How to Use ChatGPT Work Mode

ChatGPT Work is OpenAI's agentic task feature. You describe a multi-step task, ChatGPT proposes a plan, you approve it, and the task runs on OpenAI's servers. You can close your browser and return when it is finished. This is different from a standard ChatGPT conversation, where nothing happens unless you are sending messages.

Getting useful results from a Work task comes down to two things: a specific brief before the task starts, and a clear record of the scope so future runs do not require rewriting the same instructions. A tool like fenbs keeps that scope and the next step outside any single chat session.

What ChatGPT Work can do

ChatGPT Work handles tasks that involve browsing the web, reading and summarizing pages, filling out forms, conducting research across multiple sources, and compiling results. It operates in a sandboxed browser environment, not on your local machine.

It cannot access your local files, applications, or desktop. Tasks that require reading a file from your computer need you to upload the file first. Check OpenAI's current feature pages for the full list of what Work tasks can do, as capabilities change.

How to write an effective brief

Writing a clear task description is the most important step. ChatGPT proposes a plan based on what you write, and a vague description produces a plan you may not want to approve.

• State the goal precisely: "Find the pricing page for each company on this list and note the starting price for their business plan."
• Set the scope: "Research only these five companies. Do not follow links to unrelated pages."
• Describe the output: "Return a table with columns: company name, plan name, starting price, date retrieved."
• Name what needs approval: "If a page requires a sign-in to see pricing, stop and ask me how to proceed."
• Set the end condition: "Stop when you have a row for each company or have noted that pricing was not publicly available."

Review the plan before you approve it

Before a Work task starts, ChatGPT proposes a plan showing what steps it intends to take. Read through it. If a step looks wrong or broader than you intended, revise your task description and ask for a new plan. Approving a plan that includes a step you are not comfortable with means that step may run while you are away.

The plan approval is your most reliable control over the task. Use it.

Supervising a Work task in progress

Once approved, the task runs. You can monitor its progress in the ChatGPT interface: there is a status view showing what it has done and what it is currently working on. If you see it taking an unexpected direction, you can interrupt it with a message.

For tasks on sensitive data or with real-world consequences, such as filling a form that will submit an order, stay available to review the progress rather than walking away entirely on the first run.

Reviewing the result

When the task finishes, review the output against your original task description. Check that each item you asked for was addressed, that any figures or facts are plausible, and that the output format matches what you requested. If the task missed something or went off-scope, update your task description for next time.

Treat the first run of any new task type as a learning run: your goal is to understand what the task produces and what needs to be tightened in your instructions, not to rely on it for a critical deliverable.

Keeping a record between task runs

Each Work task is an independent session. ChatGPT does not remember what it did in a previous run unless you include that context in the new task. For recurring work, such as a weekly research summary or a regular audit check, save a brief note of what the last run found and what to look for next time.

That note is most useful when it lives outside the chat. If it is in a previous conversation, you have to find it and copy it into the new task. If it is in a shared record, you can link to it or paste the relevant section directly into the task description.

In fenbs, Where we left off records what was done and what comes next. Before a new Work task run, check that note first. It takes a few seconds to read and prevents the task from duplicating work or missing the follow-up from the previous run.

What ChatGPT Work does not do

ChatGPT Work does not access your local files, read your email, control your desktop applications, or take actions in accounts you have not explicitly connected. It does not remember context from previous Work task sessions unless you include that context in each new task description.

How to do this with fenbs

1. Create a free account and add a task for the job you want ChatGPT Work to handle.
2. Write the task description in the task note: goal, scope, output format, and approval checkpoints.
3. Connect ChatGPT over MCP from Integrations and ask it to read the project context and rules before starting.
4. After the Work task runs, review the result and update the task status.
5. Ask ChatGPT to update Where we left off with what was done and the next step.

For more on ChatGPT Work compared to Claude Cowork, see ChatGPT Work vs Claude Cowork. For the broader ChatGPT agent mode, see ChatGPT agent mode.