ChatGPT Dots Now Follow Custom Instructions: Set Tone and Which Actions Need Your Go-Ahead in the Thread, Plus Faster Replies, Less Jargon and Better Follow-Through

Background and the problem

Dots are OpenAI's always-on agents in ChatGPT. Powered by GPT-6 Astra, each dot has its own cloud computer, connects to apps such as email and calendar, and keeps working between conversations. Since late October, dots have added iMessage/RCS texting, replies from iOS notifications, and connections to several local computers. The hard problem for any agent that acts for you is knowing when to act and when to ask: ask too often and it is a burden, ask too rarely and it may send the wrong email or delete the wrong file. This update targets that problem.

How it works

The docs split control into two layers. Custom instructions cover style and habits, such as writing tone, how often to check in, and when to ask before acting. They belong to the dot alone, are separate from ChatGPT's own custom instructions, and can be set in chat for the dot to remember. Custom rules apply to specific actions with four levels: take action without asking, take action when you say so, ask before taking action, or hand off to you. On top of both, an automatic review runs before any action that could affect an account or share information, and decides whether it can proceed, needs approval, or must be done by the user, such as changing a password.

Benchmarks

OpenAI described the gains only as faster replies, less jargon and better follow-through. It published no latency, task-completion or benchmark numbers, so the size of the speed-up is not public. Only the feature changes can be confirmed; users will have to judge the effect themselves.

Getting started

1. Write rules straight into the dot's thread, for example "Show me a draft before sending any emails". 2. On mobile, open the sidebar, go to Customize > Behavior, pick your dot and edit how it handles tasks; update the app if Behavior is missing. 3. On web and desktop, open Settings > Personalization > Custom rules and set "Ask before taking action" or "Hand off to you" for risky actions such as messaging customers or deleting shared files. The docs warn that asking for drafted replies does not grant permission to send them, so scope instructions clearly.