What is ChatGPT Dots (OpenAI Dots)? The always-on ChatGPT agent: who can get it, how to set it up, and what to do when it gets stuck
ChatGPT Dots (OpenAI Dots) is an always-on agent inside ChatGPT with its own cloud computer, so it keeps working while your computer is off. It's included in ChatGPT Pro ($100, $200 or $500 a month, outside the EEA, UK and Switzerland) and Business Premium. Who can get it, eight setup steps, which jobs need your laptop online, why overnight work can stall on an approval, and a first message to copy.

The short version: what ChatGPT Dots is, who can get it, how to set it up, and where it gets stuck
In one line
ChatGPT Dots (also called OpenAI Dots) is an always-on agent inside ChatGPT, launched on 29 September 2026. You give your Dot an ongoing responsibility and the sources it needs, and it keeps working between conversations, messages you with results, and asks when a decision needs you. It runs on GPT‑6 Astra and has its own cloud computer and browser, so it can keep going while your own computer is off. It is worth turning on if you have work that continues while you're away: an inbox to watch, a project to keep on track, a trip that changes, or code that can be built overnight. In izumi_ren's week on Zenn, about three hours of his own instructions and checks were enough for the Dot to put in 60 to 70 hours of work and ship a small bot service with a landing page and a LINE connection. It isn't yet an assistant you can leave fully alone. It works quietly, repeated tasks can slip, an approval may not reach a task it split off, and work that needs your laptop stops when the laptop goes offline. Most people start with one ongoing job, read-only access and a few custom rules, then widen them as the results hold up.
How to set it up in eight steps
- Check that your plan and region qualify: ChatGPT Pro (Pro 100, 200 or 500, at $100, $200 or $500 a month) for users over 18 outside the EEA, UK and Switzerland, or Business Premium in every supported region. Enterprise, Edu and Healthcare get a beta an admin has to switch on. Access rolls out gradually; izumi_ren upgraded from Plus to Pro 100 and still had to wait before the Dots screen appeared. Who qualifies where is on availability, the plan and login details are on How to get ChatGPT Dots, and if you qualify but see nothing, Dots not showing or not working sorts the fixes by symptom.
- Open Dots in the latest ChatGPT mobile app, the ChatGPT desktop app (Mac or Windows), or chatgpt.com/dots in a desktop browser (mobile web isn't supported). Name it and finish the introduction. Corey Noles found the most useful first step was letting the Dot interview him about his work. You then reach the same Dot from every device, and from Slack or Teams once you add them. Getting started walks through the first job.
- Connect only the apps the first job needs. Where an app can't be limited to reading, set its plugin permission to read-only tools, and leave out any app that holds things you don't want read, as tadashi-aikawa did with Google Drive.
- Give it one ongoing responsibility: what should stay true, where the sources are, and when it should interrupt you. A daily news digest or one project's deadlines are common first jobs.
- Write custom rules one action at a time under Settings > Personalization > Custom rules. しろ's split works as a starting point: research and drafts without asking, scheduling and internal messages ask first, anything sent outside the company plus money and permissions handed off to you.
- Decide where each job runs before you step away. Research, writing and coding in a Codex cloud environment you've already set up keep going with your computer off. A local Work or Codex task needs the connected computer online with the ChatGPT app open, according to OpenAI's Tasks and memory guide, and it doesn't move to the cloud when that computer drops off. After two weeks away for a wedding in Okinawa, with his laptop sometimes closed and sometimes left open, Amon concluded that every job has to be sorted in advance: cloud-only research and writing, code for a Codex cloud environment, work that needs his own logins or apps, work that needs the screen, and steps that need his approval.
- If you use Codex, set it up as well. In a ToTheMoon episode on two weeks with Dots, Alexander Volchek noticed that his Dot hands much of its work to Codex as tasks; those tasks show up under Activity and count toward Codex limits, so @thelaggingway put Codex cloud tasks behind an ask-first rule.
- Check Activity and Scheduled once a week, and open Activity whenever the Dot goes quiet: each task there shows its progress and any request for your input. Pause stops only the Dot's main task, so delegated tasks are stopped in Activity and recurring ones are switched off in Scheduled.
Where it gets stuck, and what worked
- Work handed over at night hasn't moved by morning. Amon's tasks sat waiting for more than four hours overnight, because a split-off task showed an approval screen that never reached the conversation with the Dot, and other jobs that needed no approval stalled too. Amon's day instruction names the hours he can't reply, whether the laptop is closed, the top priority, and a time for results: cloud research and writing go ahead, anything needing approval goes on a waiting list, and at 6 p.m. the Dot hands over the draft and the open questions. He adds that an instruction alone doesn't guarantee the Dot notices a stuck task, so check Activity for requests for input yourself.
- The same approval is asked for again and again. Ren's approval didn't travel to the tasks the Dot split off; writing the permission once into each child task got the work moving.
- The automatic review refuses a step you already said yes to. Cameron64's turns were interrupted after three refusals ("rejected too many approval requests for this turn"). Retry at most once, then ask for a separate approval that names the outside service and what goes to it.
- "I'll work on it", then an hour or more of silence. izumi_ren found the Dot reports only at the end, so it is hard to stop when it heads the wrong way. Open its task in Activity to watch or correct it. The same Dot polished small details, such as the bot slipping from polite to plain Japanese endings a few percent of the time, until he hit the weekly limit several times; say plainly what is good enough.
- A message goes out early, or a draft waits forever. McAteer and the AI Never Sleep writer used the same habit: drafts first, then "send it" for that exact version. Drafts also catch a message that would go out as a new email instead of a reply, which happened to Alexander Volchek.
- A rule fires unevenly or doesn't reach the step. A plugin permission, a required confirmation or the task's own wording can decide instead. Read the refusal in Activity and ask the Dot which rule it applied; rules written too broadly were turned down outright.
- Approvals interrupt you all day. Desmond Dixon counted them among the reasons he stopped. An approval you give every day is a candidate for a custom rule.
- A screen reader can't reach the sign-in. On the first night, Taylor Arndt, who is blind, found the cloud browser's login shown as an image her screen reader couldn't use. Two weeks later a secure sign-in form let her connect LinkedIn, though finding the request in the chat still took hunting with Tab. Voice mode worked well for her from the start.
- Work on the cloud computer disappears. dafire warns that the cloud environment can reset without notice, so save projects to Space or push them out in time.
- An overnight build drifts. youer found the model sometimes adds constraints of its own, so the plan and limits need to be explicit before you leave it running.
- Memory mixes other people's details with yours. Volchek's advice is to say when a question is about someone close to you rather than yourself.
If Dots doesn't show up at all, the fixes are sorted by symptom on Dots not showing or not working; if it says "Your dot is on a break", see what that message means. Cost and the other common questions are answered briefly in the ChatGPT Dots FAQ.
What it can and can't do yet
It can:
- keep a job going between conversations and message you with results or questions
- work on its own cloud computer and browser while your computer is off, and let you take over a step
- read the apps you connect within the permissions you already gave
- draft emails, invoices and replies, and send them after your yes
- hand work to Codex and ChatGPT Work, and take voice calls; Taylor Arndt says talking things through by voice lets her get about twice as much done
- build and revise a small product over several days, as in izumi_ren's bot service
It can't yet, or not reliably:
- carry on a local task in the cloud when your computer goes offline (Tasks and memory)
- connect to more than one personal computer at a time (John Voorhees)
- run as several Dots on one account, so separate jobs share one conversation and the replies get mixed together (izumi_ren)
- run repeated tasks without slipping (Volchek, @argbeknwn)
- reach another person's Dot outside work accounts (Volchek)
- get past every captcha on its own (Reece Rogers, WIRED)
- run for personal Pro accounts in the EEA, UK and Switzerland (availability)
The jobs people hand over most are steering Codex, building apps and sites, travel planning, voice calls, small business and morning briefs.
One message to start with
Interview me for ten minutes about my work, the people I deal with, and what you should always check with me first. Then take on one ongoing job: [job]. The sources are [apps, docs, channels]. Before I go offline, tell me which parts need my computer and which can run in the cloud. Show me drafts before anything is sent, and treat "send it" as approval for that exact version only. Don't buy, delete, share files or write to someone new without asking. If one part needs my approval, park only that part, keep going on the rest, and give me one list of pending approvals at [time]. Tell me only when a deadline is at risk or you need a decision, and once a week list the approvals you asked me for and suggest custom rules for the ones that repeat.
What a Dot does that a normal chat doesn't
A normal ChatGPT chat waits for your message, answers it, and stops. A Dot is built for work that continues: it tracks progress, works out what has to happen next, and follows through as things change. You can come back days later with new information or priorities without restarting the work.
A few things make that possible, according to OpenAI's Meet dots guide and the Help Center:
- It decides when to follow up. Your Dot can pause and wake itself to continue, run reminders and recurring checks you ask for, and use background agents to work on several things at once while you keep talking to it.
- It has its own computer. The cloud computer and browser handle research, files and software. You can open it and select Take over to do a step yourself, then Return control. What people actually run on it, and what to check when the browser or a long task stalls: real uses of the Dot's cloud computer.
- It remembers. It starts with your relevant ChatGPT memory and keeps its own notes on your preferences, decisions and ongoing work, across every channel you use to reach it.
- It uses the apps you connect. Gmail, Google Drive, GitHub and other plugins work within the permissions you already granted. Connecting your own computer is optional, starts off, and needs the ChatGPT desktop app on that computer.
- It can hand work to Codex and ChatGPT Work. Those tasks show up under Activity in your Dot's profile, where you can check results and give further instructions.
- You set the boundaries. Built-in safeguards and approval checks apply from the start, and custom rules let you choose whether an action happens without asking, needs your approval, or is handed to you. Pause stops your Dot until you resume it.
Talking to your Dot doesn't count toward your ChatGPT usage limits. Work it starts in Codex or ChatGPT Work counts toward those products' limits, and your plan includes an allowance for deeper work, with extended limits for the first month after launch. The numbers and the "on a break" message are covered on the pricing and limits page.
What people actually use it for
These are the jobs people have reported most often since launch. Each link goes to a page of named, source-linked cases.
- Morning brief and daily digest, from a brief the Dot wrote for itself to overnight reports
- Email triage, from a reply stuck in spam to 53 follow-up drafts
- Steering Codex, from overnight PRs to unsticking delegated tasks
- Travel planning, from booking mistakes caught to cancelled trains
- Small business, from a first $25 customer to a client brief, payroll and ads
- Research and competitor watching, from 24/7 price tracking to hourly checks on 4,891 inns
- Meeting prep, from sales-call briefs to a transcript checked after the call
- Slack assistant, from a client channel steered to a decision to a 50-minute wait in a team channel
- Voice calls, from 10 tasks handed over at once to 3D models
Three snapshots of what people handed over in the first days
The job links above go to the longer pages. The three short reports below show what using ChatGPT Dots looks like in practice: one starts from email, one runs from a phone, and one leaves code building overnight on the cloud computer.
BuildnWrite (Korea, Pro, 1 Oct)
Before any ask, the Dot offered three tasks picked from connected email. The owner chose a hiring-evaluation document review: the Dot read the email plus two attachments, compared them with earlier messages, flagged an explanation that no longer matched a later setting, and drafted a reply without sending. It also raised an unprompted warning that a storage plan was about to expire. BuildnWrite's advice for day one: start with read and summarise permissions. Full write-up: BuildnWrite first-day review. More first-week reports are on how to use ChatGPT Dots; inbox cases sit on email triage.
Katie Fifer (Technically Curious, 4 Oct)
Entirely from her phone that day, Katie's Dot invoiced a customer (matched the amount in an email against the contract, created the QuickBooks invoice, drafted the send email), resolved two scheduling conflicts (emailed to fix them and updated the invites), and drafted a customer proposal from notes and recent proposals. Smaller ops also moved: forwarding mail, an interview invite and a software-billing check. Mistakes still happened (a Zoom link broke while updating an invite; some forwards went out as .eml and looked like spam). Full write-up: Meet Your New Work Bestie. Related pages: small business, voice calls.
youer (LINUX DO, 10 Oct)
youer uses the Dot's cloud computer to build Godot game demos: discuss the idea and constraints clearly during the day, then leave the Dot developing for ten hours or more overnight, and play the demo the next day to see whether the idea is fun. youer notes the model sometimes adds its own constraints, so prompts need to be explicit, and that connecting a local computer for coding uses allowance while cloud work does not in the same way. Thread: LINUX DO, in Chinese. Longer build stories are on building apps, websites & games; coordinator / parallel-workflow reports are on steering Codex.
What these three have in common
- It often starts from connected context, not a blank prompt. BuildnWrite's Dot opened with three email-derived offers. Katie's Dot already had calendar and mail. Asking "what can you help with?" or accepting one proactive suggestion is a common first move.
- Phone and voice keep ops moving when you're away from the desk. Katie's whole day ran from the phone. More people who get work done over a call are on voice calls.
- Overnight work belongs on the cloud computer once the plan is clear, and that computer can reset. youer left a Godot build running for 10h+. In the same LINUX DO thread, dafire describes the Dot as a coordinator that can run about six workflows at once, and warns that the cloud environment can reset without warning, so projects on it should be saved to Space (or pushed out) in time. Grumble8167's same-day report: drop a repo onto the Dot's computer to test, with multi-concurrent work.
Two asks based on these reports:
Pick one proactive offer, draft only:
From my connected email (and calendar if linked), list three things you could take on today. For each, say what you'd read, what you'd draft, and what you'd leave for my approval. Don't send, delete, or change anything until I pick one and say so.
Overnight cloud build once the plan is agreed:
Let's spend this session locking the plan for [demo / feature]. When I say the plan is clear, build it on your cloud computer overnight. Don't use my local computer. By morning, leave a short note: what shipped, what you couldn't finish, and where the files live. If the cloud environment looks like it might reset, save or push the work first.
Every job is in the use cases hub, and the workflows section turns the common ones into setups you can copy. If you already have access, how to use ChatGPT Dots walks through the first job and what people changed in their first week.
First days with a Dot: real reports
1) Access at 11 p.m., about five minutes of setup (Corey Noles, The Neuron)
In a first-day review, Corey Noles writes that Dots arrived on the account around 11 p.m. Noles named the Dot Herman, connected apps and started talking, and by the next morning was calling Herman to pick up the same conversation. The review puts the functional setup at roughly five minutes of questions, clicks and permissions, and calls letting Herman ask interview questions about the work the best step. Herman later spotted an email about a Ray-Ban warranty claim and offered to help reply or arrange a repair.
Try this first.
Interview me for ten minutes about my work, the people I deal with, and which actions you should always check with me first. Then suggest one thing you could keep track of for me this week.
2) Permissions asked for one at a time (John Voorhees, MacStories)
In MacStories' first look, John Voorhees named the Dot Puck. The article describes Puck starting with access to nothing beyond its own virtual Linux computer and asking for each tool only when a task needed it, a pace Voorhees found natural rather than interrupting. Puck sits just under "New Chat" in the sidebar as one thread that stays synced, with no model or reasoning level to pick. The main limit was that a Dot connects to only one Mac at a time, so switching to a Mac mini meant authorizing it from that machine.
3) The difference from a chat in one line (Flavio Copes)
Flavio Copes' deep dive puts it simply: a normal conversation waits, answers and stops, while a Dot takes on work that continues. You describe something that should stay true over time, point it at sources, and tell it when to interrupt you. The guide also notes that adding a Dot to Slack or Teams doesn't start any monitoring until you tell it what to watch for.
Try this first.
Here is one thing I want kept on track: [project]. The sources are [emails, doc, channel]. Tell me only when a deadline is at risk or you need a decision, and show me drafts before anything is sent.
Approvals in practice: what early users ran into
OpenAI's Control your dot guide says that before an action could affect your accounts or share information, an automatic review decides whether it goes ahead, needs your approval, or is a step you must do yourself. What early users actually ran into falls into five patterns.
1) Drafts wait for an explicit yes
Dan McAteer's Dot picked up a freelance invoice he still had to send, drafted it from the email thread, and sent it as a PDF only after he approved it. The Korean writer behind AI Never Sleep asked their Dot, 네슬, to write a price-quote email with a site link but not send it; the Dot opened its reply by saying it had only drafted the email, and after "좋아 이렇게 보내줘" ("good, send it like this") it sent exactly the approved version. Ren Gaku's four-day note.com write-up counts at least nine replies sent the same way, each by Ren or on a clear request. OpenAI's own wording says the same: asking for drafts doesn't give permission to send them.
2) Overnight work stops at the risky step
Kurt Bell went to bed with a bug and a data issue and woke up to both fixed and tested, plus a product enhancement his Dot had found and tested but not deployed without his approval, which he called the most impressive part. koh_kira's bedtime instruction on Threads sets the same line in words: keep going only within the requests and approvals already given, park anything waiting for his input and move on, and report what's done, what's left and why things are on hold after eight hours.
3) An approval may not reach split-off tasks
Ren's Dot, シエル, splits work into separate child tasks, and Ren's approval didn't always travel with them. On October 2 the same approval was requested more than ten times, and one meeting-room booking had to be asked for four times and took about seven hours. Writing the permission once, directly into the child task, got it moving. Alex Finn's open project-manager skill for Dots bakes in the same caution: never assume a child task inherits tools or skills, silence is not approval, and managing a project doesn't by itself authorize publishing, deploying, merging or spending.
4) The cost is interruptions
Desmond Dixon ran his business through a Dot for 48 hours and listed manual approvals among the four things that got in his way before canceling for now. John Voorhees at MacStories (case 2 above) found per-task permission requests natural. The difference is mostly setup: an approval you hit every day is a candidate for a custom rule.
5) The automatic review can overrule a yes you already gave
Cameron64 runs a personal project called Garden from a Dot on Android, with the Dot's cloud computer passing parts of the work to Anthropic's Claude Opus 5.5 and Haiku 5.5. When the Dot asked whether it could share Garden's source, designs, test results and deployment details with those agents, credentials and unrelated personal data excluded, Cameron64 answered "yes", and later approved Haiku in writing as well. The automatic review still rejected the calls on 10 October as sending private project material to api.anthropic.com without explicit authorization, including calls that only read the output of a process already running. After three refusals in a row the review interrupts the turn ("rejected too many approval requests for this turn"), and two fresh yeses that evening each ended in another interruption. Hours went into confirmations; the work was saved at a checkpoint and not lost. dean1012 saw a quieter version: a standing permission for recurring read-only checks of one website in the cloud browser worked at first, then was refused now and then with "no trusted user authorization for this browser access or task is present", even though nobody had withdrawn it. Pointing the Dot back to the original permission sometimes got the check through, and sometimes didn't.
Two details from OpenAI's own pages are worth knowing. The Help Center's Getting started article calls one custom-rule option "Take action if pre-approved", while the Learn guide calls it "Take action when you say so"; both mean the Dot proceeds when you asked for that action in your prompt and asks otherwise. And Pause stops only your Dot's main task: delegated tasks have to be stopped in Activity and recurring ones disabled in Scheduled.
Copyable prompts
Each prompt says which reports it draws on and which failure it is meant to avoid. Replace the brackets with your own details.
Draft first, send on "send it" (from McAteer, 네슬 and Ren; avoids a message going out early, or a draft sitting unsent because nobody said so):
For replies to [people or channel], always show me the draft first and say plainly that it hasn't been sent. When I answer "send it", send exactly that version without edits and tell me it went out. Anything I haven't said "send it" to stays a draft.
Overnight, with a stop line (from Kurt Bell, koh_kira and Ren's eight-hour run; avoids an unreviewed deploy or purchase at night, and a whole night lost waiting on one question):
I'm going offline for [hours]. Work only within what I've already asked for and approved. Fix, test and prepare, but don't deploy, publish, send, buy or delete. If something needs my yes, park it, keep going on the rest, and put it at the top of your morning report with what you'd do and why.
Carry my approval into split tasks (from Ren's repeat approvals and Alex Finn's skill; avoids being asked the same thing ten times and duplicate bookings):
If you split this into separate tasks, copy my approval word for word into each task's instructions: [action, scope, limit]. If a task still asks me for the same approval, stop that task and tell me instead of retrying, so nothing gets booked or sent twice.
Turn repeat approvals into rules (from Desmond Dixon, John Voorhees and OpenAI's four rule options; avoids approval fatigue without giving up control of risky actions):
List the actions you asked me to approve in the last seven days and how often. For each one, suggest a custom rule: take action without asking, take action when I say so, ask before taking action, or hand off to me. Keep messages to people outside [company], payments and deletions on ask first or hand off. Don't change any rule until I confirm.
When the review keeps saying no (from Cameron64 and dean1012; avoids hours of repeated yeses and an interrupted turn):
If the automatic review refuses an action I've already approved, don't retry it more than once. Stop that step, keep going on the rest, and tell me the exact refusal text, the action, where the data would go, and the approval of mine you relied on. If the work needs to send my files or data to another company's service, ask me for that as a separate approval naming the service and what goes to it.
For approval habits inside specific jobs, see email triage, invoices and money, project follow-up and steering Codex; app-level permissions are on Integrations.
Custom rules people actually wrote, and where they didn't hold
OpenAI's Control your dot guide describes custom rules as optional, one action at a time: open Settings > Personalization > Custom rules, describe the action, and pick Take action without asking, Take action when you say so, Ask before taking action or Hand off to you. The same page says a rule is an instruction the Dot tries to follow and can get wrong. A rule doesn't connect an app or a computer, doesn't remove required confirmations such as approval to use a saved login, and sits apart from Plugin permissions, which control app actions separately. If a work workspace turns custom rules off, saved rules stop applying. Below are four setups people published, followed by the reports where a rule didn't do what its author expected.
Four setups from people who wrote theirs down
One Gmail job split into six actions (入江慎吾, note.com, 6 October)
In a note.com write-up of the setup they run every day, 入江慎吾 gives the Dot one email job and writes a separate rule for each action inside it. Reading, searching and summarizing mail, and saving reply drafts: take action without asking. Labeling and archiving: ask first, then move to "take action when you say so" once the sorting has proved right. Deleting: ask first. Sending a reply: ask first, and in practice the Dot saves a draft that 入江 checks and sends by hand. New recipients, forwarding and sharing attachments: hand off, as a guard against leaks and against instructions planted inside an email. The daily 5:05 p.m. cleanup then needs one word: on one day the Dot asked whether to archive 20 of the remaining 38 emails and move 18 to Trash, and "ok" was the whole reply. Payment, security, delivery and contract notices are never deleted automatically, deleting means Trash only, and the very first run sorted a week of mail without archiving or deleting anything, to correct the categories first. 入江 adds that the Dot suggests rules of its own and points out where two rules contradict each other.
Three questions, three tiers (しろ, note.com, 7 October)
After a week, しろ sorted tasks with three questions: if it goes wrong, can I redo it alone? Does it touch someone else's time or feelings? Is the final call mine? Research and drafts passed all three and got "without asking"; scheduling and internal messages got "ask before"; anything sent outside the company, plus money and permissions, got "hand off". The reason was concrete. A drafted message to a client was polite and correct, but a little more formal than that relationship had become over time, and only しろ could hear the difference. The advice that follows is to start strict and widen one rule at a time.
A rule to protect usage, not just safety (@thelaggingway, X, 1 October)
Because work a Dot sends to Codex counts against Codex limits, @thelaggingway shared a rule set to Ask before taking action that covers "any invocation of a Codex Cloud agent, including creating, launching, resuming, restarting, or sending work to one", requires a separate approval "even if I directly request the task", asks the Dot to explain why it is needed and what it will use, and excludes "ordinary native sub-agents".
Read-only through plugin permissions, Drive left out (tadashi-aikawa, nocturne blog, 4 October)
On the first day, tadashi-aikawa set up a Dot named みみきち in ChatGPT on the web, to reach from the phone when away from the MacBook that runs a home-built team of AI agents. It got Gmail and Google Calendar but not Google Drive, which holds things tadashi-aikawa doesn't want read and nothing that would make the agent more useful. Connecting Gmail and Calendar couldn't be narrowed to reading alone, so instead of a custom rule tadashi-aikawa set the plugin permission to allow read-only tools: reads run without asking, and anything that would create a calendar event or send mail stops for a confirmation. Access to the local MacBook was held back for now. The only routine so far is a news digest three times a day. tadashi-aikawa adds that this setup trusts OpenAI's checks to work, so truly risky information and actions still deserve care. OpenAI's Connect computers and apps page describes the same option: your existing app permissions apply to the Dot, so it can read email without being allowed to send it.
Rules that only stop a job at one step are on the pages where those jobs are covered: an invoice rule that ends at the attached draft (invoices and money), a one-line "don't send email" (Slack), and a fresh Dot whose custom-rule screen defaulted to asking first (email triage).
Where a rule didn't hold
- Rules that fire unevenly. On 10 October @argbeknwn wrote that custom rules "fire inconsistently" on a day when Slack replies also needed an @mention and the Dot was slow, and that managing it was taking more time than it saved.
- A saved rule that didn't reach the step. A rule set to "Take action without asking" for merging after CI was saved, yet the merge was refused twice; the block came from wording in the pull request itself (GitHub, DmitrL-dev).
- Rules written too broadly. A Hacker News user's broad "act without asking" rules were turned down by the permissions screen (travel planner, neom).
What these reports add up to
- Use the permission layer for the hard limits. tadashi-aikawa kept Drive unconnected and made Gmail and Calendar read-only at the plugin level. A rule is an instruction the Dot tries to follow; a permission it doesn't have is a firmer limit.
- Write rules for actions, not apps. Every working set above names a verb and an object: archive these, send to new people, launch a Codex cloud task. The one broad set above was rejected.
- Dry-run before you loosen. 入江 sorted a week of mail with nothing executed, and しろ started on the strict end. A rule moves to "without asking" only after the results have matched for a while.
- Rules are also a cost control. A rule can stop expensive routes (Codex cloud tasks) as well as risky ones.
- When a rule misfires, check what actually decided. A plugin permission, a required confirmation or the task's own wording can override what the rule seems to say. Read the refusal text in Activity, restate the boundary in the conversation, and ask the Dot which rule it applied.
A copyable starter rule set
Each line is one rule: the action to describe, then the behavior to pick. Adjust the apps and contacts to your own.
- Reading, searching and summarizing my email, calendar and Drive, and saving drafts without sending them → Take action without asking
- Archiving or labeling email from senders I've listed as safe → Ask before taking action for the first week, then Take action when you say so
- Moving email to Trash, never deleting permanently and never touching payment, security, delivery or contract notices → Ask before taking action
- Sending any email or message to someone outside [company] → Ask before taking action
- Forwarding mail, sharing a file, or writing to someone I have never emailed → Hand off to you
- Creating, resuming or sending work to a Codex cloud task, with the reason and expected usage → Ask before taking action
Then ask once a week:
List every custom rule you're following, the exact actions each one covers, any two that contradict each other, and any action you asked me to approve more than three times this week that no rule covers. Suggest changes, but don't change a rule until I confirm.
What model is my Dot on? One user's DevTools reading
OpenAI's product copy for Dots names GPT-6 Astra. On 16 October rabbitdaisuke reported that Chrome DevTools WebSocket responses for their Dot showed GPT-6.1 Sol Low, with background Work disabled, and that assigned tasks stalled for more than three hours while plugin/GitHub interactions continued. The post title calls this an A/B test for $100 users; the reading comes from that one account, and OpenAI's pages still name GPT-6 Astra. The same post is covered on pricing & usage.
Related
- ChatGPT Dots FAQ: common questions answered in a few lines each
- Getting started: your first responsibility after setup
- "Your dot is on a break": what the pause means and how long it lasted for real users
- How to get ChatGPT Dots: plans, regions, login and what to do when Dots is missing
- Use cases: what people handed to their Dots, grouped by job
- Features: calls, cloud computer, apps and channels in more detail
- Availability and EU & UK availability
Copy this task brief
What ChatGPT Dots (OpenAI Dots) is, who can get it and how to set it up, where it gets stuck and what fixed it, what it can't do yet, and one message to start with, followed by what people actually handed over.Download Markdown ↓