How to use ChatGPT Work: cross-application long tasks, scheduled tasks and desktop Codex workflow

ChatGPT Work is for work that requires continuous progress: it can research and analyze information, collaborate between connected apps and files, and generate documents, sheets, presentations, reports, and Sites. Compared with one-time Q&A, the focus of Work is to allow you to check progress, add questions, adjust direction, or approve important actions during the execution process.

In July 2026, OpenAI also launched a new version of ChatGPT desktop that integrates Chat, Work and Codex, supporting Windows and macOS. This article compiles a set of usage methods suitable for real work based on [OpenAI official release instructions] (https://openai.com/products/release-notes/), and focuses on the boundaries of scheduled tasks, monitoring tasks, and local file permissions.

First choose the right entrance: what Chat, Work and Codex do respectively

These three can be used in the same project, but the nature of the tasks is different:

Entrance	suitable task	Things you need to pay attention to
Chat	Quick Q&A, writing, idea clarification	Give a clear context and confirm whether the conclusion needs to be implemented
Work	Research, organize materials, and continuously update delivery tasks	Stage goals, information sources, approval nodes and output formats
Codex	Modify code, review diffs, and handle warehouse tasks	Workspace scope, testing methods, Git changes and permissions

A simple way to judge is: if the task needs to span multiple files or applications, wait for external information, and need to repeatedly review intermediate results, use Work first; if you want to directly operate the code base or view PR diff, leave it to Codex; if an execution plan has not yet been formed, first clearly state the requirements in Chat.

Write long tasks into controllable delegates

Don’t just write “research this topic and write a report.” This can blur the scope of the task, the credibility of the information, and the standards for delivery. Better prompt words would include objectives, scope, data requirements, outputs, and approval points.

1
2
3
4


研究本周 AI 产品更新，优先使用厂商官方公告。
按发布日期倒序整理模型、API、桌面端和定价变化；排除未经证实的传闻。
先给出一页摘要和来源清单，等待我确认后再扩写为 1500 字报告。
如需访问连接的文件或应用，先说明用途并等待批准。

This arrangement has three benefits: first get an intermediate product that can be quickly reviewed; turn “official source” into an acceptance condition; and avoid accessing unnecessary external content before confirming the scope.

Use scheduled tasks to handle repetitive tasks

Work’s Scheduled Tasks can run once, repeat at a fixed time, or trigger monitoring when changes occur. It is suitable for tasks with clear rules and stable output format, such as:

Summarize last week’s product releases every Monday;
Check whether there are pending items in a shared table every day;
Monitor updates from specified sources and generate summaries after discovering new content;
Organize project progress at the end of the month and prepare meeting outlines.

Before creating, write down the trigger conditions and “what to do when there is no change”. For example, “Summary every Monday at 09:00; if there are no reliable updates, only reply without new additions, and will not supplement old news.” This prevents monitoring tasks from repeating old information in order to produce output.

For automations that write to external systems, it is recommended to only generate drafts or notifications at first. Wait until you confirm that the results are stable and the scope of permissions is correct before considering letting it perform subsequent actions.

Coordinate local files and applications on the new desktop

The new version of ChatGPT desktop puts Chat, Work, and Codex into one application. Work can use local files and desktop applications after obtaining user permission; the built-in browser can collect web page information and complete tasks with supported web tools and files.

In actual use, it is recommended to follow the following sequence:

First clarify the project directory or file set that should be accessed for this task;
Grant only the access required to complete the task;
Let Work first list the plan and the data to be used;
When it comes to code, switch to Codex to view diff, inline comments and test results;
Final human confirmation of deliverable document, report or submission.

Don’t interpret “local files can be used” to mean that all files should be handed over to the task. Minimal scope authorization reduces misuse and makes the review process clearer.

Use Codex to complete execution and review in the warehouse

Codex on the desktop supports editing Markdown and code directly, using inline comments in diff, viewing GitHub PR review feedback in the sidebar, and working with multiple repositories in one project. Suitable for turning solutions generated by Work into reviewable changes.

A common connection method is:

1
2
3


Work：根据需求和现有文档，写出实施方案、风险点与验收清单。
Codex：仅在指定仓库实现方案；每个改动说明原因，运行约定的检查，
不要提交或推送，直到我审查 diff。

This prevents “research conclusions” from being translated directly into unexamined code operations. Even if the task seems simple, the test command, configuration file scope, and whether commits are allowed should be written into the directive.

Points to note when using it in enterprises and teams

According to the official description, Work will be opened to paid plans in batches on the web and mobile terminals: Pro, Pro Lite, Enterprise, and Edu will gain access first, Plus and Business will follow later, and Free and Go will not be included in the first batch. Enterprise and Edu workspaces have a two-week preview period, during which Work is turned off by default, and administrators can opt out before automatically turning it on.

It is recommended to confirm before team deployment:

Which members can use connected applications and files;
Which tasks must retain manual approval;
Who will review the sharing permissions of reports, forms and Sites;
Do you need to encapsulate common processes into Plugins or Skills?

The original application connection will not be invalidated because the App Directory is changed to the Plugin Directory, but the data it can access and the scope of operations should still be reviewed before creating a new plug-in.

FAQ

Can Work replace Codex?

It cannot be completely replaced. Work is suitable for organizing research and delivery tasks; Codex is more suitable for viewing and modifying code, handling diffs and PR reviews. Using the two together is more controllable than cramming all the work into one conversation.

Will the monitoring task automatically make decisions for me?

Tasks can monitor changes and advance work, but important actions should have approval points. Especially when it comes to releases, external writes, deletions, payments, or access to sensitive files, ask them to report the plan and impact before deciding whether to approve.

Why can’t I see local files or applications on the desktop?

First check whether the corresponding permissions have been granted, whether the file is within the authorization scope of this task, and whether policy restrictions are imposed on the enterprise workspace. Do not extend access to the entire disk to troubleshoot a single file.

Conclusion

The correct way to open ChatGPT Work is not to “completely hand over” the work, but to split complex tasks into reviewable stages: clarify the goals, limit the scope of the data, look at the intermediate results first, retain approval for important actions, and finally complete verifiable delivery in Codex or desktop. This allows you to leverage cross-application collaboration while maintaining control over your code, files, and published results.