The Complete
Codex Guide:
From Zero to
Commanding Your Agent Codex 保姆级完整教学：从入门到指挥你的超级 Agent

The way you use ChatGPT, Gemini or Claude in the browser is: you ask → it answers → the rest is still on you. That's not AI doing your work. That's AI giving you ideas.你现在用 ChatGPT、Gemini、Claude 网页版的方式是：你问它 → 它给答案 → 剩下的还是你自己动手。这不是 AI 在帮你工作，这是 AI 在帮你出主意。

Codex runs on your own machine: it reads and writes your local files, connects to Gmail, Calendar, Canva and scheduling tools, runs several tasks at once, doesn't need you watching over it — and needs no programming background. You only need to be able to say what you want.Codex 装在你自己的电脑上：它能直接读写你的本地文件，连上你的 Gmail、日历、Canva、排程工具，同时跑多个任务，不需要你在旁边盯着，也不需要任何程序背景——你只需要会说你想要什么。

1. Find the official site找到官网 Search Codex on Google and open the official site. It detects your OS and offers the matching download. Google 直接搜 Codex，打开官网，它会自动识别你的系统版本，给出对应的下载入口。
2. Install the desktop app下载桌面端 App Open it and you'll be offered two ways to sign in. 装完打开，会提示两种登录方式。
3. Route A: sign in with ChatGPT (try this first)方式一：ChatGPT 账号登录（推荐先试这个） Least friction, and your quota simply follows your subscription. 最省事，额度直接跟着你的订阅走。
4. Route B: API key方式二：API Key Use an official OpenAI key, or a third-party relay. The real value of this route: beyond OpenAI's own models, you can drive this super Agent with other LLMs. 可以用 OpenAI 官方 API key，也可以用第三方中转 API——这条路的价值是：除了 ChatGPT 官方模型，你还能调用别的大模型来驱动这个超级 Agent。

• Far left — menus and chat entries. Bottom-left is settings, where you check your Codex quota.最左侧：功能菜单与会话入口；左下角是设置，在里面可以查看 Codex 额度。
• Middle — the largest area. Where you talk to the Agent and hand out work.中间：最大的区域，和 Agent 对话、派活的主战场。
• Right — the preview pane.右侧：预览界面。

These are the controls on the chat box you have to know:对话框上你必须认识这几个开关：

Open Plugins and note the toggle at the top left: Plugins vs Skill. One line each:点开外挂程序，左上角有一个开关：Plugins（外挂） 和 Skill（技能）。两者的区别一句话说清：

• Skill = your SOP. Turn a complex, standardised workflow into a Skill, so every run lands above the quality bar you accept.Skill = 你的 SOP。把一整套复杂的、标准化的工作流程定义成一个 Skill，确保每一次产出都在你能接受的质量线以上。
• Plugin = wiring to outside tools. It doesn't define the steps; it plugs external software into Codex so the Agent can actually operate it.Plugin = 外部工具的接线。它不定义工作步骤，而是把外部软件接进 Codex，让它能真的动手操作那些软件。

The plugins shipped out of the box fall into a few groups:官方已经提供的外挂大致分几类：

Hands-on demo: install a plugin, click "Try in chat", and one sentence is enough:实操演示：点「安装」装好插件后，点「在聊天中试用」，输入一句话就够了：

It prompts you to connect Gmail, then searches that date range, surfaces the real sponsorship offers on the first pass, and summarises them. From there you can have it draft the replies, or click straight through to the original mail.它会提示你绑定 Gmail 账号，绑定完就开始检索指定日期区间的邮件，第一轮就抓出明确的商单邀约，并整理成汇总。你可以顺手让它直接草拟回复，也可以点邮件链接跳回原文自己处理。

Top left there's Automations, similar to routine / schedule in Claude Cowork. Click New automation and describe it in plain language: what time, which project, which worktree (local or chat).左上角还有一个「自动化」功能，类似 Claude Cowork 里的 routine / schedule。点右上角新增自动化，用对话式提示词描述就行：几点跑、在哪个专案、在哪个工作树（本机还是聊天）。

Type / in the chat box and a whole row of options opens up:在对话框里输入 /，会弹出一整排功能选择：

• Call a specific Skill调用某一个特定的 Skill
• Enable and use MCP features启用 / 使用 MCP 相关功能
• Set the Agent's tone — friendly or matter-of-fact设定 Agent 回复的性格（友善 or 务实）
• Manage chat memory管理聊天记忆

The critical first step is creating your own local project. Pick an existing one in the chat box, or click "New project" — name it something like My AI Agent.正式开始用 Codex，最关键的第一步是创建你自己的本地专案（project）。在对话框里选择已有专案，或点「新增专案」，例如命名为 My AI Agent。

Once the project exists, the key artefact is AGENTS.md. It sits in the root of the project folder, Codex reads it before every job, and then acts by its rules. It directly determines your output quality.建好 project 之后，最关键的就是写好 AGENTS.md。这份文件放在项目文件夹的根目录，Codex 每次开工前都会先读它，然后按里面的规则行动。它直接决定你的产出质量。

It doesn't need to be long. Three things are enough:不用写长，说清三件事就够：

1. Who you are你是谁 e.g. you're a YouTuber — the channel name, your main content series, the audience profile. 比如：你是一个 YouTuber，你的频道叫什么、主要内容系列有哪些、受众画像是什么样的人。
2. Output language and format输出语言与格式要求 e.g. all output in Simplified Chinese; tone pragmatic, objective and careful; direct, no waffle. 比如：所有输出必须是简体中文；文案风格务实、客观、谨慎，直接干脆，不要弯弯绕绕。
3. Working principles工作原则 e.g. list a plan before executing; when done, report which files were created and where. 比如：每次执行任务前先列计划；完成后告诉我生成了哪些文件、放在哪里。

Easiest way to create it: throw all the prompts and documents you've accumulated at Codex and let it write the file for you:怎么创建？最省事的办法是把你过去积累的提示词、文档一股脑丢给 Codex，让它替你写：

What a working AGENTS.md actually contains:一份实际可用的 AGENTS.md 结构：

• Project positioning — e.g. content creation + AI agent development专案定位：例如内容创作 + AI 智能体开发
• Root directory conventions — AGENTS.md / memory / content / workflow / Sandbox根目录约定：AGENTS.md / memory / content / 工作流 / Sandbox
• Naming conventions — how files get named during content work命名规范：内容创作时文件怎么命名
• First step on entering the project — if the workspace is messy or a new subfolder has no convention, tidy the structure and fill the gaps first, then start work进入项目第一步：如果发现工作空间结构混乱、或有新子目录没有规范，先整理结构、补齐文件，再开始工作
• Content workflow — the standard route for this kind of job内容创作工作流：这类活标准怎么走
• Related links — entry points to the material you use often相关链接：常用资料的入口

Top right there's a Switch terminal button that pulls up your local terminal. If Claude Code is already installed on your machine, one short command brings it up:右上角有个切换终端的按钮，点开就把本地 Terminal 拉起来了。如果你本机已经装了 Claude Code，在这里输入一个命令就能把它叫出来：

Now one window has Codex on top and Claude Code below, both working for you. Ask "who are you" and Claude Code answers from the persona you wrote — "I'm your AI assistant, your AI double" — and lists what it can do.于是同一个界面里，上面是 Codex、下面是 Claude Code，两个 Agent 同时为你工作。随手问一句"你是谁"，Claude Code 会照着你写好的设定回答——比如"我是你的 AI 助理、你的 AI 分身"，并列出它会做哪些工作。

Your scarcest asset when working with an Agent is context. When a thread runs long, or when you want two tasks in parallel but don't want them in the same chat (messy context, conflicting work) — fork it.和 Agent 打交道，最关键的资产是上下文。当你要做一段很长的多轮对话，或者要并行推进两个任务、又不想挤在同一个对话框里（会让上下文混乱、任务互相冲突）——用分叉。

• Select the chat on the left → choose Fork在左边选中对应的对话框 → 选择分叉
• Generally, fork to local一般建议直接分叉到本机
• Result: two identical chats. Everything before the fork is the same; everything after diverges.结果：出现两个一模一样的对话框，分叉点之前的历史完全相同，之后各走各的

Go to Plugins → switch to Skill in the top left. Codex gives you a genuinely good visual list; click to install. What ships by default:点「外挂程序」→ 左上角切到「Skill」，Codex 给了一个相当不错的可视化列表，点击即可安装。官方默认提供的一批：

Skills you previously had the Agent install from the web (Lark integrations, various CLI tools) are listed here too.另外，你过去通过对话让 Agent 从网上装下来的 Skill（比如 Lark 相关、各种 CLI 工具），也都会陈列在这里。

The simplest route is to just tell Codex what you want and let Skill Creator write it:最简单的方式就是直接告诉 Codex 你要什么，让 Skill Creator 去写：

It writes the Skill as a Markdown file into your project folder. To use it later, type /IG文案 in the chat box and it knows the format. Or just say:它会自动生成 Skill 的 Markdown 文档，存进你的 project 文件夹。之后要用，在对话框打 /IG文案 它就知道按这个格式来。或者直接说：

A real output — headline: "Codex, the super Agent: you're still chatting with AI while the pros already have Agents working." The body:一次真实产出——标题是「Codex 超级 Agent：你还在用 AI 聊天，高手已经开始用 Agent 工作」，正文骨架：

• An answer is not a result. Most people use AI to ask questions, so what they get is answers, not delivery.答案不等于结果：大多数人用 AI 只是在问问题，所以得到的是答案，不是交付。
• The real gap is letting AI into your workflow. Codex is an Agent with hands: it reads projects, edits code, runs tests, organises results.真正的差距是让 AI 进入工作流：Codex 是会动手的 Agent，能读项目、改程序、跑测试、整理结果。
• ChatGPT gives advice; Codex hands in the work.ChatGPT 给建议，Codex 帮你交付。
• Four instructions that make Codex work: state the task · supply the background · define the standard · demand verification.用好 Codex 的 4 个指令：明确任务 · 补足背景 · 定义标准 · 要求验证。
• You're not asking an AI. You're managing an Agent.你不是在问 AI，你是在管理 Agent。

Once the copy is right, have it call the latest image model for the matching visual (the example used ChatGPT image 2) — a finished image lands in a few minutes.文案满意后，直接让它调用最新的生图模型出配图（示例用的是 ChatGPT image 2），几分钟拿到成品图。

Want to take the image you just made into Canva? Click ＋ → Plugins → Canva → authorise your account. The image goes straight into your Canva workspace, where every element is editable — fonts, colours, whatever.想把刚生成的图导到 Canva 继续修？点加号 → 外挂程序 → 选 Canva → 授权你的账号。图会直接进 Canva 后台，每个元素都是可编辑状态——字体、颜色随便改。

The same works for Google Slides and Figma. And an obvious extension: build an Instagram plugin so the finished visual publishes itself to your IG.同理也可以导入 Google Slides、Figma。还有一个延伸方向：未来做一个 Instagram plugin，配图生成后自动发布到你的 IG 后台。

Beyond images there are Skills for landing pages and visual decks. One sentence:除了图，还有生成落地页和可视化简报的 Skill。提示词可以是这样一句：

Watch the order it works in: it lays out the structure of the whole deck first (13 blocks in the example, positioned as "an AI-employee manual for non-engineers"), then asks you to pick one of 5 visual styles, and only then generates.它的执行顺序值得注意：先给出整份简报的内容结构（示例中是 13 个区块，重点定位为「给非工程师的 AI 员工操作指南」），再让你从 5 种视觉风格里选一个，然后才生成。

The finished report's skeleton:成品报告的骨架：

• Title: The complete Codex guide — turning AI from "answering questions" into "an employee who hands in the work"标题：Codex 保姆级教学——把 AI 从「回答问题」变成「会交作业的员工」
• What you can make with it: IG posts, Facebook ad copy, one-page sites, TikTok UGC scripts, weekly topic reports…能拿来做什么：IG 图文生产、Facebook 广告文案、一页式网站、TikTok UGC 脚本、每周选题报告……
• What Codex is, in three layers: an AI assistant that can use a computer → your content production workbench → a multi-Agent task consoleCodex 是什么，一句话三层理解：会用电脑的 AI 助理 → 你的内容生产工作台 → 多 Agent 任务控制台
• What a Project is: the box / cabinet / homework metaphorProject 是什么：工作盒子 / 柜子 / 作业的比喻
• What AGENTS.md is: the AI employee's onboarding handbook, with a sample creation promptAGENTS.md 是什么：AI 员工的入职手册，附创建提示词范例
• Three practical tricks: effort · steer (course-correct mid-task) · Fork Chat三个实操小技巧：effort（智慧等级）· steer（中途纠偏）· Fork Chat（分叉）
• A Codex vs Claude desktop comparison tableCodex vs Claude 桌面端对比表

Want more than images? You can generate video from inside Codex by calling a CLI.想要的不止是图，是视频？在 Codex 后台通过调用 CLI 的方式就能做。

Adding a new tool is a one-liner:接入一个新工具，就是一句话的事：

Codex installs it, tells you where the command entry point is and where the CLI landed, and writes a matching Skill document automatically (how to authorise, create a session, check progress, download results). That's the Agent's homework — you only send prompts.Codex 会帮你装好、告诉你命令入口在哪、CLI 装在了什么位置，并自动创建一个对应的 Skill 文档（授权方式、创建会话、查询进度、下载结果）——这些是 Agent 自己要学的东西，你本人只需要发提示词。

The actual production run:实际生成流程：

1. Give it one reference storyboard给一张参考分镜图 Any outfit-vlog storyboard found online will do. 在网上随手找一张穿搭 vlog 的分镜图即可。
2. Have it generate its own storyboard让它生成自己的分镜 storyboard Output: a 4×4 grid laying out the whole look — top, accessories, shoes and bag. 产出一张 4×4 十六宫格，把整套穿搭细节全部铺开：上装、配件、鞋包。
3. Hand the storyboard back, generate the video把分镜图交回去，生成视频 See the prompt below.见下方提示词。
4. Stitch into a finished cut拼接成片 The clips are joined into one complete piece; in the Lib TV project view, every clip's prompt and reference image is still retrievable. 各片段最终拼成一支完整作品；在 Lib TV 项目后台，每个片段对应的提示词和参考图都能找回。

Total human input: one storyboard image and one prompt.全程只输入了一张分镜图 + 一句提示词。

Codex ships plugins for build iOS App, build Web App and build macOS App. Take any dopamine-coloured music-player sketch you find online as the reference:Codex 的 plugin 里有 build iOS App / build Web App / build macOS App。这里拿网上一张多巴胺色彩风格的音乐播放器设计草图为例：

1. It plans before it builds它先梳理开发方案 No code straight away — it gives you the overall plan and asks you to confirm. 不直接写代码，先给出整体方案，让你确认。
2. Click "implement" to start点「实作方案」开始开发 When it finishes, you get a Preview screenshot. 等它跑完，会给你一张 Preview 预览图。
3. Ask for a live preview要求真机模拟预览 Reply: "let me preview and try it out directly." 回一句「直接让我预览和试用看看」。
4. It boots the iPhone simulator and deploys它调起 iPhone 模拟器并部署环境 Now you can tap and listen for real. 你就能在模拟器里真实点、真实听。

Result: it plays, it skips tracks, and the dopamine palette changes with every track, faithfully reproducing the reference mock. There's a library page for browsing by category too. That's a usable iOS App.成品实测：能播放、能切歌，每次切歌多巴胺配色跟着换，完整复现了参考视觉稿；还有 library 页面按分类浏览歌曲目录。已经是一个可用的 iOS App。

Every task above takes time to run. The point is that in Codex you can run tasks in parallel across different projects, and across several chats inside one project.前面每一个任务都要花时间跑。关键在于：在 Codex 后台，你可以在不同项目下、以及同一个项目下的多个对话框里，让 Agent 并行处理任务。

Then you go make coffee, answer messages, do something else. Come back in fifteen or twenty minutes and the tasks have finished, one after another.然后你就去泡杯咖啡、回消息、干别的。十几二十分钟回来，任务一个接一个完成。

1. Install Codex装好 Codex Search "Codex APP" on Google, download the build for your OS, sign in with your ChatGPT account. Google 搜 Codex APP，选你的系统版本下载，用 ChatGPT 账号登录。
2. Create your first project and write AGENTS.md建第一个 project + 写好 AGENTS.md Start accumulating your own context and prompts: who you are, what you do, your format requirements, your working habits. 开始积累属于你自己的上下文和专属提示词：告诉它你是谁、你做什么、输出格式要求、工作习惯。
3. Build your first Skill建第一个 Skill Think of the single most repetitive, most time-consuming thing in your week. Write down its requirements and format, and have Codex turn it into a Skill. 想一想你每周最重复、最花时间的那一件事，把要求和格式写清楚，让 Codex 帮你建成 Skill。

You don't have to do it all at once. But every step forward widens the gap between you and the people still copying and pasting by hand.不需要一次做完。但每往前一步，你和那些还在手动复制粘贴的人，差距就拉开一点。