OpenAI Upgrades ChatGPT to Execute Automated Work Across Apps

OpenAI

ChatGPT shifts from an answering assistant to an active autonomous workspace capable of managing cross-platform tasks, browser control, and code resolution.

AI-assisted summary of Get started with ChatGPT Work — original source. Check the original for full context.

Source published: 2026-08-03

Detailed notes

## Background Published on YouTube by OpenAI on July 9, 2026, this official walkthrough video introduces ChatGPT Work. According to the page description, ChatGPT Work provides a unified space across web, mobile, and the desktop app to delegate real work using connected apps, local projects, files, and browser automation.

## TL;DR The video demonstrates how ChatGPT Work moves beyond simple Q&A by actively executing complex, multi-step workflows. It showcases scheduled automations like daily briefing reports, dynamic marketing asset generation for a fictional neobank, and automated data consolidation that identifies a regional onboarding bug in Mexico. The presenter then uses appshots and Computer Use to let ChatGPT control the desktop screen, diagnose the software glitch, and submit a functional pull request via Codex.

## Atmosphere & style The presentation carries a polished, professional, and forward-looking tech-demonstration tone typical of OpenAI product launches. The narrative moves cleanly through realistic workplace scenarios, balancing smooth high-level workflow orchestration with deep-dive technical troubleshooting. It feels seamless, precise, and empowering, ending on a lighthearted note when the presenter casually asks ChatGPT to play walkout music.

## Key takeaways - [00:03](https://www.youtube.com/watch?v=GphgJjaKKhw&t=3s) ChatGPT Work shifts ChatGPT from responding to queries to actively executing tasks across web, desktop, and mobile platforms. - [00:03](https://www.youtube.com/watch?v=GphgJjaKKhw&t=3s) The platform integrates directly with everyday applications like Google Calendar, Slack, and Google Drive to access user data and context. - [00:03](https://www.youtube.com/watch?v=GphgJjaKKhw&t=3s) Users can set up recurring scheduled tasks, such as generating automated daily morning briefings with customized information. - [00:03](https://www.youtube.com/watch?v=GphgJjaKKhw&t=3s) Workflows synchronize seamlessly across web, mobile, and desktop applications, allowing users to transition between devices mid-task. - [01:42](https://www.youtube.com/watch?v=GphgJjaKKhw&t=102s) The marketing workflow supports creative asset generation, including brand-aligned launch site visuals, mood boards, and slide decks. - [02:37](https://www.youtube.com/watch?v=GphgJjaKKhw&t=157s) The platform consolidates fragmented metrics from multiple tools into single analytical views and flags data anomalies automatically. - [02:37](https://www.youtube.com/watch?v=GphgJjaKKhw&t=157s) Interactive dashboards can be created and modified in real time through direct on-screen annotations and collaboration. - [03:44](https://www.youtube.com/watch?v=GphgJjaKKhw&t=224s) Using Sites, reports and dashboards can be published as interactive web pages, shared automatically via Slack, and scheduled for updates. - [03:44](https://www.youtube.com/watch?v=GphgJjaKKhw&t=224s) The Cmd+Cmd keyboard shortcut takes an 'appshot' to provide ChatGPT with context-aware visual information about the active application. - [03:44](https://www.youtube.com/watch?v=GphgJjaKKhw&t=224s) Computer Use enables ChatGPT to take control of the cursor to navigate desktop applications, fill forms, and test interface functionality. - [05:09](https://www.youtube.com/watch?v=GphgJjaKKhw&t=309s) Developers can inspect code diffs and prepare Pull Requests directly within the integrated Codex environment.

## Detailed walkthrough ### [00:03](https://www.youtube.com/watch?v=GphgJjaKKhw&t=3s) Introducing ChatGPT Work and Scheduled Tasks

OpenAI presents a brand-new iteration of ChatGPT available across web, desktop, and mobile platforms. This update shifts ChatGPT from simply answering user queries to actively executing real work by connecting directly to user tools. The interface remains familiar, but users can now hand off substantive tasks. By granting access to applications such as Google Calendar, Slack, and Google Drive, ChatGPT can review upcoming obligations and compile daily preparation materials automatically. Beyond basic tool connections, plugins now offer role-specific bundled workflows tailored for departments like sales, data science, and marketing.

To eliminate the need for manual daily requests, tasks can be configured as recurring automations. For instance, ChatGPT can serve as a personal Chief of Staff by generating a daily morning briefing that outlines upcoming meetings alongside customized extras, such as World Cup match scores for specific teams. These workflows synchronize seamlessly across devices—a user can initiate a chat on the web interface, step away from their desk, and continue monitoring progress on a mobile device without interruption. Moving to the desktop application grants access to local files and projects, enabling users to establish context, delegate complex tasks, and re-engage only when input or review is required.

- Key integrations shown include Calendar, Slack, Drive, local project files, and role-specific plugin bundles. - Workflows automatically sync state between web, mobile, and desktop environments.

---

[01:42](https://www.youtube.com/watch?v=GphgJjaKKhw&t=102s) Marketing Workflow

To demonstrate a complex cross-functional scenario, the presenter walks through a project for a fictional neobank named Hudson as it prepares to launch its mobile app. Using ChatGPT's built-in image generation capabilities alongside brand guidelines and logos, the platform generates a wide variety of on-brand visual assets for a launch site. This capability allows teams to rapidly explore multiple creative directions while maintaining aesthetic consistency.

The project expands into an full advertising campaign where ChatGPT generates visual mood boards. Exploring conceptual avenues, such as a full takeover campaign, takes significantly less time than traditional manual brainstorming. Finally, ChatGPT converts these visual assets into a slide deck ready for team launch reviews.

- Fictional company used: Hudson (a neobank launching a mobile app). - Creative materials generated include launch site visuals, ad campaign mood boards, and presentation decks.

---

[02:37](https://www.youtube.com/watch?v=GphgJjaKKhw&t=157s) Data Analytics Workflow

Post-launch, the focus shifts to data analytics to assess early app performance. Because metrics, user feedback, team communications, and launch plans typically live in fragmented systems, ChatGPT consolidates these disparate sources. Recognizing a data-related query, the assistant automatically invokes relevant analytical skills and workflows to synthesize metrics, signups, activation rates, and onboarding funnel conversion steps.

During data exploration, ChatGPT automatically flags an anomaly: while overall launch metrics appear healthy, user activation drops significantly in Mexico due to a suspected phone onboarding issue. To share these findings effectively, ChatGPT constructs an interactive dashboard rendered in an in-app browser. Users can directly annotate elements live—such as adding a one-line summary—collaborating with ChatGPT in real time as adjustments render on screen.

- Consolidates scattered data from spreadsheets, data warehouses, and team comms into unified metrics. - Flags an operational issue specifically affecting phone onboarding activation rates in Mexico.

---

[03:44](https://www.youtube.com/watch?v=GphgJjaKKhw&t=224s) Publishing with Sites and Computer Use

To distribute reports efficiently across an organization, ChatGPT utilizes Sites to publish interactive web pages directly from the platform. With a single request, ChatGPT publishes the analytics dashboard, generates a shareable link, posts it directly to a team Slack channel, and schedules daily automated updates for the metric data.

To resolve the root cause of the Mexican onboarding bug, the workflow transitions to Codex within ChatGPT. To diagnose what customers are experiencing, the user runs an app simulator on-screen and takes an "appshot" using the `Cmd+Cmd` keyboard shortcut. An appshot functions as a context-aware screenshot that informs ChatGPT which application is active and provides context to perform actions. ChatGPT then initiates background Computer Use—taking direct control of the user's cursor to operate the app, fill out forms, and run comparison tests between valid Mexican and German phone numbers. When native app plugins are absent, Computer Use allows ChatGPT to directly manipulate desktop applications or navigate external websites using a browser extension.

- Keyboard shortcut `Cmd+Cmd` triggers an "appshot" to provide visual and functional application context. - Uses direct Computer Use to interact with simulated software and test Mexican phone number validation issues.

---

[05:09](https://www.youtube.com/watch?v=GphgJjaKKhw&t=309s) Code Resolution and Conclusion

Once the bug is identified—confirming that valid Mexican phone numbers fail validation during onboarding—ChatGPT is instructed to prepare a software fix. The proposed solution is rendered inside Codex, allowing developers to inspect code diffs directly within ChatGPT, review structural changes, and publish a Pull Request (PR) to ship the updated application build.

The walkthrough demonstrates ChatGPT Work as a unified workspace capable of delegating end-to-end tasks spanning web applications, local repositories, browser sessions, and system-level operations. To conclude the demonstration, the presenter playfully requests ChatGPT to "play some walkout music" to close the presentation.

- Automatically drafts a PR to fix phone number validation logic directly within the integrated Codex environment. - Learn more at [learn.chatgpt.com](https://learn.chatgpt.com/).

## Notable quotes > [00:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=0s) ChatGPT Work gives you one place to delegate real work across your apps, files, local projects, browser, and computer.

> [00:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=0s) In this walkthrough, see how to connect your tools, set up recurring work, move between devices, and collaborate with ChatGPT as it works toward an outcome.

> [05:09](https://www.youtube.com/watch?v=GphgJjaKKhw&t=309s) play some walkout music — Presenter

## Mentioned ChatGPT Work, OpenAI, Google Calendar, Slack, Google Drive, Hudson, Sites, Codex, World Cup, learn.chatgpt.com

[Source](https://www.youtube.com/watch?v=GphgJjaKKhw)

中文简报

OpenAI升级ChatGPT：支持跨应用执行自动化任务 ChatGPT正从问答助手转变为主动式自主工作空间，具备跨平台任务管理、浏览器控制及代码修复能力。该视频展示了ChatGPT Work如何超越简单的问答交互，主动执行复杂的多步骤工作流。演示内容包括每日简报等定时自动化任务、为一家虚构的新型银行动态生成营销素材，以及通过自动化数据整合排查出墨西哥地区的新用户引导漏洞。随后，演示者利用应用截图和“计算机使用”（Computer Use）功能让ChatGPT控制桌面屏幕、诊断软件故障，并通过Codex提交了功能性Pull Request。 OpenAI 推出 ChatGPT Work，直接整合 Google Calendar、Slack 和 Google Drive 等工具，支持在网页、移动端与桌面应用间无缝同步，并可设置每日早报等自动化定时任务。 (0:03) 在营销工作流中，ChatGPT 可基于品牌规范快速生成发布网站视觉资产、情绪板和幻灯片演示文稿，显著缩短创意探索周期。 (1:42) 针对数据分析场景，系统能自动聚合分散系统的指标并标记转化异常，同时通过 Sites 将交互式仪表板直接发布为网页并定时同步至 Slack 频道。 (2:37) 借助 Cmd+Cmd 截取上下文相关的 Appshots，ChatGPT 可通过 Computer Use 直接接管光标操作模拟器排查问题，并在内置 Codex 环境中生成代码差异与提交修复 PR。 (4:28)

Reading summary

OpenAI has introduced ChatGPT Work, an iteration of ChatGPT built to execute multi-step business workflows by connecting directly to workplace tools, automating scheduled routines, and operating desktop interfaces. Cross-Device Tool Integration and Recurring Automation (0:03) ChatGPT Work shifts the assistant from answering user queries to delegating and executing substantive tasks across web, mobile, and desktop environments. The platform integrates directly with tools such as Google Calendar, Slack, Google Drive, and local project files, while offering role-specific plugin bundles tailored for departments like sales, marketing, and data science. Ongoing tasks maintain continuous state synchronization across devices, allowing users to start tasks on desktop and monitor progress on mobile. Creative Asset Production and Funnel Analytics (1:42) In a demonstration featuring a fictional neobank named Hudson, ChatGPT combined brand guidelines with image generation to produce launch visuals, advertising mood boards, and review slide decks. Beyond creative assets, the platform consolidates fragmented post-launch data from spreadsheets, data warehouses, and team chat channels into a single funnel analysis. During data evaluation, ChatGPT automatically flagged an activation drop-off specific to phone onboarding in Mexico and built an interactive, annotatable web dashboard to display the findings. Autonomous Diagnostics via Computer Use and Codex (3:44) To share findings, ChatGPT publishes dashboards directly to the web via Sites, posts links into Slack channels, and configures recurring daily data updates. For technical troubleshooting, users can trigger an 'appshot' using the Cmd+Cmd shortcut to provide ChatGPT with active application context. Through Computer Use, ChatGPT takes direct control of the user's cursor to interact with simulated apps, navigate browser extensions, and run comparative form tests that confirm Mexican phone number validation failures. The assistant then renders code diffs in Codex and drafts a Pull Request to fix the logic. The walkthrough illustrates OpenAI's strategy to turn ChatGPT into an autonomous workplace operations layer that spans data analysis, design, and software patching, though the demonstration relied on a controlled, fictional scenario.

精炼阅读摘要

OpenAI 推出面向 Web、桌面和移动平台的 ChatGPT Work，展示了该系统如何从单纯的对话问答转向连接用户工具并主动执行端到端工作任务。多工具互联与例行任务自动化 (0:03) ChatGPT Work 扩展了传统交互模式，允许直接连接 Google Calendar、Slack、Google Drive 以及本地文件等工具，并为销售、数据科学和营销等部门提供特定角色的插件工作流。该系统支持将日常事项配置为周期性自动化任务，例如充当个人幕僚长角色，自动汇总日程安排并生成早间简报，同时可附带定制信息，如特定球队的世界杯比赛比分。工作流在多设备间保持状态同步。用户可在 Web 端发起任务，在移动端继续跟踪进度，或通过桌面端应用直接访问本地项目文件。在此模式下，用户可将复杂任务委托给模型处理，仅在需要人工审查或输入时再行介入。营销策划与品牌视觉资产生成 (1:42) 以一家名为 Hudson 的虚构虚拟银行为案例，演示展示了 ChatGPT 基于品牌规范和标志生成上线网站视觉资产的过程。模型能够在保持视觉美学一致性的前提下，辅助团队快速探索多种创意路线。除了基础视觉生成，ChatGPT 还能构建完整广告活动的视觉情绪板，大幅缩减传统头脑风暴的时间周期，并最终将创意资产直接转换为供团队上线评审使用的幻灯片演示文稿。跨源数据分析与交互式仪表板发布 (2:37) 针对应用发布后的运营分析，ChatGPT 能够整合来自电子表格、数据仓库和团队沟通记录的离散数据，对注册量、激活率以及新手引导转化漏斗进行综合分析。在数据探索过程中，模型主动标注出异常情况：整体指标表现良好，但墨西哥地区的手机注册激活率明显下跌。为呈现分析结论，ChatGPT 在应用内置浏览器中构建了交互式仪表板，支持用户进行实时标注。随后，系统通过内置的 Sites 功能直接生成可共享的网页链接，将报告推送至指定的 Slack 频道，并设置每日自动刷新数据指标的调度任务。计算机使用控制与代码修复部署 (3:44) 在排查墨西哥用户注册异常时，用户启动了应用模拟器并通过快捷键 Cmd+Cmd 触发 Appshot。该功能通过具备上下文感知能力的截屏，告知模型当前运行的程序并提供操作环境。随后 ChatGPT 启动后台计算机使用（Computer Use）功能，直接接管光标操作模拟器、填写表单，并对墨西哥与德国电话号码的验证流程展开比对测试。确认墨西哥有效电话号码存在校验失败的缺陷后，系统在内置的 Codex 环境中生成修复方案。开发者可直接在界面内审查代码改动细节（Diff），并提交 Pull Request（PR）以完成应用更新的打包发布。演示表明 ChatGPT Work 正在从单纯的内容生成工具演进为能够跨越 Web 应用、本地代码仓库与桌面操作系统的执行型工作空间，但其在实际企业生产环境中对非标准化软件操作的容错能力与安全性仍有待长期观察。

Detailed notes — zh

## Background 该视频由 OpenAI 于 2026 年 8 月 3 日发布在 YouTube 官方频道上。视频简介指出，ChatGPT Work 为用户提供了一个统一的平台，能够跨应用程序、文件、本地项目、浏览器以及电脑委派并执行真实工作，支持在网页端、移动端和全新 ChatGPT 桌面端之间无缝流转与协作。 ## TL;DR 本次演示全面展示了 ChatGPT 从单纯解答疑问向直接承担交付工作的质的飞跃。用户不仅可以通过自然语言设定定时任务并由 AI 充当数字幕僚长，还能在营销策划中依据品牌规范快速生成情绪板与汇报演示文稿。在数据分析场景下，ChatGPT 能主动捕获转化漏斗中的异常并一键通过 Sites 发布动态交互看板；面对工程问题时，则可结合 Appshot 快照与 Computer Use 物理光标接管技术进行界面复现，并自动完成代码修复与 PR 提交，构建了真正的大一统工作枢纽。 ## Atmosphere & style 整个演示节奏紧凑而充满科技感，呈现出一种前瞻且务实的专业基调。演讲者以极其沉稳、从容的口吻带领观众穿梭于各种真实而复杂的商业与工程场景之间，当屏幕上的光标在无需人为干预下自动移动接管操作时，带来了极具视觉冲击力与未来感的震撼体验。而在结尾处让 AI 播放离场音乐的幽默互动，又为整场严肃硬核的生产力展示增添了一抹轻快灵动的温度。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=0s) OpenAI 推出全新 ChatGPT Work 模式，标志着产品定位从单纯被动解答问题的聊天机器人，升级为直接跨应用调度与执行任务的工作委派中枢。 - [00:45](https://www.youtube.com/watch?v=GphgJjaKKhw&t=45s) 引入自然语言长期任务调度能力，用户可指定 ChatGPT 担任数字幕僚长自动生成每日早报，并实现网页端、移动端与桌面端的无缝跨平台协同。 - [02:37](https://www.youtube.com/watch?v=GphgJjaKKhw&t=157s) 推出 Sites 一键发布功能，无需前端工程师介入即可将 ChatGPT 生成的动态交互看板直接发布为在线网页，并支持按日自动刷新数据。 - [04:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=240s) 集成 Appshots 上下文快照与 Computer Use 物理光标操控技术，使 AI 能在无专用 API 时自主接管应用排障，并通过 Codex 直接完成代码修复与 PR 提交。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=0s) 核心产品定位质跃：ChatGPT 从被动的 Chat（聊天）模式拓展为 Work（工作）模式，底层交互逻辑转变为任务委派中枢，真正替用户交付实际成果。 - [00:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=0s) 企业级应用深度集成：开箱即用打通 Google Calendar 日历、Slack 沟通记录与 Google Drive 云端文件，实现跨系统整合与日程准备梳理。 - [00:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=0s) 垂直岗位工作流包：内置针对销售、数据科学、市场营销等专业岗位的端到端插件包，快速编排高专业度的复杂工作任务。 - [00:45](https://www.youtube.com/watch?v=GphgJjaKKhw&t=45s) 自然语言定时自动化调度：用户可通过简明指令设定长期角色与触发条件（如担任幕僚长每日汇总早报及赛况），摆脱每日重复下达任务的繁琐。 - [00:45](https://www.youtube.com/watch?v=GphgJjaKKhw&t=45s) 跨终端多端无缝协同：支持网页端发起、移动端行进中跟进以及桌面端深度执行，各端状态即时同步，提供异步“脱手”协作体验。 - [00:45](https://www.youtube.com/watch?v=GphgJjaKKhw&t=45s) 本地文件与项目感知：桌面端应用支持深度读取本地文件系统与工程工作区，左侧栏直观展示项目结构与后台运行中的任务。 - [01:42](https://www.youtube.com/watch?v=GphgJjaKKhw&t=102s) 端到端营销创意生产闭环：基于品牌 Logo 与设计手册批量生成一致风格的宣传页面与情绪板，并一键汇总导出为正式评审幻灯片。 - [02:37](https://www.youtube.com/watch?v=GphgJjaKKhw&t=157s) 多源异构数据统一聚合：无论底层数据来自数据仓库、本地电子表格还是第三方存储，均可自动汇总并生成全链路转化漏斗分析。 - [02:37](https://www.youtube.com/watch?v=GphgJjaKKhw&t=157s) 业务数据主动异常告警：在整体指标向好的情况下主动捕获细分市场（如墨西哥）异常流失，并给出手机号引导受阻的准确归因分析。 - [02:37](https://www.youtube.com/watch?v=GphgJjaKKhw&t=157s) 应用内交互与实时重构：内置浏览器允许用户直接点击动态看板进行行内批注，系统实时重构界面视觉与文本内容。 - [03:44](https://www.youtube.com/watch?v=GphgJjaKKhw&t=224s) Sites 动态看板一键公开发布：支持从对话流中一键将生成的前端交互看板发布至公网，支持团队链接共享与每日定时数据抓取刷新。 - [04:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=240s) Appshot 上下文快照机制：通过双击 Command（Cmd-Cmd）截取携带应用完整窗口语义与上下文状态的高级快照，指导代理执行具体排查。 - [04:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=240s) Computer Use 屏幕接管交互：在缺乏专用 API 时，AI 能够接管光标在本地模拟器或浏览器中进行点击和输入测试，模拟真实用户操作以复现缺陷。 - [04:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=240s) 端到端工程排障与代码提交闭环：从模拟器 Bug 复现、Codex 自动修改源码、界面清晰展示 Diff 差异，到人工复核后一键创建 Pull Request。 - [05:23](https://www.youtube.com/watch?v=GphgJjaKKhw&t=323s) 人机协作范式重塑：全面打破本地代码、SaaS 工具与操作系统视窗的壁垒，实现“全权代理自主执行 + 关键节点人工把关”的全新协作方式。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=0s) 重新定义生产力：从回答问题走向直接交付工作 OpenAI 宣布在网页端、桌面端以及移动端全面上线全新的 ChatGPT。这一次的核心演进标志着 ChatGPT 的定位发生质的跃迁——它不再只是一个被动回答提问的聊天机器人，而是通过直接连接用户的各类日常生产力工具，真正承担并执行实际工作。在展示中，用户可以从传统的“Chat（聊天）”模式直接一键切换至全新的“Work（工作）”模式。界面保持了用户所熟知的极简对话形态，但其底层的交互逻辑已转变为任务委派中枢。用户在指令中随意抛出复杂的统揽任务，例如要求它查看日历日程、Slack 讨论记录以及 Google Drive 云端硬盘文档，并以此为基础准备当天的工作要点。在后台运算的同时，演示展示了其背后的插件系统（Plugins）：这些插件不仅能够与主流应用程序建立深层链接，还打包了面向销售、数据科学、市场营销等特定岗位的专属端到端工作流。片刻之后，ChatGPT 便完整归纳出了全天日程全貌、会议背景资料以及对应的参会准备要点。 ```mermaid flowchart TD accTitle: ChatGPT Work 多应用协同示意图 accDescr: 展示 ChatGPT Work 如何整合日程、通讯、文件与插件并输出工作交付物 subgraph DataSources ["多源上下文输入"] A["Google Calendar 日历"] B["Slack 团队沟通"] C["Google Drive 云端文档"] end subgraph CoreEngine ["ChatGPT Work 核心引擎"] D["语义解析与任务编排"] E["岗位工作流与应用插件"] end subgraph Deliverables ["任务交付成果"] F["日程与会议准备简报"] G["自动化待办规划"] end A --> D B --> D C --> D D <--> E D --> F D --> G ``` - **核心定位转变**：从单纯解答疑问（answering questions）升级为替用户实际执行工作（doing work for you）。 - **工具深度集成**：开箱支持连接日历、Slack、云端硬盘等企业级应用。 - **专业工作流包**：提供针对销售、数据科学、营销等垂直工种预设的角色插件包。 --- ### [00:45](https://www.youtube.com/watch?v=GphgJjaKKhw&t=45s) 定时任务与跨设备无缝协作：打造专属数字幕僚长为了避免每天重复下达相同的准备指令，ChatGPT Work 引入了长期运行的自动化调度能力。用户可以直接以自然语言下达长期角色设定与触发条件，例如对它说：“你能否担任我的幕僚长（Chief of Staff），在每天清晨为我生成一份当天准备工作的早报？同时，在世界杯期间，帮我同步德国队的比分以及其他我关心的赛况。”完成这一句设定后，自动化任务即刻部署生效，后续无需用户反复手动提醒。与此同时，这种工作流具备极致的跨终端移动连续性。在桌面端启动的任务和对话状态，会瞬间同步至移动端手机。用户在办公室电脑前完成初步输入后起身离开，在通勤或行进途中掏出手机即可继续跟进刚才的任务。而在切换至全新的桌面端应用时，ChatGPT 则能进一步读取本地文件系统与本地项目目录。左侧栏清晰罗列了所有项目结构与正在运行的后台任务，使得用户能够把上下文交代清楚后放心“脱手”，只在系统需要确认或产出最终结果时再介入审查。 - **自然语言定时自动化**：支持以对话方式创建每日早报，结合工作备忘与用户个人兴趣（如世界杯赛事比分）。 - **跨平台多端连贯**：网页端启动、移动端跟进、桌面端深耕，各端状态即时同步。 - **本地项目感知**：桌面端支持直接读取本地文件与工作区，构建“布置任务—后台处理—就绪通知”的异步协作模式。 --- ### [01:42](https://www.youtube.com/watch?v=GphgJjaKKhw&t=102s) 市场营销工作流：从品牌规范到投放全案与演示文稿在展示更复杂的复合工作流时，演示以一家名为“Hudson”的虚构数字新银行（neobank）筹备上线移动应用为实战场景。在营销筹备阶段，ChatGPT Work 展现了从视觉创意发散到专业汇报材料生成的端到端执行力。用户仅需向其输入品牌 Logo 和设计规范手册（Brand Guidelines），ChatGPT 就能借助图像生成能力输出符合调性的上线宣传页面，并探索多种视觉方向，同时严格保持色彩与品牌元素的一致性。基于确立的品牌调性，ChatGPT 进一步将设计延展为完整的广告投放企划，快速生成不同视觉风格的情绪板（Mood Board）。“我现在构思多种创意方案所需的时间，比过去深入推敲单一方案还要短，”演讲者尤其对其中的“Takeover 视觉包揽方案”赞不绝口。在选定满意的视觉方向后，ChatGPT 顺理成章地将所有创意与方案框架转换为一套精致的幻灯片（Deck），直接作为当天上线评审会议的正式汇报材料。 - **品牌资产对齐**：依托 Logo 与品牌指南，保持批量视觉探索与生成过程中的风格统一。 - **广告概念速测**：一键生成各类户外投放与线上物料的情绪板概念。 - **全案演示生成**：自动汇总方案思路与视觉成品，直接导出可用于评审会演示的 PPT 简报。 --- ### [02:37](https://www.youtube.com/watch?v=GphgJjaKKhw&t=157s) 数据科学与 Sites 一键发布：全链路监控与动态看板当新银行应用正式推向市场后，业务视角随之切换至数据分析。为了评估上线初期的真实表现，团队需要聚合分散在各处的碎片信息——核心业务指标、用户直接反馈、内部协作讨论以及最初的发布规划。针对这一分析需求，ChatGPT Work 自动激活了专业的数据科学技能与分析流，自动抓取并整合了包括激活率、注册转化率以及完整的新手引导转化漏斗（Onboarding Funnel）。底层数据无论是来自独立数据存储、电子表格还是云端数据仓库，都能被顺畅接入。在各项指标整体向好的态势下，系统敏锐地捕捉到了异常数据：“整个上线指标看起来普遍很健康，但在墨西哥市场似乎有些不对劲。”ChatGPT 率先将该问题标出，并研判可能是手机号注册环节出现了阻碍。为了方便团队长期观测，演示者未选择制作静态幻灯片，而是直接要求 ChatGPT 生成一个动态看板。通过内置的应用内浏览器，用户可以直接点击页面进行行内批注（如添加上线概况的一句话摘要），并实时观察看板界面的即时重构。为了将分析成果同步给整个组织，系统调用了名为“Sites”的发布功能，允许用户直接从 ChatGPT 内部将看板发布为可交互网页。随后系统自动生成分享链接，发送至团队频道，并被配置为每天定时拉取最新数据自动更新。 ```mermaid flowchart LR accTitle: 数据分析到 Sites 动态看板发布流程 accDescr: 从多源指标抓取到问题告警再到应用内批注与 Sites 一键发布 subgraph Ingestion ["数据整合与告警"] M1["指标 / 反馈 / 表格"] --> M2["漏斗模型与激活率分析"] M2 --> M3["异常检测: 墨西哥手机引导受阻"] end subgraph Dashboarding ["交互看板构建"] M3 --> D1["生成应用内动态看板"] D1 --> D2["点击批注与实时渲染"] end subgraph Publishing ["团队共享"] D2 --> P1["Sites 一键发布上线"] P1 --> P2["推送团队频道并按日自动更新"] end ``` - **数据源无缝打通**：无论底层是数据仓库、本地电子表格还是自定义存储，皆可全自动汇总。 - **主动式业务归因**：自动识别特定国家地区（墨西哥市场）漏斗流失率异常，并给出疑似原因推测。 - **Sites 直连发布**：无需前端开发介入，一键将生成的前端看板发布至公网，支持团队内共享与每日定时刷新。 --- ### [04:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=240s) Appshots 与 Computer Use：从工程排错到代码提交为了解决数据看板揭示的墨西哥手机验证漏洞，用户无缝切换到了集成在系统内的 Codex 编程模块，化身软件工程师进行深度排查。得益于先前的分析，排障目标非常明确，但仍需在真实端侧确认用户实际遭遇的操作卡点。演讲者启动了移动端模拟器（Simulator），调出运行中的 App，随后按下了快捷键 `Cmd-Cmd` 截取了一张“Appshot”。Appshot 并非普通的静态截屏，它不仅捕捉应用当前的界面图像，更完整提取了窗口的上下文状态与语义信息。随后，系统触发了“Computer Use（计算机使用）”能力，在无需用户触碰键盘鼠标的情况下接管了屏幕光标。“那不是我的鼠标，那是 ChatGPT 正在操作我的应用，”屏幕上的光标自动点击输入，对比测试墨西哥与德国的手机号验证流程。在没有专用 API 插件的情况下，Computer Use 支持代理程序像人类一样操控本地桌面或借助 Chrome 插件操作浏览器，完成复杂的复制粘贴、多系统跳转以及功能验证。测试结果直接证实：当真实用户输入合规的墨西哥本地手机号时，系统的校验逻辑确实存在卡死阻断缺陷。锁定根因后，工程师下达最终指令让其准备修复补丁。Codex 随即对项目源码进行修改，并在界面内清晰展示代码差异对比（Diff）。经过人工复核无误后，工程师一键创建并发布 Pull Request（PR），准备将新版本正式打包上线。 - **快捷键调用 Appshot**：双击 Command（`Cmd-Cmd`）截取携带应用状态上下文的高级快照，指导代理采取具体行动。 - **Computer Use 物理接管**：在缺失专属 API 时，AI 能直接操控屏幕光标与点击，在本地模拟器或 Chrome 浏览器内完成复杂测试与数据录入。 - **端到端工程闭环**：从现象复现、根因定位、代码改写、Diff 审查到推送 PR，全部在统一工作区内搞定。 --- ### [05:23](https://www.youtube.com/watch?v=GphgJjaKKhw&t=323s) 真正的工作枢纽：重塑人机协作形态在演示接近尾声时，演讲者总结了这一代 ChatGPT Work 的宏大蓝图：它真正成为了用户托付实际工作的大一统中枢。无论是本地开发项目、SaaS 应用程序，还是直接在操作系统和网络浏览器内进行复杂界面交互，ChatGPT 都能提供无缝覆盖的工作流支持。为了给整场发布演示画上一个轻松诙谐的句号，演讲者对着系统调侃道：“嘿 Chat，我的演示结束了，你能放点离场音乐（walkout music）吗？”这一即兴互动展现了其贯穿严肃专业生产力与轻松日常协作的双重特质。官方表示本次演示仅是 ChatGPT Work 能力的一个快速概览，期待全球用户在其赋能下探索出更多前所未有的工作范式。 - **全场景工作委派**：打破应用孤岛与运行环境限制，将本地代码、云端工具与桌面视窗完整串联。 - **新型人机协作范式**：从命令提示词响应转为“全权代理+关键节点人工把关”的深度协作体验。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=0s) 从被动回答提问的聊天机器人，转变为通过直接连接用户的日常生产力工具，真正承担并执行实际工作。 > [00:45](https://www.youtube.com/watch?v=GphgJjaKKhw&t=45s) 你能否担任我的幕僚长（Chief of Staff），在每天清晨为我生成一份当天准备工作的早报？同时，在世界杯期间，帮我同步德国队的比分以及其他我关心的赛况。 > [01:42](https://www.youtube.com/watch?v=GphgJjaKKhw&t=102s) 我现在构思多种创意方案所需的时间，比过去深入推敲单一方案还要短 > [02:37](https://www.youtube.com/watch?v=GphgJjaKKhw&t=157s) 整个上线指标看起来普遍很健康，但在墨西哥市场似乎有些不对劲。 > [04:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=240s) 那不是我的鼠标，那是 ChatGPT 正在操作我的应用 > [05:23](https://www.youtube.com/watch?v=GphgJjaKKhw&t=323s) 嘿 Chat，我的演示结束了，你能放点离场音乐（walkout music）吗？ > [04:00](https://www.youtube.com/watch?v=GphgJjaKKhw&t=240s) 在没有专用 API 插件的情况下，Computer Use 支持代理程序像人类一样操控本地桌面或借助 Chrome 插件操作浏览器 > [00:45](https://www.youtube.com/watch?v=GphgJjaKKhw&t=45s) ChatGPT Work 引入了长期运行的自动化调度能力 ## Mentioned OpenAI, ChatGPT, ChatGPT Work, Google Calendar, Slack, Google Drive, Hudson, Sites, Codex, Appshot, Computer Use, Chrome ## Video description ChatGPT Work gives you one place to delegate real work across your apps, files, local projects, browser, and computer. It works across web, mobile, and the new ChatGPT desktop app. In this walkthrough, see how to connect your tools, set up recurring work, move between devices, and collaborate with ChatGPT as it works toward an outcome. Get started: https://learn.chatgpt.com/ Chapters: 00:00 Introducing ChatGPT Work 00:45 Scheduled tasks 01:42 Marketing workflow 02:37 Data analytics workflow 03:44 Publishing with Sites 04:28 Appshots and Computer Use 05:23 One place for real work [Source](https://www.youtube.com/watch?v=GphgJjaKKhw)

Source quotes

[0:00] 从被动回答提问的聊天机器人，转变为通过直接连接用户的日常生产力工具，真正承担并执行实际工作。 [0:45] 你能否担任我的幕僚长（Chief of Staff），在每天清晨为我生成一份当天准备工作的早报？同时，在世界杯期间，帮我同步德国队的比分以及其他我关心的赛况。 [1:42] 我现在构思多种创意方案所需的时间，比过去深入推敲单一方案还要短 [2:37] 整个上线指标看起来普遍很健康，但在墨西哥市场似乎有些不对劲。 [4:00] 那不是我的鼠标，那是 ChatGPT 正在操作我的应用 [5:23] 嘿 Chat，我的演示结束了，你能放点离场音乐（walkout music）吗？ [4:00] 在没有专用 API 插件的情况下，Computer Use 支持代理程序像人类一样操控本地桌面或借助 Chrome 插件操作浏览器 [0:45] ChatGPT Work 引入了长期运行的自动化调度能力

Source quotes

[0:00] ChatGPT Work gives you one place to delegate real work across your apps, files, local projects, browser, and computer. [0:00] In this walkthrough, see how to connect your tools, set up recurring work, move between devices, and collaborate with ChatGPT as it works toward an outcome. [5:09] Presenter: play some walkout music

Original transcript

[0:03] Today we're excited to release a brand new [0:06] ChatGPT on web, desktop and mobile. [0:09] Now ChatGPT goes from just answering [0:11] questions to doing work for you [0:13] by connecting to all of your tools. [0:16] I can switch from chat to work [0:18] and this interface should feel very familiar. [0:20] But instead of just having conversations, [0:23] I can hand off real work from here. [0:25] Take a look at my calendar, Slack [0:27] and Drive and help me prepare for my day. [0:30] While that's running, let me show you plugins. [0:33] Some of them connect to the apps you use, [0:35] others bundle role-specific workflows for sales, [0:38] data science, marketing and more. [0:39] And here we go. [0:41] It found everything about my day, the notes [0:43] and how to prepare for my next meetings. [0:45] The next thing I want is for this to happen [0:48] without me asking every time. [0:50] Let's try this. [0:52] Can you be my Chief of Staff [0:54] and give me a briefing every morning [0:56] of what I have to prepare for the day? [0:58] And while the World Cup is going on, [1:00] can you actually give me the scores [1:03] of the German team and any other [1:04] matches that I care about? [1:06] And that's it. [1:07] The automation is set and I don't [1:08] have to remember to ask again tomorrow. [1:11] And by the way, this also works on the go. [1:13] The same chat is already on my phone. [1:15] I can start on the web, leave [1:17] my desk and keep going on my mobile. [1:20] Now when I want ChatGPT to work [1:22] with my local files and projects, [1:23] I move to the desktop app. [1:25] On the left, I can see all [1:27] of my projects and everything ChatGPT [1:29] is working on, but this time [1:31] with all of my local files as well. [1:33] For me, this is a big shift. [1:35] I can give it the context, [1:37] hand off the work, and then come back [1:38] when it needs me or has something ready. [1:41] Let's look at something more involved. [1:42] I'm going to show you three different [1:44] examples across marketing, data analytics and engineering. [1:48] Let's say I'm building a fictional [1:50] neobank called Hudson and we're [1:52] getting ready to launch our mobile app. [1:54] Here's the launch site. [1:56] All of these visuals were generated [1:58] using ChatGPT and image generation. [2:00] All I did was give it my logo, [2:03] my brand guidelines, and it's incredible how [2:05] many different directions I can explore [2:07] and how consistent they are with our brand. [2:10] The next thing I want to do is translate [2:12] this into an ad campaign. [2:13] We can now create [2:15] mood boards for what this could look like. [2:17] I can now imagine many ideas in less time [2:19] than it would take for me to explore one. [2:21] I love this takeover concept. [2:24] It looks amazing. [2:25] Next, I asked ChatGPT to turn [2:27] these into a deck to share with my team. [2:29] Beautiful. Ready for our launch review meeting today. [2:36] Let's say we just launched the app. [2:38] Switching gears to data science, I want [2:40] to understand how it's actually going. [2:42] The metrics, the user feedback, team conversations, [2:45] launch plans, all are in different places. [2:47] And I use ChatGPT to bring them all together. [2:50] Since it's a data question, ChatGPT automatically [2:53] brought in the right skills and workflows. [2:56] And this is what it came up with. [2:58] Activation rates, signups, and even this onboarding funnel. [3:01] Here I'm using my own data store, [3:03] but your data could just as easily [3:04] live in spreadsheets or a data warehouse. [3:07] The launch looks healthy overall, [3:08] but something looks off in Mexico. [3:11] In fact, ChatGPT already flagged it and it [3:13] thinks it might be a phone onboarding issue. [3:16] So we'll have a look at this in a moment. [3:18] I could build this into a deck and share [3:20] this with my team, but a real-life [3:22] dashboard would be much more useful. [3:24] So I asked it to pull these metrics into this dashboard. [3:27] I can visualize it [3:28] in the in-app browser and I can click to annotate. [3:31] Add a one-line summary on how the launch is going. [3:35] I love collaborating with ChatGPT like [3:37] this and seeing the changes happen live. [3:39] Awesome. It looks good. [3:42] Now I want to share it with my team. [3:44] Sites is a great way to do this. [3:45] It lets me publish websites directly from ChatGPT. [3:49] Publish this dashboard, share the link [3:51] in my team channel, and make sure [3:53] that the metrics are updated every day. [3:55] And just like that, [3:56] we have built and shared the dashboard. [3:58] Next, it's time to switch [4:00] to Codex and put my engineering hat [4:02] on to fix this mobile app issue. [4:04] If you're already a Codex user, [4:06] it's the same Codex you know and love. [4:08] Now inside of ChatGPT. [4:10] Thanks to ChatGPT, we know where to look, [4:12] but let's confirm what the customer is actually experiencing. [4:15] I'm going to use the Simulator [4:17] and bring the app on the screen [4:18] and take an appshot by pressing Cmd-Cmd. [4:22] Review the phone onboarding in Mexico, [4:25] then in Germany to compare [4:26] and click around the main screens. [4:28] You can think of an appshot as a smarter screenshot. [4:31] It gives it the app I'm looking at [4:33] and the context it needs to take action on. [4:36] It's now triggering background Computer Use. [4:39] That is not my cursor. [4:41] This is ChatGPT using my app. [4:43] When ChatGPT doesn't have a dedicated plugin, [4:45] it can use your computer directly. [4:47] It can do the same thing [4:48] in the browser with the Chrome extension. [4:51] Copying and pasting data, [4:52] browsing sites, or checking a workflow. [4:55] It looks like customers are indeed [4:57] getting stuck when they're entering a valid Mexican phone number. [5:00] That explains why the activation rates are lower. [5:03] Let's also have it check [5:04] everything works great for other countries in the background. [5:07] Now for the final step, [5:09] let's actually have it fix the problem. [5:11] Prepare a fix for the [5:12] phone number validation bug in Mexico. [5:14] And when it's ready, I can see [5:16] the diff directly within Codex, [5:18] review the changes, and publish this PR [5:20] so we can ship a new version of the app. [5:23] So this is the new ChatGPT. [5:25] One place where you can delegate real work. [5:28] Whether that's in your apps, your [5:30] local projects, or even your computer and browser. [5:33] There's really no task [5:35] that ChatGPT cannot help you with. [5:37] Hey Chat, I'm done with my demo. [5:39] Can you play some walkout music? [5:42] This was only a quick tour [5:43] and we can't wait to see what you do with it.