引子 · 一个应用,三种工作体验Intro · One App, Three Ways to Work.
新版 ChatGPT 桌面应用把 Chat、Work 与 Codex 收到一起:Chat 对话,Work 交成品,Codex 进代码库。Codex 仍有独立视图与历史,并继续连接 CLI、IDE、cloud 与 GitHub。这一章先把合并后的入口和职责钉准,再给全书一把尺:不是「它会写多少代码」,而是「多少可验证的活能在你不守着时做完」。The new ChatGPT desktop app brings Chat, Work, and Codex together: Chat converses, Work produces deliverables, and Codex enters the codebase. Codex still keeps a separate view and history while connecting to the CLI, IDE, cloud, and GitHub. This chapter fixes the merged entrance and its boundaries, then gives the book one ruler: not how much code it can write, but how much verifiable work finishes without you watching.
你大概是这样用 Codex 的:在新版 ChatGPT 桌面应用切进 Codex 视图,贴一个报错、一段函数,看它改好,复制或接受 diff。偶尔它跑了命令、开了 PR,你觉得挺方便。一天下来,它对你仍只是一个更会写代码的对话 —— 比普通 Chat 多省下的,只是自己落地那几步。Here's probably how you use Codex: switch to the Codex view in the new ChatGPT desktop app, paste an error or function, then copy the answer or accept its diff. Now and then it runs a command or opens a PR. By day's end it is still merely a conversation that writes code better — saving only the few steps of applying ordinary Chat output yourself.这没错,但只是它最小的一层。Codex 已列入从 Free、Go 到组织档的 ChatGPT 计划,计划内用量与其它可用的 agentic 功能共池。1注 1Note 1OpenAI · Pricing —— ChatGPT Work 与 Codex 列入 Free、Go、Plus、Pro、Business、Edu、Enterprise 计划并共用用量;Free 定位 quick coding,Go 定位 lightweight coding。实际 Work 入口仍受 plan、workspace、地区与 rollout 影响。截至 2026-08-05。OpenAI · Pricing — ChatGPT Work and Codex are listed across Free, Go, Plus, Pro, Business, Edu, and Enterprise plans and share usage; Free is positioned for quick coding and Go for lightweight coding. Actual Work access still depends on plan, workspace, region, and rollout. As of 2026-08-05.更大的变化不是换了壳,而是你可以把一件复杂活拆给多个专门 agent,让主线程只守目标与决策;再把会产生独立改动的活隔离进 worktree 或 cloud,最后从 diff 与 PR 收回。Codex 的最大样子,不是一个更会写代码的模式,而是一套多 Agent 的异步编码工作流。Not wrong — but it is the smallest layer. Codex is now listed across ChatGPT plans from Free and Go through organization tiers, sharing included usage with other supported agentic features.1注 1Note 1OpenAI · Pricing —— ChatGPT Work 与 Codex 列入 Free、Go、Plus、Pro、Business、Edu、Enterprise 计划并共用用量;Free 定位 quick coding,Go 定位 lightweight coding。实际 Work 入口仍受 plan、workspace、地区与 rollout 影响。截至 2026-08-05。OpenAI · Pricing — ChatGPT Work and Codex are listed across Free, Go, Plus, Pro, Business, Edu, and Enterprise plans and share usage; Free is positioned for quick coding and Go for lightweight coding. Actual Work access still depends on plan, workspace, region, and rollout. As of 2026-08-05. The larger change is not the shell: one complex job can be divided among specialists while the main thread protects goals and decisions; independent changes can then be isolated in worktrees or cloud and collected through diffs and PRs. Codex at full size is not a better code mode but an async multi-agent coding workflow.
— I
一个应用,三种工作体验One App, Three Ways to Work.
2026 年 7 月 9 日之后,Chat、Work、Codex 进了同一个 macOS / Windows 应用,但不是三个完全相同的 tab。准确层级是:全局在 ChatGPT 与 Codex 之间切换;进入 ChatGPT 后再选 Chat 或 Work。Codex 保留独立视图与独立历史。2注 2Note 2OpenAI 帮助中心 · ChatGPT Work and Codex —— 新桌面应用的全局切换是 ChatGPT 与 Codex;在 ChatGPT 内再选 Chat 或 Work。Chat 适合快速对话,Work 适合研究与文档 / 表格 / 演示 / Site 等成品,Codex 专做软件开发。Codex 仍是独立视图与独立历史,不能作为普通体验在 web / mobile 直接选择;手机可从 Remote 接入受支持的桌面 Codex 对话。截至 2026-08-05。OpenAI Help · ChatGPT Work and Codex — the new desktop app switches globally between ChatGPT and Codex; inside ChatGPT, choose Chat or Work. Chat is for quick conversation, Work for research and finished documents / spreadsheets / presentations / Sites, and Codex for software development. Codex keeps a separate view and history and is not selectable as a normal web / mobile experience; mobile Remote can access supported desktop Codex chats. As of 2026-08-05.入口合并了,职责没有:After July 9, 2026, Chat, Work, and Codex live in one macOS / Windows app, but not as three identical tabs. The precise hierarchy switches globally between ChatGPT and Codex; inside ChatGPT, choose Chat or Work. Codex keeps its own view and history.2注 2Note 2OpenAI 帮助中心 · ChatGPT Work and Codex —— 新桌面应用的全局切换是 ChatGPT 与 Codex;在 ChatGPT 内再选 Chat 或 Work。Chat 适合快速对话,Work 适合研究与文档 / 表格 / 演示 / Site 等成品,Codex 专做软件开发。Codex 仍是独立视图与独立历史,不能作为普通体验在 web / mobile 直接选择;手机可从 Remote 接入受支持的桌面 Codex 对话。截至 2026-08-05。OpenAI Help · ChatGPT Work and Codex — the new desktop app switches globally between ChatGPT and Codex; inside ChatGPT, choose Chat or Work. Chat is for quick conversation, Work for research and finished documents / spreadsheets / presentations / Sites, and Codex for software development. Codex keeps a separate view and history and is not selectable as a normal web / mobile experience; mobile Remote can access supported desktop Codex chats. As of 2026-08-05. The entrances merged; the jobs did not:
研究、分析、文档、表格、演示、报告或 Site 成品Research, analysis, documents, spreadsheets, presentations, reports, or Sites
在真实代码库里跑命令、测试与 Git 流Commands, tests, and Git work in a real repository
Codex
写 / 修代码、跑命令和测试、审 diff、处理仓库与 PRWriting / fixing code, commands and tests, diffs, repositories, and PRs
把非软件成品都伪装成代码任务Disguising every non-software deliverable as a code task
Codex 视图只是控制台。往下还有 CLI 与 IDE 扩展这两个本地客户端、Local / Worktree / Cloud 三种执行位置、GitHub 这个团队集成层。Codex 不能作为普通体验在 web 或 mobile 直接选择;手机 Remote 只是在支持时接续桌面 Codex 对话。2注 2Note 2OpenAI 帮助中心 · ChatGPT Work and Codex —— 新桌面应用的全局切换是 ChatGPT 与 Codex;在 ChatGPT 内再选 Chat 或 Work。Chat 适合快速对话,Work 适合研究与文档 / 表格 / 演示 / Site 等成品,Codex 专做软件开发。Codex 仍是独立视图与独立历史,不能作为普通体验在 web / mobile 直接选择;手机可从 Remote 接入受支持的桌面 Codex 对话。截至 2026-08-05。OpenAI Help · ChatGPT Work and Codex — the new desktop app switches globally between ChatGPT and Codex; inside ChatGPT, choose Chat or Work. Chat is for quick conversation, Work for research and finished documents / spreadsheets / presentations / Sites, and Codex for software development. Codex keeps a separate view and history and is not selectable as a normal web / mobile experience; mobile Remote can access supported desktop Codex chats. As of 2026-08-05.Cloud 让活离开机器,subagents 让一件活在多个线程内分工。前者改变执行位置,后者改变任务结构 —— 这两件事不要混成一个「并行」按钮。3注 3Note 3OpenAI Codex 文档 · Codex cloud —— 一句话定位是「把活委派给隔离的云端环境里的 Codex」。每个云端任务在 OpenAI 托管的隔离容器里跑:检出仓库、带网跑 setup、然后 agent 阶段默认断网;跑完交回 diff,可开 PR 或追问。截至 2026-08-05。OpenAI Codex Docs · Codex cloud — delegate work to Codex in isolated cloud environments. Each task runs in an isolated OpenAI-managed container: it checks out the repo, runs setup with network access, then enters an agent phase that is offline by default; it returns a diff you can open as a PR or follow up on. As of 2026-08-05.6注 6Note 6OpenAI Codex 文档 · Subagents —— 当前 Codex 默认支持 subagent workflow:主 agent 把独立分块交给专门 agent,最后收摘要;CLI 用 /agent 查看与切换线程。适合读密集型并行,会比可比单 agent 任务消耗更多 token。eligible account 的 Work Ultra 可主动委派。截至 2026-08-05。OpenAI Codex Docs · Subagents — current Codex releases enable subagent workflows by default: the main agent assigns independent slices to specialists and collects summaries; /agent inspects and switches threads in the CLI. This fits read-heavy parallel work and consumes more tokens than a comparable single-agent task. Work Ultra can delegate proactively for eligible accounts. As of 2026-08-05.The Codex view is only the control room. Beneath it sit the CLI and IDE extension as local clients, Local / Worktree / Cloud as execution locations, and GitHub as the team integration. Codex is not selectable as an ordinary web or mobile experience; mobile Remote only continues supported desktop Codex chats.2注 2Note 2OpenAI 帮助中心 · ChatGPT Work and Codex —— 新桌面应用的全局切换是 ChatGPT 与 Codex;在 ChatGPT 内再选 Chat 或 Work。Chat 适合快速对话,Work 适合研究与文档 / 表格 / 演示 / Site 等成品,Codex 专做软件开发。Codex 仍是独立视图与独立历史,不能作为普通体验在 web / mobile 直接选择;手机可从 Remote 接入受支持的桌面 Codex 对话。截至 2026-08-05。OpenAI Help · ChatGPT Work and Codex — the new desktop app switches globally between ChatGPT and Codex; inside ChatGPT, choose Chat or Work. Chat is for quick conversation, Work for research and finished documents / spreadsheets / presentations / Sites, and Codex for software development. Codex keeps a separate view and history and is not selectable as a normal web / mobile experience; mobile Remote can access supported desktop Codex chats. As of 2026-08-05. Cloud moves work off the machine; subagents divide one job across threads. One changes execution location, the other changes task structure — do not collapse them into one 'parallel' button.3注 3Note 3OpenAI Codex 文档 · Codex cloud —— 一句话定位是「把活委派给隔离的云端环境里的 Codex」。每个云端任务在 OpenAI 托管的隔离容器里跑:检出仓库、带网跑 setup、然后 agent 阶段默认断网;跑完交回 diff,可开 PR 或追问。截至 2026-08-05。OpenAI Codex Docs · Codex cloud — delegate work to Codex in isolated cloud environments. Each task runs in an isolated OpenAI-managed container: it checks out the repo, runs setup with network access, then enters an agent phase that is offline by default; it returns a diff you can open as a PR or follow up on. As of 2026-08-05.6注 6Note 6OpenAI Codex 文档 · Subagents —— 当前 Codex 默认支持 subagent workflow:主 agent 把独立分块交给专门 agent,最后收摘要;CLI 用 /agent 查看与切换线程。适合读密集型并行,会比可比单 agent 任务消耗更多 token。eligible account 的 Work Ultra 可主动委派。截至 2026-08-05。OpenAI Codex Docs · Subagents — current Codex releases enable subagent workflows by default: the main agent assigns independent slices to specialists and collects summaries; /agent inspects and switches threads in the CLI. This fits read-heavy parallel work and consumes more tokens than a comparable single-agent task. Work Ultra can delegate proactively for eligible accounts. As of 2026-08-05.
— II
边界重画:能派多少,验得动多少The Boundary, Redrawn: What You Can Send Out — and Verify.
Codex 最省的不是敲代码的时间,而是你守着 AI 干活的时间。What Codex saves most is not typing time, but the hours you spend watching an AI work.
本书的第一条判断The first judgment of the book
所以量它的尺子得换。旧尺子问「它能写什么代码」—— 按这把尺子,它和 Claude Code、Cursor、Copilot 挤在同一条跑道上。新尺子问「有多少活可以不守着做完」—— 按这把尺子,并行有两种:一件复杂活让 subagents 分头查、主 agent 汇总;五个彼此独立的 bug 各开一个 worktree 或云端任务,你去干别的,回来挨个收。入口跟着活走:要陪跑的留在编辑器,要进脚本的走 CLI,要隔离并行或离机完成的进 worktree / cloud,要在团队流程里收的回到 GitHub。So the ruler has to change. The old ruler asks what code it can write — by that ruler it's crowded onto one track with Claude Code, Cursor, and Copilot. The new ruler asks how much work gets finished without you watching — by that ruler parallelism has two shapes: subagents investigate parts of one complex job and the main agent consolidates; five independent bugs each get a worktree or cloud task while you do something else, then come back one by one. Entrances follow the work: pair-work stays in the editor, scripts go to the CLI, isolated parallel or off-machine work goes to worktrees / cloud, and team delivery returns through GitHub.注意新尺子的后半句。派活容易,确认它真做对才难。官方最佳实践要求测试、检查与 diff 复审,再由你接受。4注 4Note 4OpenAI Codex 文档 · Best practices —— 把 Codex 当逐步配置与改进的队友;可靠性来自测试与复审:写 / 更新测试、跑检查、确认结果、审 diff 的 bug 与回归,再由你接受。官方写明 OpenAI 的 PR 全部经过 Codex review。截至 2026-08-05。OpenAI Codex Docs · Best practices — treat Codex as a teammate you configure and improve; reliability comes from tests and review: write / update tests, run checks, confirm results, inspect the diff for bugs and regressions, and then accept it. The docs say every OpenAI PR is reviewed by Codex. As of 2026-08-05.活越在看不见的线程和容器里完成,验收越是唯一兜底。本书把「验收」设为第 10 章,把多 Agent 的组队设为第 6 章:前者教你收,后者教你拆,两端必须成对。5注 5Note 5OpenAI Codex 文档 · Code review —— 本地 /review 起专门 reviewer 读选定 diff、按优先级报告发现且不改工作区;GitHub 上 @codex review 聚焦 P0 / P1,自动审查可覆盖每个 PR。截至 2026-08-05。OpenAI Codex Docs · Code review — local /review starts a dedicated reviewer that reads a selected diff, reports prioritized findings, and does not change the working tree; on GitHub, @codex review focuses on P0 / P1, and automatic reviews can cover every PR. As of 2026-08-05.Mind the second half of the new ruler. Sending work out is easy; confirming it is correct is hard. Official best practices require tests, checks, and diff review before you accept the result.4注 4Note 4OpenAI Codex 文档 · Best practices —— 把 Codex 当逐步配置与改进的队友;可靠性来自测试与复审:写 / 更新测试、跑检查、确认结果、审 diff 的 bug 与回归,再由你接受。官方写明 OpenAI 的 PR 全部经过 Codex review。截至 2026-08-05。OpenAI Codex Docs · Best practices — treat Codex as a teammate you configure and improve; reliability comes from tests and review: write / update tests, run checks, confirm results, inspect the diff for bugs and regressions, and then accept it. The docs say every OpenAI PR is reviewed by Codex. As of 2026-08-05. The more work finishes in threads and containers you cannot see, the more review becomes the only backstop. This book gives collection to chapter 10 and multi-agent division to chapter 6: one teaches you to collect, the other to split, and the two must stay paired.5注 5Note 5OpenAI Codex 文档 · Code review —— 本地 /review 起专门 reviewer 读选定 diff、按优先级报告发现且不改工作区;GitHub 上 @codex review 聚焦 P0 / P1,自动审查可覆盖每个 PR。截至 2026-08-05。OpenAI Codex Docs · Code review — local /review starts a dedicated reviewer that reads a selected diff, reports prioritized findings, and does not change the working tree; on GitHub, @codex review focuses on P0 / P1, and automatic reviews can cover every PR. As of 2026-08-05.
— III
先量一次你的活First, Measure Your Own Work.
往下读之前,先给自己的活拍一张底片:Before reading on, take a baseline photo of your own work:
01
翻两周,挑 5 件你亲手陪着 AI 做完的活Scroll back two weeks, pick 5 things you did alongside the AI by hand
不限写代码:改过的一个 bug、加过的一个小功能、清理过的一批依赖、写过的一份发布说明 —— 挑 5 件当时你一句句盯着、看它每一步的活。Not just coding: a bug you fixed, a small feature you added, a batch of dependencies you cleaned, release notes you wrote — pick 5 you watched line by line, checking each step.
02
给每件写一条「它能自己跑的检查」Write each one a check it could run itself
想象把这件活交给后台 worktree 或云端、你不在场:你能不能写出一条测试、一条命令、一个可对比的产出,来判断它做没做对?写得出、写不出、拿不准,各自标上。Imagine sending it to a background worktree or the cloud with you absent: could you write a test, a command, or a comparable output that decides whether it was done right? Mark each: can, can't, unsure.
03
数「写得出」的有几件Count the 'can' rows
那几件是你现在就能派出去、不用守着的活;「拿不准」的部分,是这本书接下来十一章要补的。读完回来重标一遍 —— 这张表就是你的前后对照。Those are the jobs you could send out today without watching; the 'unsure' rows are what the next eleven chapters fill in. Re-mark the table when you finish — it's your before-and-after.
聊天框只是入口,
活要去最合适的地方.
The chat box is only an entrance;
the work goes where it fits.
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
