How to Use OpenAI Dots: Your First Work Task

To use OpenAI dots, create your dot on desktop, choose the access the task needs, and give it a result you can check. Start with a draft rather than an action on someone else’s account. OpenAI’s setup guide lets you skip app connections and add them later. Connecting everything is not a prerequisite for learning how to give a good brief.

This walkthrough uses an illustrative weekly project update. The prompts are practice examples, not results from a hands-on test. Interface and rollout details were checked on 30 September 2026. For plan and regional eligibility, start with what OpenAI dots is and who can access it. This article focuses on the task itself, from the first request to reviewing the output.

Before setup: choose a task you understand

A useful first assignment has a source you can inspect and a result you already know how to judge. Meeting notes into a project update is a good exercise. An instruction to negotiate a contract is a poor first exercise because you may not see every trade-off until after the conversation has happened.

Use a small example that contains no confidential information you are not allowed to share. If you are testing with work material, check your employer’s approved tools and account policy first. The point is to learn the method, not to prove that the agent can reach every system you use.

Write the expected result before opening the interface. In this example, the update should contain decisions, completed actions, open questions and next steps. It should not invent a deadline or turn a suggestion into a decision. That definition gives you something concrete to compare with the eventual draft.

Step 1: create your dot on desktop

OpenAI’s access and setup guidance says to create a dot in the desktop app or desktop browser. Mobile creation and mobile web are not supported. Mobile app access follows setup when available. Rollout is gradual, so check eligibility if the entry point is missing.

Use the account you intend to keep using. In a workspace, establish whether dots has been enabled before assuming an installation problem. Do not move work documents to a personal account simply because its interface looks different. Access questions and data-sharing decisions should be settled before the first work task.

Appearance and naming can wait. A name can help you recognise the conversation, but it does not improve the task brief. Keep your attention on what you want done, which source should control the answer and who will review it.

Step 2: connect only the tools this assignment needs

For this exercise, you can supply permitted notes directly. An ongoing update from a shared folder may need that folder’s app connection instead. OpenAI’s computer and app documentation separates cloud work, app access and a connection to your own computer. Name which source the agent should use rather than leaving that choice implicit.

Before approving a connection, read the account name and available actions. “This is the project folder” is different from “use anything in my Drive”. If the task only needs the current notes, do not write an open-ended instruction to search the whole company archive. More context can introduce old versions and unrelated information.

Make a source list: this week’s notes, last week’s approved update, and the current action list. Mark the approved update as a style reference, not evidence of what happened this week. That one distinction prevents the agent from carrying old claims into a fresh report. Finding the repetitive part of your week helps you choose a repeatable assignment without handing over the whole job.

Step 3: give a brief with a visible stopping point

“Write my weekly update” leaves too much unstated. It does not identify the audience, the controlling sources or the difference between a draft and a message ready to send. Try this instead.

Use the notes and action list I supplied to draft this week’s project update for my review. Use last week’s approved update only as a style reference. Include decisions, completed actions, open questions and next steps. For each action, preserve the named owner and deadline; write “not specified” when either is missing. Do not send the update, edit the source files or contact anyone. Flag contradictions before producing a final draft.

This is a proposed brief. It is not a guarantee of what the system will do. You still need to inspect the result. The brief makes errors easier to identify because each requirement can be checked against a source or an explicit boundary.

Use the same structure for other work. A supplier comparison needs criteria. A meeting summary needs a distinction between decisions and discussion. A report needs a named audience. The principles in writing a prompt that works first try apply before you add agent features.

Step 4: inspect progress without rewriting the whole brief

OpenAI’s control guide describes Activity for reviewing tasks, files, results and requests for input. Use that view when a task needs access or a decision. A status update should lead you to evidence, not become the evidence itself.

Suppose the notes say a supplier confirmed delivery, but the action list still says confirmation is pending. Resolve that specific conflict. Tell the dot which source is current and why. “The notes were taken after the supplier call; use the confirmed delivery status and leave the cost unresolved” is more useful than “make it accurate”.

Avoid giving several corrections in vague shorthand. If you write “same as last time”, name what should stay the same: section order, tone or approval boundary. The previous draft might also contain a date or owner that should change. Clear corrections make a continuing task easier to manage.

Step 5: check the draft against the source

Read the source and the draft side by side. Do not start by polishing the language. First establish whether the meaning survived. An elegant update that changes “we might delay” into “we will delay” can cause a problem even if every name is spelled correctly.

• —Trace each decision to the notes that support it.
• —Check that action owners and deadlines were preserved.
• —Keep missing information visible rather than guessing it.
• —Remove claims carried over from the style reference.
• —Confirm the result is still a draft and has not been distributed.

Ask for a correction you can verify: “Move the supplier price discussion into open questions. The notes contain no approved price.” Then inspect the revised paragraph. Checking an AI answer when you are not the expert covers the source-checking habit when you cannot assess every fact yourself.

You can also ask for a short list of changes from the first draft. Read that list against the document rather than trusting it alone. This is particularly useful when you requested several edits and need to know whether all of them landed.

Step 6: add recurring work only when the exercise works

OpenAI’s tasks and memory guide says a recurring brief should identify what to do, timing and time zone, notification conditions and delivery destination. Ask for confirmation of the saved schedule. Do not infer that a casual request for future help created one.

Each Friday at 9am Europe/London for the next four weeks, use the agreed project folder to prepare an update draft for my review. Bring it to me in ChatGPT. If the current notes are missing or conflict with the action list, ask me rather than filling the gaps. Do not send the update to the team. Confirm the schedule, sources and end date you saved.

This example is a suggested instruction, not a task scheduled by this article. Check the confirmation and inspect the schedule in your account. Decide who maintains the source folder when you are away. An unattended draft based on missing notes is not a successful weekly process.

Decide where updates belong

OpenAI’s messaging guide explains that ChatGPT, Slack and Teams can reach the same dot, but their visible messages remain separate. Adding a contact method does not itself establish monitoring. Name the destination and the information you are comfortable sharing there.

For a first exercise, keep updates in the conversation you are using. If you later choose a team channel, specify whether the dot should bring you a private draft or post an approved update. Those are different audiences. OpenAI dots permissions explains how to keep the task’s authority clear as you add connections.

What counts as a useful first result?

Judge the attempt by corrections needed, evidence preserved and the quality of the final draft. Do not invent a productivity percentage from one run. You may discover that preparing the sources costs more effort than writing the update yourself. That is still a useful finding: the task may need a better source process before automation helps.

Keep the brief, one checked draft and a note about the errors you found. Reuse the brief only after updating its sources and dates. For larger recurring work, business process automation strategy starts with the process you can inspect before expanding it.

Coursium teaches practical AI skills on your phone. The useful practice here is simple: state the result, keep uncertainty visible, read the evidence and give precise feedback. A first task is successful when you understand what happened well enough to decide what comes next.