How to build a website or app with a ChatGPT Dot: chatgpt.site hosting, games and 57 real projects (OpenAI Dots, Oct 2026)

The short version: what a Dot can build for you, how to set it up, where it gets stuck

Is a Dot worth it for building sites and apps?

Yes for small, finished things you can open and check, and for long jobs you would otherwise have to babysit; not yet as a full replacement for your own development setup. In the cases below it does four things well. It builds and hosts small sites, dashboards and games from a plain request, often within minutes or an hour: a recipe site after a few chats, a game to play on a plane about 20 minutes after asking, a proof-of-concept web app with a Python back end in under an hour (cases 1, 11 and 14). It takes changes by voice: a website with a game asked for on a call, a personal website redesigned section by section on a 10-minute call, a company website turned into a game during a long drive (cases 3, 56 and 57). It keeps long work going while you are away: Godot demos built overnight, a coding job running around the clock, a VRChat world built all day with releases along the way (cases 24, 31 and 46). And it moves existing projects forward: running tests, preparing a submission, spotting a broken deploy, getting a mobile app ready for launch (cases 18, 19 and 20). It gets stuck when its own browser can't open what it built, on Docker and tools that need admin rights, on networks and sign-ins it can't reach, on data that only comes from paid APIs, on anything it can't see, hear or touch for itself, and on whether a game is actually fun. It suits people who can say what they want, check the result on a phone or a desktop, and correct it over a few rounds.

How to set it up

1. Check that you have both pieces. A Dot comes with Pro outside the EEA, UK and Switzerland, with Business Premium, and with Enterprise once an admin turns it on (How to get ChatGPT Dots). ChatGPT Sites, which hosts what it builds at a chatgpt.site address, is available on Plus and Pro and in supported Business and Enterprise workspaces, not on Free or Go (OpenAI Help).
2. Say which computer it builds on. Each Dot has its own cloud computer with a browser, files and a shell, and keeps working there while your laptop is closed (Introducing dots). Ask it to build there unless the project needs your machine, and to tell you first what is already installed. Alex Xiang's Dot had started on his laptop until he said no (case 27), and youer found that coding on a connected local computer is what used up his allowance (case 31).
3. Test the connections before any code. Have it try, from its cloud computer, the database, company service or package site the project needs. Yuji Nakakoshi's cloud computer couldn't reach oracle.com, so the build moved to his own Mac (case 49).
4. Put the brief in files. DigitWaves AI kept the plan, the requirements and the rules ("no made-up reviews", "the contact form sends nothing") in markdown files in one folder, with one image as the design reference (case 38). For a bigger job, settle the plan in a regular ChatGPT chat first and hand the Dot one project at a time (cases 45 and 46). If it will ship through a store, make the store's rules step one, as Maya Bello's Dot did with Apple's icon-testing rules (case 37).
5. Ask for it on ChatGPT Sites, kept to you. A new Site is limited to its owner and workspace admins until you change who has access, and in Enterprise workspaces public publishing stays off until an admin turns it on (case 25). Every deployment URL is a live URL, so ask the Dot to save a version for you to review before it deploys a change (OpenAI Help). If other people will sign in, ask how the back end decides who sees what, because a sign-in button isn't access control (case 26).
6. Tell it how to sign in for testing, and how to report. Simon Smith's Dot couldn't get past his passkey (case 53). Ask it to list separately what it clicked through in a real browser, what it only tested with scripts, and what still needs a real device (cases 23 and 27).
7. For long jobs, set the rhythm and the limits. DaisukeDaisuke used Custom rules that pre-approve the operations the job needs and a schedule that nudges the Dot every 15 minutes (case 24); Xenoah asks for a push, a release and screenshots as the work goes (case 46). Keep deploying and publishing on Ask before taking action, as Chase AI's Dot deployed only after he approved (case 12). Building or editing a Site counts toward your Codex usage, so check chatgpt.com/settings/usage before a long build (case 44).

Where it gets stuck, and what worked

1. "Tests passed" without a real click-through. AI Arte's Dot said its cloud browser couldn't open localhost, so only simulated tests had run (case 23), and DaisukeDaisuke hit the same wall (case 24). What worked: have the Dot install Firefox (case 24), or serve the unchanged file somewhere else, note its SHA-256 hash and click through it there (case 23). For a desktop app, ask for test results as screenshots (case 54).
2. Docker, sudo and Unix sockets. sebastian73's .NET stack compiled and passed 193 backend tests, then stopped at those limits, and omaredy09's Docker Compose project hit the same wall (case 32); lll9p needed fallbacks because the cloud computer has no zenity (case 28). What worked: keep cloud builds to things that run without admin rights, such as Godot demos, single-file apps and Sites, and keep multi-service stacks on your own computer.
3. Networks it can't reach, and a computer that falls asleep. What worked for Yuji Nakakoshi: a Codex project on his own Mac, with the Dot starting the Codex task and the database connection named in the request (case 49). Keep that computer awake while you talk to the Dot from your phone: Nakakoshi points out the Mac can go to sleep mid-instruction, and Alex (@uhak_crew) codes by voice only because his computer stays on (cases 36 and 49).
4. Sign-ins and permissions. Lennox Saint's first hurdle was getting into staging, and after that it went smoothly (case 18); Allison Johnson had to grant a shocking number of permissions before her Dot could rebuild her site in its back end, and then it worked (case 56). What worked: settle up front how it signs in for testing, and expect a round of permission prompts the first time it works on your computer.
5. Data, sound and touch it can't produce. Himanshu V's trip app stopped at about 70% because places, train lines and flight delays come from paid APIs (case 50); Aframe93's Dot changed a Windows audio app it couldn't hear (case 30); touch feel, heat and frame rate on Alex Xiang's game still needed a real phone (case 27). What worked: ask early which data it can't supply and which checks need real hardware, then do those yourself, and ask for proof from the device itself, as Fadi AlZuabi got with the time, the failed first flash and what the chip doesn't do (case 33).
6. Look-alikes instead of the real thing. Bub imitated a liquid-glass button instead of using the component library's code (case 38). What worked: ask which parts came from the real component, skill or connector and which were approximated, and inspect one yourself.
7. A game that runs but isn't fun. よりむー's Dot proved with logs that every game started, scored and restarted, and kept adding moves that didn't change how it was played (case 40). What worked: play it yourself and tell it which decision you want to be different each time; talk the rules through by day before an overnight build (case 31).
8. Long jobs that stop or go quiet. On its own the Dot tended to stop once the task in front of it was done (case 24), mocisly's cloud coding task stopped with "You've hit your usage limit" (case 44), and DigitWaves AI couldn't tell whether Bub was working or waiting (case 38). What worked: Custom rules plus a 15-minute schedule (case 24), a usage check before the job (case 44), and asking it to say each time it is waiting on you (case 38).

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

It can now:

• Build a small website, web app, dashboard or browser game and host it on ChatGPT Sites, kept to you, shared in a workspace or published.
• Take changes over a phone call or a Slack message and come back with the next version.
• Redesign a site you already have, and with access to your computer, rebuild the new design in the site's own back end.
• Keep a long build going overnight or around the clock on its own cloud computer, with releases and screenshots along the way.
• Run the tests on an existing project, prepare it for submission or launch, and notice a broken deploy.
• Make Godot games and Blender models and animations, and put an image on a small device connected to your computer.
• Turn a rough ask into a brief for Codex or another coding agent, and check what comes back.

It can't yet:

• Reliably open a page served on its own cloud computer in its default browser.
• Run Docker or tools that need admin rights on its cloud computer.
• Sign in with your passkey, or reach every network and paid data source a project needs.
• Hear audio, feel touch controls, or judge whether a game is fun.

One message to start with

Build me a [website / small web app / browser game] for [purpose] on your own cloud computer, not on mine, and host it on ChatGPT Sites, visible only to me. Before you write any code, tell me what's already installed there and check that you can reach [any service or data it needs]. Keep anything that sends, publishes or spends as a demo until I say so. When it's ready, send me the link and list separately what you clicked through in a real browser, what you only tested with scripts, and what you couldn't check. Then suggest the next three changes you'd make.

Quick answers

What is chatgpt.site?

chatgpt.site is the address used by ChatGPT Sites, which creates, hosts and shares websites, web apps and games without a separate deployment setup (Sites guide). A Dot builds on its own cloud computer and can publish the result there, so you don't need hosting of your own. Sites is available on Plus and Pro and in supported Business and Enterprise workspaces, not on Free or Go, and a Site can be kept to you, shared inside your workspace, or published to anyone (OpenAI Help). To try it, tell your Dot what site you want and that it should be hosted on ChatGPT Sites; the projects below show what people got back.

How do I ask a Dot to build a website?

Say what it's for and that it should be hosted on ChatGPT Sites, then correct it in follow-ups. @ai_dev_diary got a personal recipe site after a few chats, with ChatGPT sign-in on the admin side, and Mykhailo Patsan simply said "make me a website with a game" on a call. For a business site, DigitWaves AI put the plan, the requirements and the rules (no made-up reviews) in markdown files in a folder and steered his Dot round by round. Keep it visible only to you until you've checked it.

How long does a build take?

Small things come back fast. Mustafa Ergisi was playing his plane game about 20 minutes after asking, @ai_dev_diary's proof-of-concept web app with a Python back end took under an hour, Alistair Barr's "Strange Endings" site was up in about five minutes, and Fadi AlZuabi's avatar was on an ESP32 screen 38 minutes 36 seconds after the ask, failed first flash included. Bigger jobs take days: Jeon Seung-gi's personal ERP got stuck partway, youer leaves Godot demos building for 10 hours or more overnight, and DaisukeDaisuke kept a coding job running around the clock with a 15-minute schedule.

Who can see what my Dot builds?

A Site can be kept to you, shared in your workspace or published to anyone; in Enterprise workspaces public publishing stays off until an admin turns it on, which is why Alistair Barr's site stayed private. A sign-in button isn't access control, as Kyle of KodeLab points out: Sites passes the visitor's email address to the back end, which still has to decide who sees which notes. And if you sign in with a passkey, the Dot may not be able to test past it, as Simon Smith found.

Does building with a Dot use my Codex usage?

Work the Dot did on its own cloud computer often didn't: @ai_dev_diary used no Codex tokens, sorilove's playtest loop didn't touch the Codex limit, and lll9p saw no allowance used as long as Codex wasn't involved. youer found that coding on a connected local computer is what used it up. Coding the Dot hands off does count: mocisly's cloud coding task stopped with "You've hit your usage limit". Building or editing a Site counts toward your Codex usage, while running and hosting it, including traffic and storage, is counted separately (OpenAI Help). See Pricing and usage.

Can it make games and 3D models?

Yes. Beto's Dot shared a retro Snake game as a link, Alex Xiang's built a Godot phone game with 36 automated checks and a first preview release, and Batraz Botsiti got Blender models and an animation from one prompt; Xenoah has a Dot build a VRChat world all day, with GitHub releases and screenshots along the way. Judge the fun yourself: よりむー's Dot shipped games that started, scored and restarted, but couldn't yet come up with the fun. More Blender work is on the cloud computer and video & animation pages.

Where do builds get stuck?

The Dot's cloud browser couldn't open a page served on its own machine for AI Arte and DaisukeDaisuke, so ask what it really clicked through. Docker, sudo and Unix-socket limits stopped sebastian73's .NET stack and omaredy09's Compose project. Yuji Nakakoshi's cloud computer couldn't reach oracle.com, Himanshu V's trip app stalled at about 70% because the data comes from paid APIs, and Aframe93's Dot couldn't hear the Windows audio it was changing.

Which plan do I need?

Sites is available on Plus and Pro and in supported Business and Enterprise workspaces. 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, read Dots in the EU and UK first. Everything else a Dot can do is on the features page, and if it goes quiet mid-build, start at Dots not showing or not working. For blog drafts and scheduled posts rather than sites, see content pipeline.

1. Websites

1. A personal recipe site, hosted, after a few chats (ai_dev_diary)

Korean product manager @ai_dev_diary asked their Dot for a recipe website just for them. After a few conversations it was built and live at a chatgpt.site address, with ChatGPT login on the admin side so only they can request recipe changes. What surprised them was that they needed no hosting of their own and, in their words, used no Codex tokens.

2. An apartment you can walk through in 2D and 3D (xucidal)

@xucidal's Dot Sol built "Plan to Place — Apartment 01", a site where you explore a furnished sample apartment in 2D and 3D, take a camera tour, change the light, and switch between two interior palettes. It is hosted on ChatGPT Sites.

3. A website with a game, asked for on a phone call (Mykhailo Patsan)

Mykhailo Patsan (in Ukrainian) set up his Dot, named it Ernest, called it, and said "make me a website with a game". A little later he was playing the Snake to Earn game it built. After that the Dot moved on to meeting notes, client documents, presentations, negotiation prep, and putting his gym sessions on the calendar.

4. Website changes picked up by voice at 2:37 a.m. (Matt Maher / Dippy)

Matt Maher called his Dot Dippy at 2:37 in the morning to check on existing website work without getting pulled back in. The video shows the real changes coming back, along with the corrections he still wanted. He credits his folder setup and a connected computer for giving Dippy the context to start.

5. A photography site reviewed, then rebuilt from scratch (Chris Verzwyvelt)

Chris Verzwyvelt (@ChrisCoffeeAi) had his Dot browse his wife Hannah's live photography website, list its real problems, and then rebuild the whole site to see whether it could do better. The video shows the review and the rebuild on screen.

6. A clearer website and an Android dictation widget (AI Mastery Hub / Leo)

The AI Mastery Hub named their Dot Leo and shows three projects: making the AI Mastery Hub website clearer, building and refining an Android dictation widget, and keeping a daily log of what works. They are open about wrong turns and fixes along the way, and say they still review important decisions and check results.

2. Dashboards and tools for your own work

7. An owner-only dashboard for 12 projects (Zate Berg)

Zate Berg had his Dot build a main dashboard as a Site with owner-only access, tracking the status of his roughly 12 projects. He also has it build plugin "gateways" that connect the dashboard to other things, including one that autoplays the classic game Bard's Tale on his laptop through DOSBox-X so he can manage the party remotely.

8. A personal ERP that maps everything he works on (Jeon Seung-gi)

Korean creator Jeon Seung-gi (@soo_sang_han) built a personal ERP with his Dot alone: a home screen with a project map covering his store, marketing, lectures, development, and personal schedule, where each project opens its own tasks, dates, and files. He says they got stuck partway and only just pushed through.

9. A project dashboard sorted by what needs attention (Tiffany Kyazze / Pixel)

Tiffany Kyazze (@techtiff.ai) named her Dot Pixel and asked it to track every open project across her ChatGPT conversations. It sorted them into active, waiting, and needs-attention, then built a dashboard that put her "Post dots video" task first, all in one thread she could check and redirect.

10. A Decision Lab built overnight, unprompted (Morgan Linton)

Morgan Linton (30 Sep 2026) woke up to a Decision Lab his Dot had built for VulcanBench while he slept, without being asked. He was using it the same morning and said it felt like a new employee trying to impress him.

11. A proof-of-concept web app with a Python backend in under an hour (ai_dev_diary)

The same product manager, @ai_dev_diary, gave their Dot a long list of requirements for a pilot web app for a new project. It delivered in under an hour with a clean design, almost no errors, and Python backend code as well as the front end.

12. A lead-intake dashboard, deployed after the go-ahead (Chase AI / Winston)

In a sponsored reel (#ChatGPT_Partner), Chase AI asked his Dot Winston to pull inquiries from several sources into one system. The video shows Winston building and testing a Lead Intake dashboard on its cloud computer and deploying it once he approved. The leads in the demo are sample data.

13. A CRM that lives in the chat and texts him about deals (Oren)

In a sponsored post (#ChatGPT_Partner), Oren (@orenmeetsworld) (29 Sep 2026) used his Dot to build a CRM that lives in his chat and texts him when one of his deals needs attention.

For a smaller money tool, the finance page has a per-project expense tracker a Dot built from card and bank statements in about 15 minutes.

3. Games

14. A game app to play on the plane, in about 20 minutes (Mustafa Ergisi)

Mustafa Ergisi (@mergisi) was on a flight when he asked his Dot in Slack to build something fun for the plane. About 20 minutes later he was playing his own game app.

15. A Snake game, a new Xcode project, and a Blender dog (Code with Beto / Oso)

Developer Beto (@codewithbeto) called his Dot Oso from his phone and gave it three jobs: a 3D model of his Great Pyrenees in Blender, a new Xcode project for an app, and a retro Snake game, which it finished and shared as a link. His full hands-on video shows the tasks handed over by Slack and phone.

16. A 3D game, then a website for it (Aaron / AI-native builder)

Aaron gave his Dot a small brief: a ceramic courier, three energy cores, and a rocket to bring home. The Dot built the 3D game, then added night lighting, smoother cameras, a website, and a flight replay. He notes the website is a local preview rather than a public deployment, and that Codex handled the local recording and rendering while he directed.

17. Game rules tuned in a playtest loop (sorilove)

@sorilove (in Korean) uses their Dot to refine the rules of a game in development: play the build, give feedback on what is fun, change it, and check again. Doing that loop in Codex used up a lot of their usage, they wrote; in the Dot it did not touch their Codex limit.

4. Shipping and keeping things running

18. A ChatGPT extension tested and prepared for submission (Lennox Saint)

Lennox Saint has had his Dot running his ChatGPT extension project for several days: running tests and getting it ready for submission largely on its own. The first hurdle was getting into the staging environment; after that, he said, it went smoothly and he was far less involved.

19. It saw the deploy break first, then counted leads in production (André Prado)

Brazilian creator André Prado (in Portuguese) tested his Dot on his real projects. It noticed his deploy had broken before he did and counted his course's leads directly in the production database. He also shows what was confusing.

20. A mobile app made ready for launch (Peter Yang)

Peter Yang (29 Sep 2026) had been testing his Dot for a few days. Among the results he listed: it got his mobile app ready for launch and flagged a five-figure revenue opportunity. His video walks through it.

5. Tools people built around their Dot

21. A command-line tool to message your Dot from the terminal (Zeal Chou)

Zeal Chou (in Chinese) shared a CLI for ChatGPT Dots that shows its status, reads its messages, and sends it messages from the terminal; the screenshot shows a message sent from the CLI getting a reply. It is packaged as a skill so other coding agents can talk to the Dot through it, with an optional MCP server.

22. A KakaoTalk connector, so the Dot only flags what matters (Ellis)

One of the first things @superstarcong (in Korean) did was name their Dot Ellis and build an MCP connector that lets it read their KakaoTalk messages. The agreement with Ellis: only tell them about messages they absolutely cannot miss.

For more on what people install and run on the Dot's machine itself, including a desktop app for calling your Dot and taking over its mouse, see the cloud computer page.

6. Checking that the build really works, and keeping a long job running

A Dot can tell you that the app is done and the tests passed. These two owners went further and checked what that report actually covered. Both also found the same gap: the browser on the Dot's cloud computer would not open a page served on the machine itself (localhost).

23. A ToDo app the Dot built but couldn't click through itself (AI Arte)

The Japanese tech site AI Arte (1 October 2026, updated 9 October) asked a Dot, from the Codex app on Windows, to build a small ToDo app with HTML, CSS and JavaScript only, on its own cloud computer, without delegating to Codex, connecting the owner's PC, publishing or pushing. The Dot delivered one todo.html file with add, complete, delete, saving in the browser and a refusal of empty entries. It then reported that its cloud browser was blocked from opening localhost and the file directly, so only simulated tests had passed, and it said plainly that the app had not been checked in a real browser. The owner served the unchanged file locally, recorded its SHA-256 hash so it was clear which version was tested, and had a separate Codex chat click through it: adding, completing, deleting and reloading all worked. A follow-up request added an "undo delete" button, and that version was checked the same way. One more detail from the first chat: the Dot suggested a private plan the owner had discussed earlier in ChatGPT, before it had been mentioned to the Dot, and the owner told it to leave that out and use only public information.

24. A coding job kept running around the clock on the cloud computer (DaisukeDaisuke)

DaisukeDaisuke (Qiita, in Japanese, 4 October 2026) hands the Dot reverse-engineering work on a game and wrote up about 60 hours with it. On its own, the Dot tended to stop as soon as the task in front of it was done. After three days of trial and error, DaisukeDaisuke settled on three settings: a custom rule saying that when "no-stop is allowed" is given, the Dot must not stop; a custom rule that pre-approves the operations the job is likely to need; and a schedule that, every 15 minutes, tells the Dot to push the open items forward. With that, the job had been running 24 hours a day at the time of writing. Two other findings: the default browser on the cloud computer couldn't reach localhost, so DaisukeDaisuke had the Dot install Firefox; and writing code with the Dot alone was fairly slow. DaisukeDaisuke also reports that the cloud workspace was wiped about every 35 to 42 hours, and set up a restore the Dot runs on its own, described on the cloud computer page.

If your build depends on existing Codex chats on your own computer and the Dot can't reach them, Jeremy Devz's "bridge task" workaround is on Codex orchestration.

7. Sign-in, sharing and workplace limits

25. A site that stayed private because the workplace said so (Alistair Barr, Business Insider)

Alistair Barr writes Business Insider's AI Insider newsletter and got a Dot, which he named alidot, at an OpenAI training session through the company's ChatGPT Enterprise account (Business Insider, 10 October 2026). His first job was a site he had been meaning to make for years: a Tumblr-style feed of real, strange deaths, each one checked, kept up to date. About five minutes later "Strange Endings" was up with six source-checked cases. It was private, and the Dot said automatic updates weren't on yet because access to the site's update endpoint was blocked. When he asked for a public link, the Dot checked and reported that the workspace doesn't allow public links or invitations to people outside it, so an admin would have to turn on external sharing first. An OpenAI spokesperson told him that on a paid personal ChatGPT account the site would have been easy to put live. The newsletter part of the same test is on the content pipeline page.

26. A notes app with a database and ChatGPT sign-in (Kyle, KodeLab)

Kyle, who writes the KodeLab blog, got access on a Pro 100 plan on 30 September and asked the Dot for a private notes web app with a front end, a back end and a database (KodeLab, in Traditional Chinese, 30 September 2026). The result was deployed with ChatGPT Sites to a *.chatgpt.site address, with a sign-in page offering "Continue with ChatGPT" and a notes interface behind it. Kyle points out two things. The site is hosted by ChatGPT Sites, not served from the Dot's cloud computer. And a sign-in button isn't access control: when someone signs in with ChatGPT, Sites passes their email address, and possibly their name, to the site's back end, which still has to check who may see which notes. Kyle's look inside the cloud computer itself is on the cloud computer page.

8. Choosing where it builds, testing desktop apps, and handing work to other coding agents

These six owners cared less about what the Dot built than about where the work ran and who checked it. One moved a build off his laptop with a single sentence, one used the cloud computer as a test bench for a desktop app, one made the Dot the planner for two other coding agents, one found where that test bench stops for Windows audio, one left game demos running overnight on the cloud computer, and one got most of a .NET stack compiling before Docker and privilege limits blocked the rest.

27. A Godot phone game, moved off the laptop and onto the cloud computer (Alex Xiang)

Alex Xiang, a former Baidu and Weibo engineer, recorded 2 minutes 50 seconds of a session on the evening of 30 September and wrote it up on his blog; he notes that the write-up itself was produced with an AI tool from that recording (zicode.com, in Chinese). His ask was one sentence: build "One-Key Gravity" (一键引力) in Godot, for phones, polished and fun to replay. The Dot planned an Android landscape touch version and a Windows trial build and reported that it had already started on his laptop. He replied "don't do it on this machine". The Dot stopped straight away and asked only whether to move to its cloud computer or another computer. Once he chose the cloud, it checked first that Godot and Java were already installed there and said it would no longer use his laptop.

He then asked it to sync the code to D:\work\games on another machine named alex. The Dot could see alex in its list, but reported that it was Offline and that task access hadn't been turned on, so it kept building in the cloud. He didn't finish the steps to bring alex online, so that sync never happened. When he took control of the cloud computer and checked it himself, it was an 8-core slice of an AMD EPYC 9V74 with 9 GB of memory, Debian 13, a 32 GB disk with 3.5 GB used, and no GPU.

The game went up as an MIT-licensed GitHub repository with a first playable release, v1.0.0-preview, carrying 10 files: a compact ARM64 Android test build of about 26 MB, a larger generic test build, two source packages, TEST_REPORT.md and SHA256SUMS.txt (release page). The Dot's 36 automated checks covered restarting, pausing and resuming, duplicate taps and level simulation. The release notes say the game ran and its screens were checked on Linux, that Android and Windows had not been tried on real devices, and that the Android build uses a temporary test signature. The Dot said the same thing in the chat: touch feel, heat and frame rate still had to be checked on a real phone. When it later prepared a short vertical video, it said it would label his sidebar screenshot as a screenshot rather than mock up a chat screen and pass it off as a recording.

28. A desktop app tested function by function, with a manual written along the way (lll9p)

lll9p, the author of MicroSeg, posted on linux.do on 10 October about a use they had found (linux.do, in Chinese). They uploaded a desktop app with a graphical interface, had the Dot test each function on its cloud computer and write a report, and had it write the user manual at the same time. Their observation is that this didn't use up any of their allowance as long as Codex wasn't involved. The only snag was that the cloud computer has no zenity, the small tool many Linux programs use to show dialog boxes, so they had to add fallbacks for it.

29. The Dot as planner, Claude Code doing the work, Codex checking it (genkunjc)

genkunjc wrote up an experiment on Zenn on 7 October: the Dot decides what to do next, Claude Code does the work, and the Codex CLI reviews the result, with no one copying text between them (Zenn, in Japanese). The Dot drops request files and a policy.json into its working folder. A small runner on the PC picks them up every 10 minutes, starts Claude Code on each one unattended, has Codex audit the result, logs time and cost, and types a short "request finished" note into the Dot's chat. The Dot sets the pace through the policy file: whether to pause, how many jobs run in parallel (1 to 4), the model, whether to audit, the timeout and the order. Pushing, publishing, sending and spending are blocked in the runner's code, and values outside the allowed range are clamped and noted.

In about a day from 6 October the loop handled 21 requests: 16 started unattended and 5 passed to a Claude Code chat genkunjc had open. Real requests took about 10 to 26 minutes each. Codex passed 8 and sent 9 back for fixes, so close to half of what Claude Code had called finished didn't survive a second look. The pitfalls genkunjc lists: the Dot can only write inside its own task folder, which changes per task (Documents/Codex/<date>/<task>/), so the runner has to watch a wildcard path; and on Windows the scheduled runner couldn't see the claude command because it had been installed inside the Claude desktop app's sandboxed package, so it needs the full path. genkunjc calls it a trial that works but is still shallow, and doesn't recommend that everyone install it.

30. A Windows audio app the Dot could change but couldn't really hear (Aframe93)

Aframe93 posted in the OpenAI Community on 11 October about building a Windows audio application with their Dot, which hands the coding to delegated Codex tasks (OpenAI Community). The Dot can read the code, make changes and run synthetic tests on Linux. What it can't do is run the Windows audio stack: in the cloud environment their workflow uses, Wine wouldn't start, because the local inter-process communication it needs was restricted. Aframe93 explains why that gap matters. The app sounds robotic at small buffer sizes and still crackles at larger ones, yet the synthetic tests pass, so passing them proves nothing about the real audio path. Because the Dot can't reproduce the problems, Aframe93 keeps installing builds and testing them by hand. The post asks OpenAI for an optional Windows test environment with virtual audio loopback and automatic runs across buffer sizes and sample rates, and accepts that real hardware would still need testing.

31. Godot demos discussed by day, then built overnight on the cloud computer (youer)

youer (linux.do, in Chinese, 10 October) uses the Dot to make game demos and quickly check whether an idea is fun to play. The demos run on the Dot's cloud computer in Godot. The pattern that worked for them: talk through the idea during the day, then leave the Dot to develop for 10 hours or more overnight, and play the demo in the morning. They note that the model sometimes adds its own constraints or settings, so clearer prompts or a few rounds of discussion help before the long run. Coding on a connected local computer, they say, uses up allowance; staying on the cloud computer avoids that.

32. A .NET and Blazor stack that compiled, then hit Docker and privilege walls (sebastian73)

sebastian73 (OpenAI Community, 1 October) tried to use the Dot's cloud computer as a full development environment so the home PC could stay off. The project was a .NET 10 application with a .NET Core backend, a Blazor frontend, SQL Server, and Playwright. Dependencies restored, the backend compiled, and 193 backend tests that don't need SQL Server passed. The full stack did not: the machine is Debian 13 without sudo or write access to /usr/bin and /opt; Docker's host socket was denied; rootless Docker failed on missing privileged UID/GID helpers (Operation not permitted); Unix socket creation returned EPERM and also blocked the Blazor build-task host and Chromium, though local TCP binding worked. sebastian73 asks for a supported way to run isolated Docker workloads with documented limits and private previews, and notes the same setup worked on Grok Bot. In the same thread, omaredy09 reported the same wall for a Docker Compose project (database plus Node.js services): Docker wasn't usable under the privilege limits, so they couldn't reuse their existing development setup.

9. Hardware, smart glasses and a bridge between two computers

These three owners built something outside the browser: a tiny screen, a pair of glasses, and a link between two machines. The Dot did the setup work in each case, and each owner said plainly what the result does and doesn't do.

33. Its own avatar on an ESP32 screen, after a failed first flash (Fadi AlZuabi)

Fadi AlZuabi (@fadi.zuabi) (8 Oct 2026, also posted as an Instagram reel) asked his Dot to put its avatar on the round display of an ESP32 board. The first flash failed. The Dot diagnosed the problem, installed a fix and flashed the board again, and 38 minutes 36 seconds after the ask the screen showed "hello, I'm Dot". He is clear about what it is: the chip only shows the avatar and the greeting, the AI model isn't running on it, and the voice in his video is synthetic.

34. A Dot on Even G2 smart glasses, with voice input and notifications pushed to the glasses (Yusuke Ito)

Yusuke Ito (@y_5k110) (in Japanese) got his Dot working on Even G2 smart glasses. A long press starts voice input, and replies can be read aloud. Speech is turned into text on his own main machine, so there is no per-use charge for it, and the read-aloud uses the iPhone's built-in voice, which is free. These sessions are kept apart from his desktop sessions, so they don't show up there, but the Dot can still pull in their context. He says usability still needs work, but the original goal was met. In a later post he calls the core feature the fact that it isn't one-way: the glasses also receive notifications from the Dot.

35. A bridge between two remote computers, set up from a video transcript (Abraham Ventura)

Abraham Ventura (@ceo_abraham) (10 Oct 2026) had watched a workaround video by Jeremy (@jeremydevz) (8 Oct) about letting a Dot reach local Codex and ChatGPT tasks, threads and chats. He gave his Dot the transcript, and it set up a bridge. He reports that two-way task communication now works across two remote computers, and calls it a workaround while he waits for a fix from OpenAI. For more on Dots working with local Codex tasks, see Codex orchestration.

10. Coding by voice from a phone

36. From planning to final code by voice, through Relay MCP (Alex / @uhak_crew)

Alex (@uhak_crew) (Threads, in Korean, 11 Oct 2026) now vibe-codes without typing. His loop runs from his Dot through Relay MCP into coding sessions in his IDE (Gemini, Claude and Codex), then back through Relay MCP to the Dot. He plans with the Dot and goes all the way to the agreed version of the code by talking from his phone; the only condition he mentions is leaving his computer on. He calls the Dot a perfect work assistant and expects the same setup to feel just as natural once the phone gives way to AI glasses. The post doesn't say which app or project he built this way.

11. Planning an App Store icon test on a call

37. An App Store icon test planned on a call, starting with Apple's rules (Maya Bello / Side Quest Slayer)

Maya Bello (YouTube, 12 Oct 2026) makes an app called Side Quest Slayer. Earlier in the chat her Dot had sent her a video in which another app tested a robot icon and a brain icon against its original icon; the brain won, although that team had expected the robot to. On a call she asked whether the Dot remembered sending it, and what steps an icon test for her own app would take. Before suggesting any design, the Dot checked Apple's icon-testing requirements, because an alternate icon may need to be included in an approved app build before a test can start. She asked it to put the steps in the chat so she could see them, then had it look up Side Quest Slayer's current icon and compare similar journaling and memory apps with visible ratings, to find icon ideas worth testing. A smaller first job in the same video was an animation she started from her phone; she asked for one revision because she wanted it to feel more celebratory, with the hands going up and down. The video ends before any test runs, so there is no result yet. Her verdict: worth trying if you already have the plan, but she wouldn't upgrade just for Dots. The call itself is covered in voice calls, case 35.

12. A business website built with steering

38. A restoration-company website, built from a markdown brief and steered round by round (DigitWaves AI / Bub)

DigitWaves AI (YouTube, "ChatGPT Dots Built This Website — My Honest Verdict"; short version on Instagram, 7 Oct 2026) named his Dot Bub and gave it a dedicated episode folder on his connected Windows PC. Markdown files in that folder held the plan, the website requirements and the rules for a fictional water-damage business, Still Bank Restoration: a navy masthead, a cyan help button, a full-width living-room hero, three service sections, a demo contact form that sends nothing, and no made-up reviews or performance claims. One image served as the design reference. Bub first turned the brief into a build prompt, then worked through his requests one at a time. It used the Higgsfield connector to turn the hero picture into a 5-second clip with gentle fades, added pictures for each service, and made each text and picture column ease in on its own as you scroll. He had it run the Impeccable design skill to strip out the generic AI look, then use components from 21st.dev. The first pass only added a small button effect, so he browsed the library himself, picked a connected-card design and a team layout, supplied the portraits, and had Bub adapt them; the cards became a restoration diagram near the top of the homepage. He also pointed the hero at homeowners with water damage, which produced the headline "Water damage. Take back your home." and a "Tell us what happened" button. When he inspected a liquid-glass button, Bub admitted it had approximated the look instead of using the library's actual code. The real code made the buttons blend into the background, so they went back to solid cyan. He then checked the finished site on desktop and on a phone.

What still needed him: the wording, the design checks, confirming that real code had been used, and asking for status, because nothing showed clearly whether Bub was working or waiting for him. Bub hit permission issues on the video side of the project, so he moved that part to a different model. Bub also identified a Windows file-transfer compatibility problem but did not fix it, so he downloaded the files himself. His verdict: less preparation than his earlier workflows, light usage, and Bub becomes his daily driver for website builds.

13. A small public site built with Dots and Codex

39. A gallery of Dots characters you can restyle and export (liuzhihang)

liuzhihang (X, 13 Oct 2026) built Dots · Make It Your Dot, a character gallery, with a Dot and Codex. You pick a character and adjust its look, then either export it as a transparent PNG or copy its prompt and keep creating in ChatGPT. The site runs at a chatgpt.site address, switches between English and Chinese, has search and sorting, and takes new characters through GitHub, where sign-in is required and submissions are public. Two days earlier, liuzhihang had reported a limit on macOS: the Dot could read the Codex tasks it created but not the Codex chats liuzhihang had started by hand, covered on Dots with Codex.

14. A game studio run by the Dot: games that worked, but no reason to play again

40. Browser games the Dot planned, built and tested, judged on whether they were fun (よりむー, @yorimuu_)

よりむー (note, in Japanese, 3 October 2026; X post, 13 October) told a Dot it was the head of a small game studio and handed it browser-game development, including the choice of what to make. The plan was three days, 2–4 October, with one playable browser game finished by 11 p.m. each day, the 6.1Sol model to stay inside the weekly usage limit, no extra budget, and nothing published without the owner's approval. A game counted only if it could start, the player's judgement or skill changed the result, a result screen appeared, and the player could try again. From three drafts the Dot picked 宙縫い (Sora-nui): predict where moving stars will go and catch them inside a triangle of three points within 75 seconds, while avoiding red stars. Start, result and retry all worked, and the Dot found and fixed a hard-to-see Start button on its own. When the owner said it still wasn't fun, the Dot added a move that charges up amber stars and slashes them all at once, which did feel good. Then it added a comet move, but the owner never needed it, because the old charge-and-slash routine was safer. A rotating-parts puzzle and a compact tactics game followed; the puzzle came down to trying each part's four orientations in turn, and the tactics game showed its winning line on the first play. Along the way the Dot analysed eight well-known games from published design writing, developer interviews and player reviews, without having played them. Its last prototype, 星火の夜 (Hoshibi no Yoru), is a falling-block puzzle that sets off fireworks on chains. It came with browser test logs: 24 pairs placed at random reached at most a one-step chain, and fixed inputs reproduced two- and three-step chains. The owner's first question on playing it was how it differed from Puyo Puyo. The experiment ended early, on day two, 3 October. The owner's conclusion was that the Dot drove development on its own and coded well, but couldn't yet come up with the fun from scratch, and that "works as specified" and "fun to play" have to be checked separately.

15. A music map built with Dots

41. City Pop Atlas: Japanese city pop as a graph of artists, records and credits (ヴィンス, @antifragile77)

ヴィンス｜AI最前線 (@antifragile77) (13 Oct 2026, in Japanese) built City Pop Atlas with OpenAI Dots, and tagged the post with Codex as well. The site lets you follow Japanese city pop through artists, albums, songs and the people credited on them, laid out as an interactive graph. The features he lists: search and filters by name, era and type; a graph that shows how people and works connect; embedded Spotify players and links to YouTube; explainer articles with illustrations; and a switch between Japanese, English and Chinese. He pitches it as a way in, from one favourite song to records and musicians you don't know yet. The site is live.

16. From a messy product ask to a Codex session

42. hi.new registration: a one-line ask became a Codex brief and a live session (浦田勇樹 / @yuki_urata)

浦田 勇樹 (@yuki_urata) (13 Oct 2026, Japanese, with screenshots) asked his Dot, in very rough terms, to register hi.new and, if the form asked for a name or email, to use Tachikoma for those fields. The Dot turned that into a detailed Codex brief: use a supported browser on a Mac mini, verify the owner email, limit the job to registration and issuing a continuous-use token, do not install the CLI, do not talk to other bots, do not expose the secret token in chat or screenshots, and stop if new terms or costs appear. It then started a Codex session with that brief. He called Dots a strong orchestrator for Codex work.

17. A PDF editing engine's remaining tasks

43. Remaining work on a PDF editor, tracked down to a weekend finish line (はじめ / @hajime2ai)

はじめ (@hajime2ai) (13 Oct 2026, Japanese, with a task list screenshot) had asked Dots to clear the remaining tasks on a PDF editing engine (public version v82, with a Web UI ready to integrate). The open list covered verifying the Web UI in real browsers (including Japanese input), freer edits on real estate flyers and Office-tagged PDFs without breaking other glyphs, image and shape replacement that keeps layout, stronger save and regression tests, and cleaner API / CLI / MCP UX for humans and agents. He wrote that quite a few items were already cleared and it looked finishable that weekend.

18. A long build stopped by the usage limit

44. A cloud coding task the Dot started, stopped by "You've hit your usage limit" (mocisly)

mocisly (@mocisly) (13 Oct 2026) uses a Dot for ongoing software development. At around 17:25 UTC that day, a cloud coding task the Dot had started stopped with the error "You've hit your usage limit. Visit chatgpt.com/settings/usage to purchase more credits or try…", and mocisly raised it in a reply to @thsottiaux on X. As the error shows, coding work a Dot hands off counts against your ChatGPT usage. Marko Galesic asked the same day for a way to choose whether the work threads his Dot launches use that usage or API credits, so an ambitious cloud project doesn't use up what he needs for everyday tasks. Before you hand a Dot a long build, check what is left at chatgpt.com/settings/usage, and see Pricing and usage for how the plans differ.

19. Plan in a regular ChatGPT chat, then hand the work to the Dot

45. Decide in a normal chat, confirm briefly with the Dot, then let it work (わたしさん, @tokiwa_app_jp)

わたしさん (@tokiwa_app_jp) (13 Oct, Japanese), who makes the TOKITEM app, replied to someone who disliked waiting for the Dot's slow answers by describing how they work: decide things first in a regular ChatGPT chat, pass the decision to the Dot as an instruction, talk with it briefly to confirm what it will do, and then leave it to get on with the work. They had also had the Dot work on a second app in parallel, but stopped because running both at once confused them. They add that asking the Dot to work on its own computer, rather than through Codex or Work, helps keep token use down for now. If your Dot feels slow to talk to, try the same split: settle the plan where answers come quickly, then hand the Dot one project at a time.

20. A VRChat world: ChatGPT for the brief, Dot for 3D all day, Unity at home

46. Concept art and a stage brief in ChatGPT, then the Dot builds 3D all day with GitHub releases (Xenoah, @shuminchuuu)

Xenoah (@shuminchuuu) (14 Oct, Japanese) posted the pipeline that had been most stable for building a VRChat world with AI only. First they brainstorm in regular ChatGPT and export concept art plus a stage-setting markdown file as a zip — they call this step the important one and say to go carefully. Then they hand that package to a GPT-6 Dot and leave it building 3D for the whole day. Two habits they keep: the Dot pushes to GitHub and cuts a release as it goes, and every progress report and Git update comes with screenshots attached. While it works they stay in the Dot chat to reject bad results and throw in new ideas. At home they pull the repo onto their own PC, have Unity MCP assemble it locally, then upload the world. What makes it work is the split: a careful brief, a long unattended Dot build with release checkpoints, and a local Unity finish.

21. More Sites builds steered day to day

47. An anime "what to watch" site on ChatGPT Sites, built entirely through the Dot (Joe Hu, @hubeiqiao)

Joe Hu (@hubeiqiao) (14 Oct) built a side project that answers what anime to watch, where to watch it, and what to watch next. He says it was built alongside his other projects, entirely with ChatGPT Sites through his Dot, after days of back-and-forth fixing the same things over and over — enough friction that he felt the gap with Opus 5.5. The live site is anime.hubeiqiao.com. It is a personal tool he keeps shipping and improving, not a one-shot demo.

48. A Sites workroom as a kanban board, talked to through Dots and MCP (るでぃ, @Ludi_uni)

るでぃ (@Ludi_uni) (14 Oct, Japanese) replied that they made a workroom on Sites and turned it into a kanban board, then keep work moving by talking to it through Dots and MCP. MCP event hosting still has to live elsewhere; everything else in that setup stays on OpenAI. Here Sites works as the Dot's visible task board rather than only as a public website.

22. Sites as a back end for your Dot, and where builds hit walls

49. An Oracle APEX app built by talking to the Dot on an iPhone (Yuji Nakakoshi, @yuji_nkks)

Yuji Nakakoshi (@yuji_nkks) (14 Oct, Japanese), an Oracle APEX advocate, wrote up the whole setup on a blog post. The plan was to build on the Dot's own cloud computer, which already had a browser and Java 21, but it couldn't connect to oracle.com even though openai.com and microsoft.com worked, so that idea was dropped. Codex Cloud didn't work either: as far as Nakakoshi could tell, it only allowed outbound HTTPS, while SQLcl needs a TCPS connection to the Autonomous AI Database. So Nakakoshi made a Codex project on a MacBook at home, installed Oracle's APEX and database skills for Codex, turned on "your local computer" when creating the Dot, and told it to start a Codex task in that project and to connect with sqlcl -name adb_apexdev. After that the work went by phone call from an iPhone: the requests were a regional report and a product chart from the Sales History sample data, and the app's URL sent to the chat when it was ready. The notification came, and the finished app opened and signed in on the phone. On the way, calls kept cutting out and had to be redialled many times. There was no transcript of what was said, the Codex task didn't appear in the side menu, Computer Use asked for approval again and again, and the post points out that the Mac can go to sleep while you are giving instructions from the phone.

50. A Japan trip app that got 70% of the way to Wanderlog (Himanshu V, @himanshu_v92)

Himanshu V (@himanshu_v92) (13 Oct) built a Japan trip app with Dots and showed it at home. The reply was: "I already use Wanderlog, what's the point?" So the next ask to the Dot was to add Wanderlog's features, and it added most of them in minutes. What it couldn't supply was the data: places, reviews, train lines, flight delays and map routes all come from paid APIs. Himanshu puts the result at about 70% of the way there, and the conclusion is that what is hard to copy is the data and the reach, not the app itself.

51. One agent wrote the app, the Dot built it and caught a bug (cradrrat)

cradrrat (14 Oct) had another agent, Argon, write the app and gave it to their Dot to build, then posted the Dot's note: it was packaging the app with macOS's built-in WebKit, so the app wouldn't need a server running in the background, and it had found a problem in the code it was given. Opening a draft and then cancelling it would delete the draft, so it was fixing that as well.

52. Two Sites with MCP: a knowledge graph in chat and a shopping list the Dot looks after (Emmanuel Galanos, @egalanos)

Emmanuel Galanos (@egalanos) (14 Oct) first built a Site with an MCP app UI that shows a knowledge graph inside the chat. Then came a second Site with a web page and an MCP endpoint, so the Dot can keep a shopping list up to date. What Galanos would like next is for Sign in with ChatGPT to also let the Site's worker call the OpenAI API, so the Site could make decisions and process what gets typed into the page.

53. A Sites-backed plugin the Dot couldn't fully test because of a passkey (Simon Smith, @_simonsmith)

Simon Smith (@_simonsmith) (14 Oct) uses Sites as an MCP server for custom plugins built for personal use through a Dot. The trouble is sign-in. The Dot couldn't fully test the app because it couldn't sign in to the Site, which Smith signs in to with a passkey. The next day another sign-in problem stopped the plugin from reaching the MCP endpoint. Smith wants an easier way to make a Sites-backed custom plugin without fighting with authentication.

54. A long background port of gimp-painter, with UI tests reported as images (しーげっち, @seagetch)

しーげっち (@seagetch) (14 Oct, Japanese) has had Dots porting gimp-painter to GIMP 3.0 in the background for a long time. The Dot is asked to report UI test results as images rather than text, and in the latest screenshot something had started working. The habit is worth copying: for a desktop app, a picture of the screen tells you more than "tests passed".

55. Blender models and an animation from one prompt (Batraz Botsiti, @batrazbotsiti)

Batraz Botsiti (@batrazbotsiti) (14 Oct) had a Dot build Blender models and an animation from a single prompt. Because the Dot runs on its own computer, the post points out that it can repeat the job every week while the owner's machine is off. The stated aim is new game creatures every week without hiring a 3D artist. The post doesn't say how many fixes the models needed.

23. Redesigning a site you already have, by voice

56. A personal website redesigned on a 10-minute call, then rebuilt block by block in its back end (Allison Johnson, The Verge)

Allison Johnson (The Verge, 2 Oct 2026) tested a Dot she named Dotty McDotface on a $100-a-month Pro account, and says it came into its element when she gave it something fully under her control: her personal website, which she calls embarrassingly outdated. The Dot pulled together a new, more streamlined design as a prototype. Instead of typing out her changes, she called it and went through them section by section for about 10 minutes, then walked away to do the dishes until a notification said the next version was ready: not perfect, but better. More rounds followed in a browser chat on her computer and by voice-to-text in the phone app while she tidied the house. To put the design live, she gave the Dot access to her own computer. It needed what she calls a shocking number of permissions, but it worked: she watched it click around the website's back end and re-create the prototype block by block. Her verdict on the result: the new design is basic, but miles ahead of the complicated, out-of-date version it replaced.

57. A demolition company's website turned into a game on a four-hour drive (James Costello / D)

Demolition professional James Costello told OpenAI's ChatGPT Pro Community newsletter (How experts are using their dots, 15 Oct 2026) that he updated his website on a call with his Dot, named D, during a four-hour drive to work. He decided to make the landing page a game, move all the real company information to a second page, and asked for a cartoon look. He says the final game is exactly what he wanted. His reasoning: in his industry no one really looks at company websites, and now people may visit just to try the game.

More asks you can copy

A small website:

Build me a website for [purpose], with [3] sections: [list]. Host it on ChatGPT Sites, keep it visible only to me for now, and send me the link with a short list of what you'd improve next.

A dashboard for your own work:

Look through my open projects in [Slack / Gmail / Drive / these chats] and build me a dashboard that sorts them into active, waiting, and needs my attention. Keep it owner-only and update it every weekday morning.

A quick game:

Make a small browser game I can play on my phone in [5] minutes: [idea]. Keep the controls simple, send me the link when it runs, and tell me one change that would make it more fun.

Keep an existing project moving:

Here is my project: [repo or link]. Run the tests, tell me what is broken, fix what you safely can, and list anything you need from me before it is ready to [launch / submit].

Check what it built (from cases 23 and 27):

When you finish, list separately what you opened and clicked in a real browser, what you only tested with scripts or simulations, and what you couldn't check. Tell me the file name and its SHA-256 hash, and if your browser can't open the page, say so instead of reporting it as tested.

Build on the cloud computer, not mine (from case 27):

Build [project] on your own cloud computer, not on any of my computers. Before you start, tell me which tools are already installed there and what you'll need to add. When it's done, list which platforms you actually ran it on, which you only tested with scripts, and which still need a real device.

Overnight Godot or game demo on the cloud (from case 31):

Discuss [game idea] with me until the rules and feel are clear. Then build a playable Godot demo on your own cloud computer overnight (at least [10] hours). Do not use my local computer. In the morning, tell me how to open the demo, what you changed from our plan, and what still needs a real-device check.

Put something on a small device and confirm it (from cases 33–35):

I've connected [board, e.g. an ESP32 with a round display] to [computer]. Put [image or text] on its screen. If a flash fails, tell me the error and what you changed, and try again at most [2] times. When it works, tell me how you confirmed it on the device, how long it took, and what the device does and doesn't do on its own.

Plan an App Store icon test (from case 37):

My app is [name] ([App Store link]). Before we design anything, check Apple's current rules for testing an alternate app icon, including whether the icon has to ship in an approved build first. Then show me the current icon, compare [5] similar apps with visible ratings, and propose [2] icon ideas worth testing, with the steps in order in this chat. Don't submit anything to App Store Connect.

Build a business site from a brief in a folder (from case 38):

The brief for [business] is in [folder] on [my computer]. Read the markdown files there first and turn them into a build prompt for me to approve. Build the site on ChatGPT Sites and keep it visible only to me. Keep the contact form as a demo that sends nothing, and don't invent reviews or numbers. When you use a component library or skill, tell me which parts use its actual code and which you approximated. Check the site on desktop and mobile, and tell me each time you are waiting on me.

A long 3D or world build with release checkpoints (from case 46):

I'm giving you a zip with concept art and a stage-setting markdown brief. Build the 3D world from that brief on your cloud computer for the next [N] hours. Every [60] minutes: push to GitHub, cut a release, and send me a short progress note with screenshots of what changed. Don't wait for one final dump. When I say stop, write a handoff note for pulling the repo into local Unity.

A Sites workroom used as a kanban board (from case 48):

Create a private ChatGPT Site that is my workroom kanban for [project]. Columns: backlog, doing, blocked, done. Keep it visible only to me. When I message you here, update the board and tell me what changed. Host any MCP event plumbing elsewhere if Sites can't; keep the board itself on OpenAI.

Check the network before the build (from case 49):

Before you write any code for [project], test from your cloud computer whether you can reach [database host / company service / package site] on the ports it needs. Tell me which connections worked and which failed. If one failed, suggest whether to build on [my computer] instead, and wait for my answer.

A small Site with MCP that the Dot keeps up to date (from cases 52–53):

Build a private ChatGPT Site for my [shopping list / reading list], with a simple web page and an MCP endpoint you can use to add, tick off and remove items. Keep it visible only to me. Before you say it works, tell me how you signed in to test it, and if you couldn't get past my sign-in, say so instead of reporting it as tested.

Redesign the site you already have (from cases 56 and 57):

Here is my current website: [link]. Build a more streamlined prototype of a new design, visible only to me, and then I'll call you to go through the changes section by section. When I approve the prototype, apply it in my site's [website builder or CMS] on my computer, block by block, and tell me which permissions you need before you start. Don't publish anything until I've checked the result.

Related

• Use cases hub · Codex orchestration · Cloud computer · Video editing & animation · Small business
• Dots + GitHub · Voice calls · Getting started · Pricing and usage
• How to get ChatGPT Dots · Dots in the EU and UK · Features · Dots not showing or not working · Content pipeline