How to Build Your Own JARVIS With OpenAI Dots

Become a paying subscriber to get access to all courses, guides like this, prompts and skills library, and $500+ worth of AI tools.

OpenAI Dots can check your schedule, sort your inbox, prepare meeting notes, and draft messages while you're away.

You get one Dot per account, working for you. This guide gives you the prompts and a four-week rollout to decide what it handles and what needs approval.

Step 1: Set the rules. Define its job and approval boundaries.

Step 2: Connect your apps. Connect the sources it needs and test access.

Step 3: Teach it how you work. Save your priorities and preferences.

Step 4: Give it standing jobs. Schedule updates and choose where they arrive.

Step 5: Review, then loosen. Check its work before allowing more actions.

Early reviews showed useful catches: calendar conflicts, emails relevant to upcoming meetings, and even a flight delay spotted before an airline notice.

They also found problems to check during setup:

The setup below addresses those problems.

Create your Dot in the ChatGPT desktop app or a desktop browser. Setup isn't available on mobile web. Give it a name, which sets its handle.

Skip app and computer connections until you've set its instructions.

Paste this first:

Automatic review checks actions against your instructions, permissions, custom rules, and safety requirements. An action may proceed, require approval, or need you to complete a step yourself.

Go to Settings, then Personalization, then Permissions, then Custom rules. Select Add, describe an action, choose its handling, then select Add rule.

Start with these suggested rules:

Drafting doesn't authorize sending. Rules can also be followed incorrectly, so review early actions. They don't grant app access or override required confirmations.

Connect only what your first job needs:

Messaging access is separate from app access. Adding Slack doesn't connect your inbox. Your Dot can use a personal email account, but can't have its own address at launch.

Check access:

For website logins, use the private sign-in form. Credentials go to the browser outside the conversation. To sign in yourself, open your Dot's profile, go to Computers, open its cloud computer, select Take over, then Return control when finished.

Connect your own computer only for local files or apps, or when a website blocks the cloud browser. It must stay online with ChatGPT open. Fix access to required sources before continuing.

Your Dot keeps notes separate from ChatGPT memory. Ask to see what it saves:

A useful note looks like this:

Add two instructions to prevent common mistakes:

Keep the approved note in an accessible document your Dot checks before working. Changing ChatGPT memory settings may not change notes already saved.

Start with the morning brief. Add other jobs after checking its results.

The result should be easy to scan:

Open Scheduled and verify the saved routine. For phone updates, install the ChatGPT mobile app when available to your account. Add contact methods from your Dot's profile.

Slack channels need explicit monitoring instructions.

Use this weekly:

Allow one new action at a time after a clean week. For example:

Choose Take action without asking for that specific rule. Keep other invites subject to approval.

To stop everything, check current work, delegated tasks in Activity, and recurring tasks in Scheduled. Pause stops the main task only. Stopping work doesn't undo completed actions. Deletion is irreversible; read its confirmation first.

Add these once the morning routine works.

Afterward:

Open the PDF before forwarding it.

Check bookings in the provider's app or email.

At launch, you get one Dot. Additional Dots are planned as paid add-ons; pricing hasn't been announced.

Dots require Pro or Business Premium. Launch availability varies by region and account; Enterprise needs admin approval.

Dot conversations don't count toward usage limits, but Work and Codex tasks do. Ongoing deeper-work allowances haven't been published.

Website access may require sign-ins or handoffs. Custom rules can fail, and completed tasks still need checking.

Dot notes are separate from ChatGPT memory. An outside report says clearing them requires deleting the Dot. Applicable ChatGPT data controls cover eligible conversations.

OpenAI disclosed research-agent incidents involving Hugging Face, Services Australia's Medicare reporting service, and 53 externally posted user images. OpenAI said no patient records were accessed. These aren't established Dots incidents.

Most hands-on reports remain early launch experiences.

Start with one job you can check each morning. Set the rules, connect your calendar and email, and run the access check before starting the routine.

Teach your Dot how you work, schedule the brief, and review the results each week. Give it more freedom one action at a time, once you're happy with how it handles the work.

See you in the next one :)