I spent months building a team of AI agents.
Then I became their assistant.
They felt more like new hires than effective bots. I kept explaining what we’d already decided, checking whether the work was actually done and correcting assignments that shouldn’t have reached my desk.
I’d built a team to take work off my plate. Why was I spending so much time managing theirs?
I built those agents in Codex. I also tried Claude Cowork and Perplexity. I was willing to learn the tools and spend time making them work. Most people don't want that to become another job.
That's what interests me about Dot: the possibility of getting useful help without first learning how to build a team of agents. Could someone on my team use it? Could Lauren? Could someone who reads this article set one up and hand it a real responsibility? That's the promise I want to test.
That’s the problem I’m trying to solve as I set up my ChatGPT Dot. Before I give it more responsibility, I want decisions to carry forward and ownership to hold after the conversation ends.
The problem goes deeper than a noisy briefing. An assistant can find a real request, summarize it correctly and still give me the wrong job.
It might read the original email and miss the reply that settled it. Or find an assignment without the later meeting decision that changed who owns it. The source is real. The recommendation is already out of date.
Giving me a link to that first email doesn’t fix the recommendation. I still have to reconstruct what happened afterward. The agents have retrieved information, but I’m still doing the work of figuring out what it means for today.
Remember what changed
Being copied on an email doesn't make me responsible for every next step. If a meeting assigns routine execution to someone on the team, I don't want the afternoon briefing quietly assigning it back to me.
More history only helps if the assistant can distinguish a current instruction from the one it replaced. That's where the new-hire feeling comes from: I'm repeating context I thought we'd already established.
For my morning brief, I want roughly 150–220 words: usually no more than three material items and two actions for me. That's a limit on the report's length. The research still needs to be thorough.
The afternoon brief is the test
There’s already an inventory of existing AI tasks to work from. I want those routines reconciled before adding more. A new tool shouldn’t leave two briefings giving me different instructions about the same assignment.
The useful test is one actual workday. Start with the morning decisions. Prepare for the meeting using the latest source. Then check whether the afternoon review understands what happened between them.
If I answered the email, that request should disappear from my action list. If the meeting assigned the next step, the responsible person should stay responsible. If something changed enough to need another decision, explain what changed.
These are the conditions I want to test. I’m not claiming the full sequence has already worked.
What has worked is narrower. One automated read test in AppFolio and Bloom completed on September 29 without asking me to sign in again or provide input during that run. That’s one successful test. Recurring reliability still has to be established.
Vendor-insurance instructions have been saved, but their behavior in the actual workflow remains untested. An accounts-payable software correction has been merged, but that doesn’t establish that the live queue is clear.
I don’t want “saved” reported as “working.” Those distinctions tell me what’s ready for another test and what still needs proof before I rely on it. If the afternoon brief brings back the same resolved item, the workflow needs correction. Adding another schedule won’t settle the decision it missed.
Set up your own Dot around one useful job
You don't need to spend months building agents to learn from what went wrong with mine. Start with something you already know how to evaluate. A meeting brief. A weekly household plan. A list of unanswered client questions. If you can't recognize a useful result, you won't know whether the setup is helping.
1. Create your Dot, then give it a job
As of September 30, OpenAI says to create your Dot in the ChatGPT desktop app or on desktop web and follow the onboarding prompts. Mobile access comes afterward where available. Availability varies by account and workspace. The official setup guide has the current details. You can name it. More importantly, tell it what responsibility you're trying to hand over.
“Help me be more productive” gives it very little to work with. “Prepare me for tomorrow's owner meeting using the current email thread and the last meeting's decisions” gives it an assignment you can inspect.
2. Show it what good looks like
A new hire needs more than a job title. Giving an assistant a long description of your life won't teach it how you want a particular assignment handled either. Start with the current process and one example of work you'd be happy to receive again.
Explain what makes that example good. Maybe it makes the recommendation clear before presenting the options. Maybe it separates an owner's decision from the team's next step. Include the exceptions. Where does the written process stop being enough? What can the assistant decide, and when does it need another person's judgment?
For someone on our Coastline team, that could mean the relevant Whale procedure, a strong completed example and the approval boundaries for that assignment. Use the current version and explain which instruction replaces an older one. Don't upload every handbook and hope the assistant sorts out the contradictions.
At home, the equivalent could be a trip-planning routine, your preferences and an itinerary you liked. Explain why it worked. Otherwise, the assistant may reproduce the layout and miss the judgment behind it.
Try this:
Use this process and completed example to prepare a first version of [assignment]. Follow the current process. Where it leaves a material question unanswered, point out the gap and recommend a way forward. Don't invent a rule to fill it.
Review the result. Give a specific correction: “The recommendation belongs at the top,” or “Getting an estimate isn't approval to order the work.” Then give it a different case. Can it apply the correction without you repeating it? Can it recognize an exception instead of copying the example blindly?
That's what I mean by training here: instructions, examples, practice and feedback. Uploading a handbook doesn't establish that the assistant can follow it.
3. Connect a source and prove it can use it
Connect the app needed for that first assignment. In ChatGPT, review available connections under Settings → Plugins and complete the service's authorization steps. A work account may also need administrator approval. OpenAI's app setup instructions explain the process.
Then test a record you recognize. Ask it to find a specific thread, read the latest reply and explain what remains open. Check the answer against the source yourself. A connected-account badge doesn't tell you whether it read the message that changed the decision.
Your Dot has a cloud computer; connecting your own computer is optional. Don't make a complicated technical setup the first assignment unless the work actually needs it. See the current computer-access guidance.
4. Give it room to work
Explain what it can handle and which decisions belong to you. Make this specific to the job, rather than burying it under a page of general restrictions.
For meeting preparation, it can research and prepare the brief. Sending the resulting message to an owner is a separate action. For a household plan, it can compare options and organize the day. A booking or purchase needs whatever approval you've agreed to provide.
Here's a starter instruction you can adapt:
Help me take [one recurring responsibility] off my plate.
Use [the connected sources] and the context below. Before recommending action, check the latest replies and decisions. Keep work with the person responsible, and don't ask me to decide something I've already settled.
Give me [the useful output] by [the time or trigger]. Lead with what needs my attention and your recommendation. Link the evidence that matters.
You can [the actions you want handled]. Bring me [the specific decisions you want to retain]. If a source is unavailable, explain how that changes the answer and continue the work you can do.
Relevant context and corrections: [your role, people, commitments and preferences].
Replace the brackets with your actual situation. This is an assignment, not a magic prompt.
5. Change something and see whether it keeps up
A first answer can look impressive. The more revealing test comes after you give it a correction. Tell it a decision has been made or a task has changed hands. Later, ask for the updated brief. Does it carry that change forward? Can it show what remains open without making you explain the whole story again?
For a Coastline team member, the test could be a repair estimate that gets approved between two reviews. The next brief should follow the approved work, not request the same approval. For a household plan, change the agreed departure time. The next version should adjust the plan rather than preserve both times and leave you to sort them out.
6. Make the useful result repeat
Once the result is useful and the correction holds, ask your Dot to run that assignment on a schedule. Review recurring work under Scheduled in its profile. OpenAI's scheduling instructions describe the available controls.
Check existing schedules first. You don't need two assistants reminding you about the same settled question. Then review what the next run actually produces. A schedule that exists and a routine you can rely on are different accomplishments.
Start with one job. Give it the context that matters. See whether your answer survives the next conversation. That's a useful first setup whether you're running a company or trying to make next weekend easier.
And it's the standard I'm applying to the team of agents I spent months building: help me move the work forward, without making me their assistant again.