An AI assistant becomes useful when it understands what you are trying to achieve. But useful context and unlimited access are not the same thing. The practical question is: how do you get meaningful help without sharing more information—or handing over more control—than the task requires?
In Chloe Shih’s Office Drama interview, published on 8 October 2026, OpenAI’s Holly Li discusses using dots as an ongoing thinking partner, for monitoring developments and preparing work for review. The conversation also addresses privacy, permissions and the limits of safeguards. This guide draws on those themes; the prompts and recommendations below are AGENTICA IX’s own practical suggestions.
Watch the practical-use discussion, or jump to the privacy section.
1. Start with one useful, low-risk job
Choose a task where you can easily judge the result: turn rough notes into a plan, compare public information, or prepare a draft. Avoid beginning with your most sensitive records or an instruction to manage everything.
A useful first prompt: “Turn these three priorities into a plan for tomorrow. Ask about missing deadlines, flag assumptions and give me the next action for each.”
OpenAI’s getting-started guide describes dots as agents that can work towards a goal with context, connected apps and ongoing feedback. Available tools and access still matter for any particular task. Read the getting-started guide.
2. Give a brief, not just a command
A short, well-bounded brief should explain the result you want, which sources are relevant, when you need it and where human review belongs. This reduces guesswork without requiring a long prompt.
For example: “Using this approved project summary, draft a client update for Friday. Highlight missing facts. Keep it here for review; do not send it or include information from other projects.”
Voice can also help you untangle a problem. Finish by asking for a written summary of decisions, assumptions and next steps, so a useful conversation becomes something you can check.
3. Separate access from permission to act
Before connecting an app, ask what the task actually needs. A meeting brief may need a small set of documents rather than a broad collection of customer records. Where practical, provide a relevant extract with unnecessary identifying details removed.
For recurring work, begin with notification-only monitoring: “Check the approved project source each weekday. Tell me if a deadline changes, with the source and time checked. Do not contact anyone.” Add action authority deliberately after you have seen useful, reliable results.
Treat client, employee and other people’s information as a separate responsibility. An account being yours does not automatically make every record in it appropriate for an AI task. Follow your organisation’s approved tools and data-handling requirements.
4. Understand which privacy control does what
OpenAI’s current FAQ distinguishes several controls. Existing plugin permissions are shared across dots, ChatGPT, Work and Codex. Disconnecting a plugin stops new access; it does not erase information already retained. Personal-plan model improvement follows the “Improve the model for everyone” setting; Business, Enterprise and Edu data is not used for training by default. Opting out does not prevent every form of limited human review.
The FAQ also says individual dot memories cannot currently be directly viewed, edited or deleted. Deleting the dot removes its own context, while separately stored files, conversations and ChatGPT memories need separate consideration. Check the current documentation before relying on a deletion or retention assumption. Read OpenAI’s privacy, security and safety FAQ.
The practical lesson: decide what is suitable to share before connecting or uploading it. “Stop accessing this” and “remove everything previously retained” are different requests. Li discusses this distinction directly in the interview’s Gmail-disconnection example.
5. Keep review at the point of consequence
Specify the destination and limits when approving external work. “Draft a response” is a useful first step; a message sent to the wrong person is a different kind of outcome. Check recipients, attachments, confidential details and factual claims before consequential communication.
OpenAI documents Custom Rules and action checks, but these do not eliminate mistakes or malicious instructions embedded in external content. Supported secure sign-in flows keep credentials out of model context; that protection does not cover passwords pasted into chat. Use the supported sign-in process. See the documented safeguards.
6. Ask for evidence, not just reassurance
A polished answer is not proof that a task finished. Ask what was actually checked, what remains uncertain and where you can review the output. For a document, inspect the file. For a scheduled check, verify the scope. For a website change, inspect the actual page.
Useful follow-up: “What did you complete, what evidence confirms it, and what still needs my decision?” Keep the answer focused on outcomes rather than a stream of routine activity.
A practical starting point
Pick one repetitive task this week. Give your dot a clear outcome, only the information it needs and an explicit review boundary. Check the result before expanding its role.
The goal is less mental load and better follow-through—not maximum automation. Start small, keep sensitive information deliberate, and expand only when the results justify it.
Editorial note: Interview themes are paraphrased; example prompts are original and are not quotations from Holly Li. Product documentation checked 10 October 2026. Features, settings and access may change. This article is independent AGENTICA IX commentary, not an OpenAI endorsement.