GPT DotFIELD GUIDE
Home/Workflows
Workflows

How to use a ChatGPT Dot as a project manager: setup, status boards and 19 real project follow-ups (OpenAI Dots)

Yes, a ChatGPT Dot (OpenAI Dots) can follow a project for you. Start with the short version: what it does well, six setup steps, where it gets stuck and what worked, and one message to start with. Then 19 real project follow-ups, from an App Store review watched through to approval to boards the Dot keeps current, plus what to tell it before an overnight run.

GPT Dot · Updated

The short version: what a Dot can do as your project manager, how to set it up, where it gets stuck

What it does well, and where it falls short

Four things come through in the cases below. A Dot keeps a running list of what's open and brings back only what needs you: わいへい's Dot watched an App Store review through to approval and asked before announcing it, and Tiffany Kyazze's reminds her what she has started. It notices what slipped or doesn't match without being asked, like the candidate Ted's Dot found booked on two dates, the visa application Igor Pogany's flagged after eight quiet days, and the three different end times Naufaldi Rafif Satriya's spotted. It keeps a board you can both read, as 坂本匡史, AiHank, @srsuemori, Dejan Simic and BumbleV AI each set up. And it keeps work moving between your decisions: Creer's two projects, Alex Finn's milestone-by-milestone skill, and Kurt Bell's overnight fixes that stopped short of deploying.

The weak spots are just as consistent. A Dot can stop a whole project over one open item, ask permission too often at first, let an outside board fall behind until it next reads it, and lose a night's work to a usage pause, a cloud workspace reset or a computer that went to sleep.

Set it up in six steps

  1. Create or open your Dot in the ChatGPT mobile app, the desktop app or desktop web (getting started). If it should work on your own computer, allow access from the desktop app on that computer (Getting started with your dot).
  2. Connect only the places where the project actually lives: the plan document, the task board (Asana, Linear, Notion, Planner or a Google Sheet), the email threads and the chat channel. Setup guides are on the integrations page. Permission to read an app may not include permission to change it or send messages, so check each app's allowed actions.
  3. Describe the project in one message: the finish line and date, the owners, which source wins when two disagree, where updates go, and what the Dot may do without asking.
  4. Ask for one baseline tracker first, open its source links, and correct owners and dates.
  5. Decide how it should check back. You don't need a fixed schedule for every check: OpenAI's Meet dots guide says you tell your Dot what to keep current and which changes deserve your attention, and it decides when to follow up. When you do want a fixed check, give a time zone and an end date and have the Dot confirm it. OpenAI's own example is a planning channel checked each weekday at 9 AM Central for four weeks, with a message only when a deadline is at risk or a decision is needed.
  6. Keep an eye on what it is carrying. In the desktop app, its profile lists work under In progress, Scheduled and Completed. To change a scheduled task, select it in Recent activity or open the Scheduled section, where you can change its repeat schedule, time or completion notifications.

A useful update looks like this: a design approval slips two days, and the Dot shows the old and new date, the milestone it blocks, who can unblock it, and a draft follow-up for you to review. Unchanged items stay in the tracker instead of in your messages. To practise first, TakeAICourse's project-tracking exercise walks through a small workshop plan with owners, deadlines and one dependency.

Where it gets stuck, and what worked

  1. It stops the whole project over one open item. Creer's Dot admitted it had done exactly that. Tell it to keep going on anything that doesn't depend on you, and to stop only for technical judgement, final approval or permission changes.
  2. It asks permission too often. わいへい and his Dot edited its settings to say what needs no approval. Name what it may do freely (reading files, running tests in named folders) and what always needs you (publishing, submitting, announcing, paying). On mobile, Customize, then Custom rules, lets you set an action to Take action without asking, Ask before taking action or Hand off to you.
  3. The board and the Dot drift apart. 坂本匡史 found that the Dot sees a change on the board the next time it reads it, not the moment you press a button. Ask it to read the board at the start of every check and tell you what it changed, keep "scheduled on the board" apart from a schedule the Dot has confirmed, and remember that finishing one run of a recurring task doesn't stop the series.
  4. "Done" hides what you still have to test. BumbleV AI split the dashboard into needs my approval, needs my testing, completed and in progress, and asks for test steps with the address, the buttons and the expected result. A visible last-update time tells you how fresh the board is.
  5. Work stops while nobody is awake. c2chq's overnight job ended "on a break", bshane54 waited about 13.5 hours, nervouna lost a cloud folder, and Creer's office PC went to sleep. Have the Dot save progress after each step to Google Drive or GitHub, and keep a connected computer online with the ChatGPT app open, or keep the work on the Dot's own cloud computer.
  6. It can't reach a mounted drive. Creer's Work kept going back to a Google Drive for desktop mount it couldn't read until the Dot sent it to the drawings saved locally. Point it at local copies, or connect Google Drive as an app.
  7. You approved something and it is still blocked. Open the affected task and look for a separate sign-in, connection or approval request. Asking for a draft doesn't authorise sending it.
  8. Something went out that you didn't expect. Rohan Attravanam's Dot texted a prospect overnight. If nothing should leave without you, say "draft it, don't send it" in the request.

What it can do, and what it can't yet

It can follow several projects at once, use background agents and start new Codex threads, and pick up new information without restarting the work (Meet dots). With local access on, it can create Work or Codex tasks on your computer, and it can create cloud tasks in a Codex cloud environment you have already set up. It can read and update a board you connect and draft follow-ups for you to approve.

It can't yet keep an outside board in live sync: it sees changes when it next reads the board. Each task has its own conversation and doesn't automatically receive everything you have told your Dot, so give corrections in the affected task. Work that needs your computer stops when that computer is offline.

One message to start with

You're the project manager for [project] until [end date]. The sources are [plan], [board], [email threads] and [channel]; if they disagree, [primary source] wins and you show me the other one. Start with a baseline tracker (milestone, owner, due date, status, next step, source link) and a list of what you still need from me. After that, tell me only when a milestone slips, a dependency blocks the next step, or a decision needs me. If one item is waiting on me, keep going on everything that doesn't depend on it. Draft any follow-up messages for me to review; don't send them.

Quick answers

Can a ChatGPT Dot be my project manager?

Yes, for projects that run for days or weeks. わいへい's Dot submitted his iOS app for App Store review, checked back half a day later, noticed the approval on its own and asked before announcing it; it also keeps a kanban board updated every day. Alex Finn runs a published project-manager skill that moves a build milestone by milestone, and Tiffany Kyazze uses hers to remember what she has started and what still needs her.

How do I set it up?

Describe the project in one message: the finish line and date, the owners, which source wins when two disagree, and where updates go. Ask for a baseline tracker first, then a recurring check with a time zone and an end date, and have the Dot confirm it (Meet dots). In the desktop app, its profile shows what it is carrying under In progress, Scheduled and Completed (Getting started with your dot).

Which boards and apps can it follow?

Whatever you connect. Ted's Dot read Asana, Google Drive and Gmail together; Pat Simmons's Dot reached his Linear-style Code Queue board through an MCP server; BoomBigNose saved action items to Google Sheets; 坂本匡史 and AiHank had the Dot keep a task app or an HTML page current. For code projects, see Dots + GitHub; for shared files, Dots + Google Drive; for updates, Slack or Microsoft Teams.

Will it notice what slipped?

Often, without being asked. Ted's Dot found the same candidate booked on two dates in a hiring sheet; Igor Pogany's flagged a visa application with no update in eight days; Naufaldi Rafif Satriya's spotted an end time that differed between his calendar, an email and the rundown; Andrey's pointed out a 1 hour 15 minute connection right after he bought tickets.

Why did it stop, and how do I keep it going?

Two common reasons. One is a pending decision: Creer's Dot admitted it had treated one open item as a reason to stop a whole project, so tell it to keep going on anything that doesn't depend on you. The other is a usage pause: c2chq's overnight job ended with "on a break" (see Your dot is on a break). Kurt Bell's Dot fixed and tested overnight but stopped short of deploying, which is the line you want.

Which plan and apps do I need?

A Dot comes with Pro ($100, $200 or $500 tiers) outside the EEA, UK and Switzerland, with Business Premium in any supported region, and with Enterprise once an admin switches it on (How to get ChatGPT Dots; in Europe, see Dots in the EU and UK). The full list of what a Dot can use is on the features page.

Real examples, by job

Make the Dot the project manager

Alex Finn, "Project Manager" skill (YouTube, 1 October 2026). In his video, Alex shows his own Dot running a custom skill he published as project-manager-skill. The skill takes a software, app, or game project from an agreed vision through small milestones: it hands each milestone to an existing implementation thread, tests the exact build that comes back, runs fix-and-retest loops, and keeps progress, decisions, screenshots, and QA evidence in a private project Space. Publishing, deploying, merging, and spending each need their own approval. You install it by uploading the SKILL.md to ChatGPT and invoking @project-manager.

Tiffany Kyazze, @techtiff.ai (Threads, 30 September 2026). Tiffany turned her Dot into a project manager: it keeps track of what she's working on, makes sure she doesn't forget what she has already started, and tells her what still needs her attention. She describes it as a partner that knows what's already handled.

Aria, @ariapramesi (Threads, 1 October 2026). Aria says his Dot is now his PM "for multiple things", running a modified version of automation code from Kun Chen's repository plus content collection. In the replies, as summarised by Made with Dots, he adds that it runs his fleet of websites hands-off.

わいへい, @yhei_hei (note, 6 October, and Threads, 8 October 2026, in Japanese). わいへい, an indie developer, runs several projects at once: an iOS app called カルマロック (Karma Lock), a Unity game, a Roblox game, and a podcast and newsletter. Before Dots he was the project manager himself, giving Claude Code and Codex instructions one session at a time. In his note he says his only request was to move Karma Lock's development and marketing along "nicely". The Dot submitted the app for App Store review, set itself a check for half a day later, noticed on its own that the app had been approved, and asked for his approval to announce it on X. When he lost track of what was in progress, he asked for a kanban dashboard; the Dot built one and now updates it every day, and anything blocked comes to him as "this is blocked, I need your decision". On Threads he sums it up: he sets the schedule and the milestones, and the Dot picks up what he forgot and asks for approval where it's stuck. Not everything went smoothly. At first it asked permission too often, so he and the Dot edited its settings file together to say what needs no approval. When his Mac mini locked, the Dot complained that it couldn't see the screen, so he changed the lock settings. It also tends to hand work to Codex first, while he prefers Claude Code for creative game work, which the Dot now calls with claude -p. He thinks it pays off for people juggling several projects, and much less for someone with only one.

Copyable ask:

You're the project manager for [project] until [end date]. The sources are [plan], [board], [email threads], and [channel]; if they disagree, [primary source] wins and you show me the other one. Keep a tracker with milestone, owner, due date, status, next step, and source link. Tell me only when a milestone slips, a dependency blocks the next step, or a decision needs me. Start with a baseline tracker and a list of what you still need from me.

Copyable ask (from わいへい's report and Creer's second day below):

You're the project manager for [project A], [project B] and [project C]. I set the milestones and dates; you keep the work moving between them. Keep a kanban board with columns for to do, in progress, blocked and done, and update it every morning. When something is blocked on me, tell me in one line exactly what you need me to decide. If one item is waiting on me, keep going on anything in the same project that doesn't depend on it: waiting on one decision doesn't mean the project stops. Don't ask me before reading files or running tests in [folders]; do ask before anything is published, submitted, announced or paid for.

Keep delegated work visible

@srsuemori, labour and social insurance consultant in Ehime, Japan (Instagram, 6 October 2026). He asked his Dot to build a dashboard app so the work he has delegated to it is readable at a glance by people. The screenshot he posted, titled "the current state of delegated work", shows 0 tasks in progress, 3 scheduled, 5 waiting on confirmation or materials, and 5 recently completed. He says the to-dos, deadlines, and who has the ball are all visible on one screen. (Caption translated via Made with Dots.)

Pat Simmons and "László" (AI for Mortals, 1 October 2026). Pat's AI agents track client work in Code Queue, a Linear-style board he built with an MCP server and almost never looks at. In his review, László found Code Queue on his Mac, checked the MCP connection, read the API, and scheduled a status summary in Slack every weekday at 9 PM. On a call it told him 137 issues were in progress and 122 were in review, most of them stale. He asked it to propose which to close, send the list in Slack, and close them once he approved; about a minute after he hung up, a "Code Queue cleanup" message arrived.

Adrian Murray, @adriancmurray (Threads, 5 October 2026). Shortly after his five-hour Claude session restarted, his Dot sent him a status read-out on one of his projects without being asked.

Copyable ask:

Every weekday at [time] [time zone] until [end date], send me a status summary for [project] in [Slack DM or ChatGPT]: newly overdue items, items waiting on someone else, items due in 7 days with no next step, and anything stale for more than [N] days. If you think items should be closed, send the list and wait for my yes. Confirm the schedule.

Give the Dot a board you can both read

坂本匡史 (Masashi Sakamoto), AI solopreneur in Fukuoka (note, 4–8 October 2026, in Japanese). A week into handing work to a Dot, Sakamoto's question had become "what is it actually carrying right now?" Asking works, but asking for the list after every new job got tedious. In an 8 October post, Sakamoto describes TaskDock, a task app built with ChatGPT Work after weighing Asana, ClickUp and Notion: tasks grouped by project, a list and a Kanban board, Markdown details with images, and recurring tasks with weekday, time and time zone. The Dot then got operating rules: register every job it is asked to do, update the status as the work moves, mark anything that needs Sakamoto's check, write the reason and the next step for anything that is waiting, and look first so nothing is added twice. After a connection check, the Dot filled TaskDock with the work it had been tracking in conversation, sorted by project. When Sakamoto marked a task done in the app, the Dot noticed the next time it read TaskDock and closed the matching work on its side. Sakamoto is precise about the limits: this is not an automatic sync, and the Dot isn't told the moment the button is pressed; a schedule entered in TaskDock is labelled separately from "the Dot's schedule confirmed"; and finishing one run of a recurring task is not the same as stopping the series.

A post from 7 October shows the kind of follow-up that made the board worth having. When new email continued earlier work, the Dot said so in Slack: a cloud bandwidth overage came to $5 and was already paid, while the gap between the usage CSV and the billed quantity was still unreconciled, which it stated plainly. A meeting that had been entered for 3–4 AM got a question ("is this 3–4 PM?") rather than a silent change.

AiHank, @hank.huang0516 (Threads, in Traditional Chinese, early October 2026). The Dot often seemed to vanish, and it was hard to tell what it had done all day, so AiHank set up a Gantt chart to maintain together with it. The Dot deploys and keeps updating an HTML workspace page that lists every task mentioned in their chats. Under each task: the decisions the manager needs to make, a detailed status (goal, progress, finished items, blockers and items not yet accepted, next step) and the manager's comments. Anything stuck or waiting on a decision is flagged. Rather than explain a general method each time, AiHank lets the Dot propose the fix; the manager approves it or leaves a comment to change direction. AiHank compares it to an earlier multi-agent Kanban setup.

Creer, a civil-engineering design firm (note, 1 October 2026, in Japanese). Creer handed one Dot two development projects; the first day is on the cloud computer page. In a follow-up that evening (full version on Creer's site), Creer tested whether the Dot could also take over the "continue" messages. Each Work reached a stopping point every 10 to 20 minutes. With no new instructions, the Dot collected the results; when Creer opened it, there were 3 notifications and a report that kept the two projects' progress apart and stopped wherever a decision or extra permission was needed. Stalled Works then started again after Creer had only opened the report, without typing "continue", though Creer notes that the order of events on screen doesn't prove opening was the trigger. Creer then told the Dot to keep going on anything that needs no decision without waiting for notifications to be read, and to stop only for technical judgements, final approval or permission changes. The Dot confirmed, and both Works moved again. Other stops were mixed in: a path error ("invalid turn/start params: AbsolutePathBuf deserialized without a base path") when a new Work started, approvals passed from the Dot to a Work not counting the way Creer's own did, and a message the Dot asked Creer to relay to a Work that couldn't be copied out (right-click opened an emoji menu, and a screenshot failed too). While one project was blocked, the other kept moving. Watching the usage meter, Creer saw it drop about as fast as it does with two Works alone. Whether the Dot keeps going all night with nobody opening notifications was not yet tested.

Creer, the next day (Creer's site, published 5 October 2026, in Japanese). That night the test stopped for a simpler reason: the office PC went to sleep, and the work that used its local environment stopped with it. Creer's third report follows the two projects from about 7:30 a.m. to 3:58 p.m. on 2 October. Before a desktop-app update, the Dot saved each Work's state and resume point, then carried on from there after the update. When one Work marked itself "final verification complete" after a screen clean-up, Creer told the Dot that standing still was the biggest risk. The Dot replied that nothing showed the project had run out of work, picked the next item from the existing development plan, and later that morning all 189 tests passed. On the other project, a Work went back to a shared Google Drive mount it already knew it couldn't read; the Dot cut that path and sent it back to the drawings already saved locally. Around 11 a.m. a Work reached a reading of the drawings it couldn't settle and sent Creer a single question, while it kept working on parts that didn't depend on the answer. In the afternoon, with Creer's go-ahead, a parent Work split two independent checks between subagents and merged the results at 12:42, keeping "design intent not confirmed" in the record. At 3:58 p.m. both projects had stopped again. Asked why, the Dot said its own progress management had fallen short: it had treated one pending item in each project as a reason to stop the whole project, although both still had work that didn't depend on it. Creer hadn't yet confirmed whether it restarted them on its own.

Dejan Simic, Adaptive Technologies (blog, 7 October 2026). Dejan gave his Dot the Notion sprint board he already used and added a column for the agents. His standing instruction has two parts: check every new card in To do and tell him if an agent couldn't do the work from the description alone, and once he moves a card into With agents, start on it, hand it to Codex and move it to In progress when a draft is ready. The Dot checks the board every 15 minutes between 8am and 6pm, so moving a card is the whole hand-off. More of his setup is on the Notion page.

BumbleV AI (Built with BumbleV AI, 14 October 2026). Working on several projects at once, BumbleV AI kept rereading conversations to tell a change that had been implemented from one they had actually tested. Their Dot, Wondubu, still takes requests in conversation, but progress now lands in Daily Worklog, a dashboard built with ChatGPT Sites that groups seven projects under four headings: needs my approval, needs my testing, completed and in progress. The screenshot they shared, last updated 14 October at 19:20 (UTC−06:00), shows one approval, six tests (three still waiting on preparation), ten completed items and two in progress. For anything that needs a hands-on check, they ask for the address or version, the buttons to press, the steps in order and the expected result at each step. They also point out that the visible update time matters, because opening the dashboard doesn't mean every detail was checked a moment ago.

Copyable ask (from 坂本匡史's, AiHank's, Creer's and BumbleV AI's setups):

Keep [task board or shared page] as the list of everything you're doing for me. Add each job I give you, update its status as it moves, and mark anything that needs my check. For anything waiting, write what it's waiting on and the next step. Look before adding so nothing appears twice. When I mark something done there, treat it as done the next time you read the board, and tell me which items you changed. Keep "scheduled on the board" separate from "schedule confirmed in Scheduled". Split what needs my approval from what needs me to test it, and for each test give me the address, the steps and the expected result. Show when the board was last updated. If a task needs no decision from me, keep going without waiting for me to open your notifications.

Catch what slipped or doesn't match

Ted, @tedzted (Threads, 30 September 2026). Ted gave his Dot access to Asana, Google Drive, and Gmail. In his post, written in Chinese, he says that while they were talking about something else, it noticed an interview in Asana for the next day, searched his Google files, and found two rows in the recruitment spreadsheet that looked like the same candidate with the same 3 PM slot on different dates. HR confirmed it was one person and one form hadn't been updated. The same Dot also pulled a Wi-Fi renewal date and receipt-lottery prizes to claim before 6 October out of his email.

Igor Pogany, The AI Advantage (YouTube, 30 September 2026). In his first 24 hours with a Dot, the Dot used his connected plugins overnight and at 8 AM raised two things on its own: an upcoming US trip, and an O-1 visa application with no status update in eight days, with a suggestion to follow up with his assistant. (Summary via Made with Dots.)

Naufaldi Rafif Satriya and "Sherly" (Threads, 30 September 2026). On day one, right after setup, his Dot flagged that an event's end time differed between his calendar, an email, and the event rundown. Next he wants it to turn emails into action items and collect issue context before handing work to Codex; he still reviews emails and posts himself.

Andrey, @atars (Threads, in Russian, 12 October 2026). On his first day with a Dot, right after he bought plane tickets, it pointed out without being asked that one ticket included checked baggage and the other didn't, and that the connection was only 1 hour 15 minutes. Then it offered to deal with a matter that had been hanging for a week: it looked into it, prepared the replies and sent them. (The tickets are also on the travel page.)

Turn meeting notes into action items

BoomBigNose (YouTube, Thai, 2 October 2026). In his walkthrough, he attaches a meeting-notes file and asks his Dot to separate confirmed decisions from proposals, list action items with their sources, and save the result to Google Sheets. A second case tracks a project's deadlines and the permissions each step needs. His advice is to start with one task and check the result before adding more; in the video, the Google Sheets login was still a step he had to complete.

AI 서대표, @AISeoceo (YouTube, Korean). While on holiday abroad he couldn't attend the next day's meeting, so he only asked his Dot to work out the context and pass on what the team needed. In his three-day review, the Dot opened a Slack thread asking the team for what was needed, then checked the meeting time and the decision that came out of it, reported back, and saved it to memory.

Copyable ask:

Here are the notes from [meeting]. Separate what was decided from what was only proposed. List each action item with owner, due date, and the line in the notes it comes from, and add them to [sheet or board]. Leave owner or date blank if the notes don't say, and list those gaps for me. Check back on [date] and tell me which items moved.

Pick up a project you were already working on

Sungyun Kim, @sykimesq (Threads, Korean, 30 September 2026). After he named his Dot and connected his local machine, it continued the development project he had been working on. When it offered to write a prompt, he told it to do the work itself; it checked the local connection and started running Codex. It also went through his Google Calendar, found overlapping events, and asked whether to sort them out. He says it feels like a personal assistant and a project manager at once. For more on handing coding work to Codex this way, see Codex orchestration.

Overnight runs: what to tell your Dot before you sleep

Handing a Dot a job at bedtime is one of the most common things early users try. The reports below come from X, Threads, the OpenAI Community and GitHub, grouped by what the person said before going to sleep, where the Dot stopped, and what went wrong while nobody was awake to step in.

What the handover said

キラ, @koh_kira (Threads, Japanese, 2 October). Before an eight-hour sleep, キラ posted the exact message that gets handed over each night: continue each session's remaining tasks within the requests and approvals already given, restart sessions that stopped, switch to a different approach when stuck, set aside anything waiting on input and keep working on other things, send notifications only for important changes, and after eight hours report what was finished, what is left and why anything was put on hold.

Himanshu Jha (X, 1 October). The bedtime ask was "Surprise me by morning", with three rules: spend nothing, leave existing projects alone, and keep going all night. By 9 a.m. the Dot had published three new GitHub projects, opened an audited draft PR, passed 193 targeted tests and sent three DevRel applications. Himanshu Jha handled the sign-ins and permission approvals.

Eduardo, @martiano (X, 2 October). Before going out for a walk and dinner, Eduardo left the Dot merging about 15 PRs, waiting for CI and resolving conflicts on Eduardo's own computer, then building the app, updating packages, and sending a phone notification when everything was ready to take over.

Ren Gaku, @rengakuai (Threads, Japanese). After four days using the Dot "Ciel" as a work secretary, Ren Gaku handed it eight hours of autonomous work overnight, and flagged four things to watch: usage limits, memory, approvals and where the data ends up.

Where the Dot stopped, and where it kept going

Kurt Bell, @softypapa (X, 30 September). Kurt Bell went to bed with a software bug and a data issue and woke up to both fixed, tested and written up in a morning report. The Dot also found and tested a possible product enhancement, then stopped short of deploying it without approval. Kurt Bell called that last part the most impressive.

Rohan Attravanam (Threads). Rohan Attravanam gave the Dot a list overnight. By morning it had sent a follow-up text to one prospect on its own and was partway through emailing images to another. Rohan Attravanam was happy with that, but it shows that an overnight list can include sending, unless you say otherwise.

What went wrong with nobody awake

c2chq (OpenAI Community, 5 October). Before bed, c2chq asked the Dot to find jobs that matched their qualifications and put them in a Google Sheets tracker. It never finished and never sent an update. In the morning the Dot answered that it was "on a break" after hitting an abuse-prevention limit. In the same thread, OpenAI Support asked affected users to email support@openai.com with a screenshot, the task, when it stopped and their time zone.

bshane54 (same thread, 7–8 October). bshane54 logged the clock on the same "on a break" message: it started at 7:43 a.m. PDT and the Dot came back at 9:15 p.m., about 13.5 hours later.

nervouna (openai/codex GitHub issue #50577, 3 October). A project folder under /workspace/shared on the Dot's cloud computer held 230 files at about 20:25 UTC and was gone by 08:46 UTC the next morning. Part of the work had to be rebuilt.

Himanshu Jha, again. Partway through the same night, the cloud workspace reset. The Dot rebuilt the lost files, reran the checks and saved everything again.

Андрей, @andrey_zwart (Threads, Russian, 30 September). The Dot found and removed unneeded subscriptions, but also closed every app on the connected Mac several times and put the Mac to sleep, which is the opposite of what an overnight job on your own computer needs.

What these reports have in common

Three of these reports describe finished work waiting in the morning (キラ, Himanshu Jha, Kurt Bell), and Ren Gaku says eight hours of autonomous work went through. The good nights share three things: the request named what was off limits (spending money, existing projects, deploying), it said what to do when the Dot got stuck or needed input, and it asked for a report with done, remaining and on-hold items. The bad nights had nothing to do with the prompt. A usage pause, lost cloud files or a local computer that went to sleep stopped the work, and in c2chq's case nothing partial was left to look at. That is why the prompts below ask for progress to be saved somewhere you can see it while the job runs, not only at the end. If you are wondering how the first-month usage allowance is described, see the updates page. For how approvals behave in general, see Approvals in practice.

Overnight asks you can copy

Each one draws on at least two of the reports above and names the problem it is meant to avoid. Replace the parts in brackets.

Bedtime handover (from キラ, Himanshu Jha and Kurt Bell; avoids a Dot that oversteps or stalls silently):

I'm going to sleep for [8] hours. Keep working on [project or task list] within what I've already asked for and approved. Don't spend money, don't touch [projects or folders to leave alone], and don't deploy, merge or publish anything; if you find something worth shipping, test it and leave it for me. If you get stuck, try a different approach. If something needs my input, put it aside and move on. Only message me about [what counts as important]. When I wake up, give me a report: what's done, what's left, and what's on hold and why.

Save as you go (from nervouna and Himanshu Jha; avoids losing a night's work when the cloud workspace resets):

As you work tonight, save the files and a short progress note to [Google Drive folder or GitHub repo] after each finished step, not only at the end. In the morning report, list where each output is saved. If any files are missing when you come back to them, tell me what you rebuilt and what you couldn't recover.

Leave something to look at (from c2chq and bshane54; avoids waking up to nothing after a usage pause):

Put results into [Google Sheet] as you find them, one row at a time, so the sheet is useful even if you have to stop early. Write the time of your last update in the first row. If you're paused or stopped by a limit, the first thing you tell me when you're back is where you stopped and what's still missing.

Nothing goes out overnight (from Rohan Attravanam, Kurt Bell and Himanshu Jha; avoids messages being sent while you sleep):

Tonight, don't send any texts, emails or messages to anyone and don't submit forms. Draft them instead and put every draft in the morning report with who it's for, so I can approve or change each one. If something needs a sign-in or an approval, stop that item and list it for me.

Working on my computer (from Андрей and Eduardo; avoids the Dot interfering with your own machine):

You can use my computer for [the repo or app] tonight. Don't close other apps, log out, or put the computer to sleep. If the computer stops responding, keep going on your own cloud computer where you can, and tell me in the morning what couldn't be finished there. Send me a notification on my phone when everything is ready for me to take over.

For more overnight coding setups, see Codex orchestration. For how the Dot's own machine works, see the cloud computer page, and for morning reports that arrive on a schedule, see Morning brief. If a job stops with an error rather than a pause, start at Dots not showing or not working.

Related

Copy this task brief

Maintain the status of [project] through [end date].
Sources of truth: [plan], [issue board], [named conversation channels], [named email threads].
On conflicts, prefer [primary source] and show the conflicting source beside it.
Create a tracker with milestone, owner, due date, current status, next action and source link.
Each update must separate: newly added, changed, completed, overdue, waiting-on a decision, and at risk (due in 7 days with no next action).
Check at [frequency, time, time zone] until [end date]. Confirm the recurring task.
Notify me in [channel] when a milestone slips, a dependency blocks the next step, or a decision needs me.
Keep routine unchanged checks quiet.
For each open decision, give the options and your recommended next step.
Draft suggested follow-up messages for my review; do not send them.
Start with a baseline tracker and list the information you still need.
Download Markdown ↓

Sources & further reading