.md 中用 @AGENTS.md 引入共享内容。
尽量共用内容，接入差异分别处理
我会尽量让几家工具读同一份 Skill，把实体放在共享目录，再为需要专属入口的工具建链接，免得复制三份以后，各自长成不同的性格。建立链接前先检查已有文件，别直接覆盖。
内容仍需审查。Claude Code 特有字段、写死的工具名，换到 Codex、Pi 后未必适用。任务说明、参考资料和普通脚本相对容易共享，宿主特有行为则要适配。
我之前的《Claude Code 基础分享》整理过 Rules、Skill、MCP 和共享目录的使用经验，相关配置还要结合当前版本文档核对。
偶尔检查一下，到底读的是哪份
我自己的 Sessions Viewer 汇总了 Skill 和 MCP 配置，平时用它检查实体目录和引用关系。同一份 Skill 的多个入口，与内容已经分叉的几份副本，需要分别处理。
同名 MCP 也可能配了不同参数。图中的 token 是缓存估算的定义开销，配置状态不能证明当前会话已经连通，仍要实际调用核对。不用管理界面，直接检查文件也可以。
九、装完以后，怎么验收？
拿同一组图片做个对照：先直接给要求，再带上 Skill，看看有没有少漏步骤、结果是否更容易核对。还可以试试缺少密钥、文件已经存在、上传失败时，它能不能停在该停的位置。
正常流程能跑通以后，继续试误触发、工具断连，以及用户临时改需求。这些情况更容易把设计上的疏漏翻出来。
下载来的 Skill 要连同引用文件和脚本一起看，MCP 服务也要核对来源与权限。写着“安全”的说明书，不能替实际执行的代码担保。先在测试目录、测试账号或合适的隔离环境中试，尤其留意删除、覆盖、上传和读取凭据的动作。
我做过一个收尾报告 Skill，叫 ask-report-html。因为任务一长，我总要重新翻记录，找出改了什么、测了什么、还有什么没验，翻得比查 bug 还像在查案。
具体设计和安装方式在这篇关于 ask-report-html 的实战回答。
Skill 负责内容与生成方法，规则说明何时调用，Workflow 要求到了任务交付阶段再整理。换会话继续时，保存的任务记录也能帮助恢复进度。
没有浏览器验收，就写“未运行”。HTML 文件生成成功，证明不了业务通过；报告再漂亮，也不能替没跑过的检查签字。和上传成功后才清理文件一样，完成标准必须能核对。
我会继续看重复解释和返工是否减少，失败时有没有留下足够信息，结果是否更容易验收。装了多少工具，替代不了这些判断。
十、今天就想试，可以从哪儿开始？
挑一个手头的小任务，先用现有 Agent 做。它在哪儿让你反复解释、哪一步总要人工接手，再补相应配置。
总要纠正同一条要求，就写规则；一类任务反复交代步骤，就整理 Skill。现有能力拿不到外部信息时，再接 MCP。长任务出现确认条件和失败分支，就把 Workflow 写清楚。隔一阵子回来还要重讲的已确认信息，再考虑保存到记忆。
具体可以这么做：
- 写一份简短的项目指令，只放真正经常用的约定。
- 把一个已经做过、下次还会做的任务整理成 Skill，写清适用范围和交付物。
- 如果需要外部接入，只接一个服务，核对实际返回结果。
- 给会花钱、覆盖文件或影响外部系统的步骤加确认，并测试失败时是否会停下。
- 下次换新会话，检查关键约定能否找到，发现过期信息就更新。
先做到前两步也很好。等任务需要时再加，省得本来想让 AI 帮忙写代码，最后给自己新添一份配置运维的工作。
总结
放回一次开发任务里，Rules 留下长期约定，Skill 保存这类活的做法，MCP 接入工具和数据，Workflow 交代顺序与确认条件，Memory 帮下次会话找回已确认的信息。它们可以配合，也可以只用其中一两样。
用得好不好，还是看任务有没有少漏步骤、结果能不能验收，出错时能否如实停下。先拿一个真实小任务试，比对着配置列表挨个打勾更容易看出效果。
两个 Skill 附件下载
- skill-designer 下载：帮你梳理 Skill 的适用场景、内容结构和检查点。
- 微缩动画 Skill 下载：文中新街口案例使用的技能，先看样帧，确认后再出片。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
I used it between Claude Code, Codex and Pi-Bolt, often: I just wanted AI to help write a page, and the configuration had to come up with a ranking word, like a test.
Taking the "Page Realization Based on Figma Design", MCP can access the draft, Skyll specifies how the picture will be handled, Workflow arranges the order of confirmation and execution, Jules limits which documents to remain unmoveable, and Memory leaves the next engagement that is needed.
They can be used together, with a general division of labour as follows:
The names of the products will overlap. A Skill can contain Workflow, and a client may also place a rule file in the “Memory” menu. What it does is judged by the timing, the maintenance and the conduct of the execution, and by the fact that the name of the menu is easy to recognize.
The following Agent is an AI assistant who can read and write files and call tools; the host is running its client.
I. Starting with the processing of the drawing picture
I often export asset from Figma, compress PNG, upload Tencent cloud COS, copy the address and put it back in code. The page hasn't started. I've been a courier for the pictures.
When you meet a page, you have to repeat it several times. Then I made it into image-auto-upload Skill, with two scripts, one compressed PNG, one upload COS, and then asset processing into the page development.
I was reading the draft with Figma MCP. Images of how to compress, where to pass, and how to return the link are described by Skill; confirmation of asset list before uploading is the task Workflow. Keys may not appear in codes and logs, but are suitable for long-term rules.
The actual call for compressed services and COS SDK is a script. AI to determine what steps the current mission needs to take to run through the host tool.
If there's a script to do the job, and Agent can run the command, you can use it first. There is no need for an MCP layer for each script to complete the configuration. The picture has not yet been uploaded. One more service, which is not very economical.
The old text uses the method of uploading successful clean-up maps. Now design a similar process for beginner, and I will keep the source file, verify the remote results and then clean it up as agreed. Even if the key is placed in a non-submitted environmental file, the key will be secured from logs, screenshots and tool outputs.
II. RULES: Write down repeated corrections
If you start programming with AI, I'll write a few rules first.
“Don't go along with the other issues,” “The original changes to the workspace, not the “test without running”, you do,” which may have been said many times. I don't usually expect a surprise to be able to rebuild half a warehouse.
Such requirements are suitable for inclusion in long-term instruction documents. Claude Code is common in CLUDE.md, Codex and Pi, and AGENTS.md. The client determines how the file is found and where it is loaded.
These can be written first:
# Project collaboration engagement
- Check the workspace before amending it and keep changes that are not relevant to the current mandate.
- Replicate existing project components and dependencies; explain the reasons before adding additional reliance.
- The failure of the interface must keep the wrong information and do not impersonate success with empty data.
- Runs the relevant check after the change and records the pass, failure and non-run items.
- There is no explicit authorization to hold, submit, push codes.
Each article corresponds to a specific act. Problems also make it easier to check which ones were not complied with. I'd rather write this, but I'd rather give AI the title of the world's top engineer first.
Don't let all the rules follow every mission.
The preferences that apply to all individual items are at the user level; team-shared build commands and directories are agreed to at the project level; only the specifications of the interface catalogue are limited to that extent.
Claude Code, for example, can put the path code in.claude/rules/api.md:
---
paths:
- "src/api/**/*.ts"
---
When you change the interface, use the existing request envelope.
Do not repeat the server address in the page component.
This is the specific syntax of Claude Code, which, instead of other tools, looks for other ways of supporting it.
Split multiple files and check how they are loaded. If all the norms are directly introduced in the main document, the model may be read in large quantities at a time. In order to reduce the amount of context that is not relevant, the path range supported by the client or the needs-based read-out mechanism cannot be moved to a single file.
The rules have to be cleaned up long enough. The item is changed from npm to pnpm, and the old command should be updated; the temporary barrier is limited and will be checked for retention after completion.
The “prohibit” in Markdown is only an order for actions that are indeed impermissible. permission, sandboxes, command checks and CI can be intercepted at the appropriate link, where they can be stopped, which requires verification. The three times in the document that the word “prohibited” is written, there is no automatic additional door-bargaining.
iii. Skill: This kind of work. Next time do this.
You can't get how Skill wrote it, I'll start with skill-designer. It helps streamline the division of labour between scenes, entry notes and reference materials. In cc, pi, codex and others aggent, the following can be requested:
/skill-designer
Help me sort out the page Question as a Skyll.
Confirms the application of the scene and the input, distinguishing read-only analysis from changes in the code.
Lists the treatments when the connection fails and the check points before delivery.
Indicate what is placed at the portal, which is broken into reference documents and which is suitable for scripting.
This is the Skill I used myself to install a location that the client can find. The terminology of the different tools is different, and there is a comparison later.
Once the draft is out, it will still have to be tried with real task. Let's see if we're wrong to trigger or miss steps and if we can stop when we fail. The catalogue looks well and works well, and has to be examined separately.
What's worth doing, Skill?
The same set of requirements began to be pasted repeatedly, or a return to work left a reusable solution, which I would consider as Skill. Next time, at least you don't have to go to archaeology in the chat record.
Photographs can be uploaded, code reviewed, and certain types of interface questions can be sorted. The task entrance and completion criteria are clearly defined before deciding which documents are required.
Common catalogues are:
page-check/
├── SKILL.md
├── references/
│ └── checklist.md
└── scripts/
└── check-output.mjs
SKILL.md indicates when and by what steps. Detailed check items can be placed in references to stabilize duplicate calculations and validate to the script. Skill can be done without script, and there is no need to build an empty folder to make the catalogue look professional.
First, Skill, it's good to write down the trigger and the delivery.
This is a simple example of a page-screening display that is not yet a validated generic barrier tool:
---
name: page-check
Other Organiser Applicable to user request to repeat page failure; not used when reading code only.
---
1. Confirm the anomalies observed on the target page, operating environment and users.
Read the project start-up and validation instructions and then open the actual page.
3. Record the recurrence steps and check relevant DOM, console and network requests.
4. Reasons for positioning against codes. In the absence of a change in the delegation of authority, only analytical conclusions are given.
5. Modified authorization to repair and rescale.
Reasons for delivery, scope of changes, validation results and unverified items.
When the browser connection fails, indicate the blockage; do not write code extrapolation as a live recurrence.
See details/checklist.md for further reading when entering the page.
It explains what to do and how to deal with the failure of the connection and the absence of a change in the authorization.
It would be preferable to include both cases of application and exclusion. The “optimal development experience” is too broad to determine when the model should be used. On a large scale, an order may be delayed for a full medical examination.
"Step-by-step disclosure" is where you read it.
In the Pi-Bolt that I use, the model first sees the name, description, and path of Skill, deciding to use it and then read the text. The text can guide it by reading the corresponding details at a certain step.
Codex and Claude Code also follow the lines of load on demand, which differ from the field, Discovery Rules and context budget.
I used to do micro-animation, Skill, to separate the style, reference image processing, motion check, audio and export. The scene was designed without reading the entire coding instructions, and then releasing it in the official film.
In the article about Skill and Prompt, New Street micro-animation, I wrote, "It's not enough to open a file, and the entrance has to be clear about the timing."
As for "Skill is not Prompt," the text is, of course, the instructions for the model. I keep descriptions, information, templates and scripts for such missions and when to read them. Change the file name, the model won't grow a brain.
MCP: When external access is required, it will be matched
MCP, in its full name, Model Context Protocol, can be understood as a set of protocols for client connection tools and data services. Browser services provide access to pages and requests, document services provide search and reading, and design tool services provide draft information.
Once the connection has been successful, the method, target environment, login status and permission still need to be confirmed. The browser tool opens the page, and there is a set of practices to be followed for replicating questions, checking requests or cross-codes.
When is it worth pretending?
It is hoped that AI will be able to check for mature MCP access when you check the running page, read the draft, and query the document service.
If you only read local files, run tests, process several images, and the client is already able to read the files and run the commands, you will first use the ready capability. CLI, SDK, HTTP API may meet the demand.
This desktop application, which I developed with Tauri, has an easy-to-peer pit: the development must start with the dev-mcp feature when checking the actual interface through the MCP bridge. Ordinary development orders can also open applications, but bridge capacity is not included. Look at the window, it's not working. It's a good place to doubt life.
This start-up requirement is suitable for inclusion in the project, Skill, for subsequent validation to read the instructions.
Let's start with a read-only service.
For example, Codex can configure the official document MCP in ~/.codex/config.toml:
[mcp_servers.openaiDeveloperDocs]
url = "https://developers.openai.com/mcp"
Configure to check the state of connection, and then allow it to check a document and check the return content and source. This official document-only service does not get your database or cloud account permission.
Claude Code, Pi have different formats and positions, and do not stick the original TOML to JSON.
Some MCPs are linked by remote web sites, and some require local start-up of service programs, often referred to as stdio through standard input. The latter may rely on environments such as Node, Python and also process service-end installation and document access authorizations.
When a name appears in the list, you continue to check whether the client has read the configuration, whether the service is connected, whether the required tool can be called and whether the call result comes from the correct environment. The ability to read public pages does not mean access to the back of the login; it does not mean access to the testing library or to write permission to the production bank.
In particular, the browser will see which window and session it links to. I'm used to reusing the Visible Browser to facilitate environmental confirmation. Whether or not to separate the browser or support the connection depends on the service.
Chome DevTools MCP: Do not crowd me out of the web page.
I've seen this before, a new AI session, and Chome plays the connection authorization; let AI open the web page, and it cuts the tab I'm looking at. Instead of an independent browser, the login is separated. I wanted him to check a page, so I had to get him a seat.
The complete configuration and script at the time were: chrome-devtools-mcp, authorization to bomb and grab tabs? I'm so step by step.
That was the practice of MacOS, Codex CLI and the official package 1.12.0, respectively, which addressed these issues:
--autoConnect selects the connection target; custom-written quiet starter relays calls, reuses connections and binds backstage operations. The quiet here is an extra fit code, and none of the official "quiet switches" cover these things.
Creates a page in the background: true, selects a task page without stealing the front desk with the bringToFront: false. In addition to writing rules to remind me, I had fixed these two parameters in a suitable script to reduce leakage.
User permission is still required for the first connection. Authorization may also be requested once the Chrome, connection or resident process is restarted. Client approval of tool call, and Crome allowing remote debugging, are two layers of permission, which cannot replace the latter.
When matched, you need to open another AI session test: whether to reuse the original process, whether to open the backstage and whether the original tab is still selected. The current session was successful once, but the inter-session session has not yet been validated. On a daily basis, it is also important not to stop the process of sharing each mission.
In this configuration, the MCP provides browser operations, matching the use of the Skill introduction, Rules restrains the user from moving the original tab, and adapts the code to connect and fix parameters. The starter relies on the internal module of a specific package version and upgrades need to be verified together and cannot be deployed directly to all systems.
The services are much better equipped and context costs are considered. Tool definitions, call results and logs may all take up space; some client support tools are found or delayed loading, and the actual cost cannot be calculated by MCP numbers alone. The services that will be used to keep the current task are difficult to locate.
V. WORKFlow: Explanatory order and confirmation conditions
Some produce the Markdown step of duplicate call, known as Workflow, and some provide the workflow engine that preserves the status of the executive branch. Organizes a set of steps in Skill, also designing workflow.
So, Skill can install workflow: The former is a capacity kit, the latter describes the implementation process. When it comes to specific products, it also depends on what it calls Workflow.
This is illustrated by the fact that the actual work can be found in an animated response to the new street that was quoted earlier. The order of my production is:
Read reference diagrams, confirm style requirements
Generate four actual frames of the same animation
Wait until I confirm the current version
Run a check and re-render the whole film.
Reconfirm if there's a change in the picture or the plot.
"Let me see first" is too general. A concept or a frame? Do you think you're gonna approve the whole movie? With the change of shot, is the old authorization still valid? If a compliment were to be accompanied by a production authorization for the entire film, it would be too much.
I later completed the check of approval records with the version and checked the approval for the current version before the film was released. Markdown indicates the order to be followed and the code check rejects an ineligible request at the designated executive entry point.
Development missions can also be organized as follows:
Revert the problem and record the evidence.
Recognition of causes and scope of modifications
Restoration and completion of relevant tests
I'll take the original step.
Check diff, deliver validation results
The identification of steps that would broaden coverage, incur costs, cover data or affect external systems. I can't stand the fact that the text has been stopped five times; the part that can be completed consecutively allows it to be done consecutively.
Where there is a need for unmanned, fixed branches or reliable retrys, the parts suitable for determination of execution can be handed over to scripts, CI or workflow engines, with AI handling the steps to be judged. The list of natural languages does not necessarily cover these requirements alone.
MEMORY: Don't let me start over.
After a few days, I still hope AI will find an agreement that has been confirmed: for the time being, a project cannot be upgraded, an environment can only be started in a specific way, or an analysis can be made before the code can be modified.
This information can be saved to memory. Stable preferences, recognized decision-making and remaining effective environmental constraints are appropriate for the long term. It would be more appropriate to record the day or the progress of the mission if one of the documents that was read today and one of the tests was not running. Everything is permanently stored, and finally it has to be flipped in it.
We'll have to try and find it.
The document is called MEMORY.md, which does not mean that the model is read in its entirety, much less that the content is correct. Depends on which parts the client loads, whether they are retrieved on demand, when the new record becomes effective and how old information is modified or deleted.
In Pi-Bolt, I used a pi-memory extension to keep long-term records, daily logs and to-dos separate, providing tools for reading, writing, searching and forgetting. Search also relies on an optional qmd search tool; semantic search of similar content by meaning requires corresponding indexes and models. This is my extended configuration, and the installation of Pi does not automatically attach this whole set of memories.
Claude Code has its own auto memory, and the manually maintained CLUDE.md is not written in the same way and for the same purpose. Turning to Codex, check the corresponding version and configuration support. Both sides can read Markdown, and it's a lot less than automatic syncment.
In a client that already has memory capabilities, this can be done:
Remember: In this project, there is no clear mandate not to temporarily save or submit codes.
Please save it as a long-term agreement to the project and tell me where it was written.
Do not save the test account number, key or current round temporary log into long-term memory.
Then a new session, asking it what it agreed to, where it read it, then checking the saved documents. One cannot be reassured by just one word of “remember” that my memory is too weak to be so assured.
Leave the scope and cause, and remember to clean up.
A dependency limit can be recorded as follows:
Scope: Payment dependency on current project.
Conclusion: the current version is maintained for the time being.
Reason: The compatibility of the new version has not yet been validated.
Re-evaluate: confirm whether or not to upgrade after compatibility verification has been completed.
This is a sign that the actual reason and the approver have to come from a genuine discussion and that the model cannot be used to justify itself. When a decision or environment changes, the old record is updated and deleted, or the same error is revived with memory.
Rules and memories also intersect. “Unauthorized non-refoulement” may be preserved as a personal preference, or may be written into rules that require long-term enforcement. Memory emphasizes preservation and recall, and the rules emphasize what it is expected to comply with.
Re-verification is still required to operate the production environment. Memoory can tell the last time, but today's environment may have changed.
VII. Orders, scripts, Hooks, what does it matter?
The same slash entrance, with different tools, may also have different mechanisms behind it.
Pi's Prompt Template can turn common tips into /review. Just a little bit less of a review reminder to use a template; reference and complete processing are required before considering Skill.
Claude Code has merged custom Comands into the Skills system, and the old directory is still compatible. Don't look at it all as a single mechanism.
Scripts perform fixed logic. Document summary, data conversion, format validation, this sort of work is clear, and I'm willing to get Agent to call on a ready script to save every rerun.
Hooks executes assigned commands or checks at the time of an incident and is suitable for alerting and intercepting at specified points. Some of the access points that Hook can cover need to be validated in conjunction with the host. The writing of "tool call pre-inspection" in Skill does not automatically generate a Hook; neither does it make it clear that all operations are safely isolated.
The plugins can be packed, connected and other resources for easy distribution. The expansion also allows for registration tools or response to events and increases host capacity. When these capabilities are required for the task at hand, it is not necessary to make the editor a tool gallery in advance.
Eight. Pi, Codex, Claude Code, how do we land?
Common local entry points are as follows. ~ is the user home directory; the configuration in the project directory is valid for the corresponding item. The loading behaviour is still affected by the version and settings.
My machine started with pi-bolt, modified Skill and refreshed with /reload. Upstream Pi, Pi-Bolt, different versions, and installed extensions will vary, and I do not necessarily use all of Pi's default configurations.
Claude Code is also currently in a position to support reading AGENTS.md, influenced by the version, whether the project has a CLAUDE.md and related settings. When multi-tool sharing norms are used, check the actual loading source; if compatible, you can introduce shared content with @AGENTS.md in CLUDE.md.
Share content as much as possible, and access differences are treated separately
I will try to get a few tools to read the same Skill, put the entity in a shared directory and link the tools that require an exclusive entrance, so that they do not have a different character after three copies. Check existing files before creating links, do not directly overwrite them.
The content still needs to be reviewed. Claude Code has a field, a name for a dead tool, which may not be applicable when changed to Codex, Pi. Mission statements, reference materials and ordinary scripts are relatively easy to share, while host-specific behaviour is appropriate.
My previous Claude Code Basic Sharing collated experience with the use of Rules, Skill, MCP and shared directories, and the configuration was to be checked with the current version of the document.
Check once in a while. Which one is it?
My own Ssessions Viewer aggregates the Skill and MCP configurations that are used to check entity directories and references. Several copies of the same Skill ' s multiple entry points, separated from the content, need to be processed separately.
The same name MCP may have different parameters. Token is the definitional cost of cache estimates in the figure, and the configuration does not prove that the current session is connected and still requires a physical call check. It is also possible to check files directly without a management interface.
Nine. When you're finished, how about acceptance check?
A comparison of the same set of images: direct requests, then Skill, to see if there are any missing steps and whether the results are easier to check. It is also possible to try to stop at the stop if the key is missing, the file already exists and the upload failed.
Once normal processes can run, the test error triggers continue, the tool is disconnected and the user ' s need for temporary modifications continues. These conditions make it easier to turn the gaps in design.
The download of Skill is to be read with the reference file and script, and MCP services are to check the source with permission. A statement of “security” cannot be secured by the code actually executed. Test first in a test catalogue, test account number or suitable quarantine environment, paying particular attention to deletion, overlay, upload and readable actions.
I did a final report, Skill, called ask-report-html. Because of the long mission, I'm going to have to go over the record, find out what's changed, what's measured, what's untested, and turn it over like a case.
Specifically designed and installed in this field response on ask-report-html.
Skill is responsible for content and method of generation, rules on when to call, and Workflow requires to be sorted at the task delivery stage. The task record kept can also help restore progress when the exchange continues.
Without browser acceptance check, write " Not running." HTML files have been generated successfully and cannot prove business approval; reports are so beautiful that they cannot be signed for unrun checks. As with the clean-up of documents after successful uploading, completion criteria must be capable of checking.
I will continue to see whether repeated explanations and returns are reduced, whether sufficient information is left in case of failure, and whether the result is easier for acceptance check. How many tools are installed to replace these judgements.
Where can we start today?
Pick a small task at hand to do with the existing Agent. It's where you're going to explain it over and over again, and it's going to have to take over manually and reset it.
The same requirement has to be corrected, the rule written; the first type of task is repeated, the Skill is organized. If external information is not available from existing capacity, then pick up MCP. When a long mission comes up with the confirmation of conditions and the failure branch, write it out. The confirmed information that comes back for a while to repeat is considered for memory.
This can be done specifically:
- Write a short project directive, with only truly frequent agreements.
- Collapse a task that has already been done and will be done next time into Skill, with a clear scope of application and delivery.
- If external access is required, only one service is available to verify the actual return results.
- Confirmation of steps that would cost, overwhelm documents or affect the external system, and testing whether to stop in case of failure.
- Next time a new session will be held to check whether key agreements can be found and to update the information when it is found to be overdue.
It's good to do the first two steps. When the mission needs to be added, save AI from wanting to help with the code, and finally add a new job to the configuration.
Summary
Back in a development mission, Rubles left a long-term engagement, Skill kept this kind of activity, MCP access tools and data, Workflow spokes and confirmation terms, and Memory helped the next session to retrieve confirmed information. They can cooperate or use only one or two of them.
Whether it's useful or not, if the mission is missing steps, if the result is acceptance check, if it's wrong or not. Take a real little task first, which is easier to see than ticking the configuration list.
Two Skill Attachment Downloads
- Skill-designer download: help you comb the applicable scene, content structure and check points of Skill.
Microfiche Animation Skill Download: Skills used in the new street case, sample frame, confirmation and later film.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.