GPT DotFIELD GUIDE
Home/Workflows
Workflows

How to set up a content pipeline with ChatGPT Dots: blog drafts, scheduled posts and 38 real examples (OpenAI Dots)

What a ChatGPT Dot (OpenAI Dots) can do in a content pipeline, seven setup steps (one source, a short brief, drafts in a named place, publish rules, your own sign-ins, then a schedule) and eight places it gets stuck, with fixes. 38 real examples, from WordPress drafts approved before publishing to three Xiaohongshu posts a day.

GPT Dot · Updated
OpenAI Dots product demonstration
Image source

The short version: what a Dot can do in your content pipeline, how to set it up, where it gets stuck

Is a content pipeline worth handing to a Dot?

Yes, when you already have the material and keep the final say. In the reports below, a Dot did four things well. First, it turned one source into many pieces: Stephanie Nyarko's transcript came back as clips, a LinkedIn post, a newsletter blurb and a pinned YouTube comment; Manuel Suarez got a captioned reel cut from a webinar on his ride home; Lena got a 46-second Reel and a five-slide carousel from a video of 2 minutes 23 seconds; and Itopan turned an old article into an MP3 of about 12 minutes 43 seconds in his cloned voice, then into a waveform video. Second, it put drafts straight into the tools people publish from: a WordPress draft about ten minutes after the request for the Libecity member, a note draft with a thumbnail for Shohei Nonaka, and a newsletter test draft for Alistair Barr in about five minutes that his editor called "95% good". Third, it kept posting on a schedule under written rules: 超级峰's Dot runs three Xiaohongshu slots a day with a daily cap, and Aruru Seisakusho gets a Monday report on what went out. Fourth, it caught publishing risks nobody had mentioned: Kurt Bell's Dot checked Wix's subscriber notifications before moving old posts, and Diana's Dot told her unprompted that the page it published was private. It gets stuck on the first draft, which often reads like a template, on the last step before a post goes live, on sign-ins and uploads, and when it rewrites more than you asked. It suits people who will read every piece before it goes out; the time saved is in the gathering, cutting and formatting, not in the final read.

Setup steps

  1. Pick one job and one source, and connect only the apps that job needs (getting started; integrations). Stephanie Nyarko says "help me with my business" is a wish, not a job: she connected Google Drive, where the finished pack goes, and held off on Slack until she wanted the Dot working there.
  2. Write a short brief: the job, the audience, your platforms, two samples of your writing or your blog rules, and who decides. Stephanie's brief ends "you prepare, I decide". Namiki Kani got publishable articles only after the brief named the outcome, the sources, the condition for moving to the next step, and who acts when the work stops.
  3. Ask for one complete package as drafts in a place you name. OpenAI's creator example in Introducing dots has the same shape: clip moments, show notes and social posts drafted for your approval, with your edits carried across the set. The Meet dots guide shows the title or tone being refined from a phone before anything goes out. If you mean your own site, name it and say whether you want a draft, a private page or a public one.
  4. Review every piece for facts, quotes, timestamps and voice, and send corrections in the same conversation. Itopan's waveform video took three tries, and Luckyfairy asked twice for a badly drawn hand to be redrawn. Write the corrections back into the brief so the next package starts from them.
  5. Set the publish rules before anything goes live. Custom rules can set an action to Ask before taking action or Hand off to you (Dots privacy FAQs). Aruru Seisakusho's list is a good model: new posts only; no deleting, editing, replying, DMs or follows; no client names; ask when unsure; list each day's posts with their URLs.
  6. Do the sign-ins yourself. At Aruru Seisakusho a person took over the Dot's computer to sign in to each service, then the Dot checked it could reach every posting screen and reported that nothing had been published or scheduled. Barr entered his password through a pop-up and read out an Okta code; his employer had to approve that access first.
  7. Make it recurring only after a few packages have passed your review. Give each slot a time and time zone, a daily cap, and a rule for missed slots, as 超级峰 did with three Beijing-time slots, at most three posts a day, and no make-up posts. Keep topic research and publishing as separate tasks, and ask for a weekly report. Saved tasks are listed under Scheduled, where you can pause one or change its time (Getting started with your dot).

Where it gets stuck, and what worked

  1. The first draft reads like a template. bytenote's rewrite came back generic, Pat Simmons's first newsletter-growth summary was vague, and one of Kani's published articles had a broken structure. What worked: asking for specific tactics instead of a summary, giving writing samples, and planning one review round before anything is posted.
  2. It rewrites more than you asked, or explains instead of rewriting. 陳盟升 asked for contradictions in the site copy and got the whole site rewritten in generic text; bytenote sent a detector screenshot and got an explanation until he wrote "rewrite". What worked: say whether you want a list of suggestions or a new draft, and say that feedback means "redo it".
  3. "Publish" ends up somewhere else. Diana's review page went to ChatGPT Sites at a chatgpt.site address, private by default, not to her WordPress site. What worked: name the site and the state you want: draft, private or public.
  4. The job stops just before going live. Kani's Dot stopped at uploading images to Wix, and Kani had to ask more than once whether anything was published; Nonaka's reply stopped at the send confirmation. What worked: Nonaka checks three things separately (was it written, did it land in the draft, was it saved), and Luckyfairy's Dot opened each post signed out and compared it with the approved draft.
  5. Uploads stall or fail. One of Luckyfairy's image uploads stalled for about 25 minutes, a PNG was over the 2 MiB limit, and an image on Naver landed in the wrong place. What worked: checking size limits and file types first, converting to lossless WebP when a file is too big, and checking what is already uploaded before retrying.
  6. A sign-in or credential breaks the run. 超级峰's publishing stopped one evening because the cloud computer was missing the Feishu credential; 象先志's Dot hit a Xiaohongshu login pop-up and its research came out thin; Luckyfairy found signing in to Google on the Dot's computer from a phone clumsy. What worked: sign in yourself at setup, have the Dot stop and hand over at a captcha or expired login, and have it check existing records before picking up again.
  7. You can't see what it's doing. ひーさん's status board sometimes showed nothing in progress while work was running, and AI整理パパ found that writing "Codex" in the owner field didn't mean the job had reached Codex. What worked: have the Dot open what came back before marking anything done, label every draft with a version name and date, and ask it to refresh the board when it looks stale.
  8. Side effects outside the post. Publishing old posts could have sent Kurt Bell's subscribers one email each, and Aruru Seisakusho's Dot couldn't see posts already scheduled by another tool. What worked: the "Before you publish more than one post" prompt below, and treating proposed posting times as suggestions.

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

What it can do today:

  • Turn one recording or transcript into clip moments, posts, timestamps and a newsletter blurb.
  • Cut a captioned reel or short from your files and send it to your phone.
  • Draft a blog post and save it as a WordPress or note draft with a featured image.
  • Fill a newsletter tool from a Google Doc once your employer has approved the access.
  • Post in fixed slots with a daily cap, then report what went out each week.
  • Make carousels and cover images at a set ratio, and change one image without touching the rest, as 象先志 confirmed with checksums.
  • Check your site read-only, fix one problem with your approval, and confirm the page still loads.

What it can't do yet:

  • Write in your voice on the first try.
  • See posts you scheduled with another tool.
  • Get past captchas and sign-ins on its own.
  • Apply writing rules saved in a project every time, as Kani found.
  • Keep its own status board reliably up to date.
  • Replace your read before something goes out under your name.

A message to start with

Here is one source: [recording, transcript or voice note]. Build one package for [audience] on [platforms]: [five clip moments with start and end times, three posts, a newsletter blurb]. Use these two samples of my writing for voice: [links]. Quote me exactly and list any claim you couldn't check, with the reason. Save everything as drafts in [folder, or my site address and CMS], not on a new site, and send me the list with links. Don't post, publish, upload, schedule or email subscribers. After I've reviewed it, tell me what you'd change in the brief before we make this a repeat job with a time zone, a daily cap and an end date.

Quick answers

Can a ChatGPT Dot run a content pipeline?

Yes, from one source to many pieces. Stephanie Nyarko handed her Dot a YouTube transcript over a phone call and got a repurposing pack with clips, LinkedIn posts and timestamps. Manuel Suarez got a captioned Instagram reel cut from a webinar recording on his ride home. Dayu built a newsletter from more than 330 of their own blog posts. The setup steps are at the top of the page.

Can it publish to WordPress or post on a schedule?

Yes, once you say yes. The Libecity member's setup ends at a WordPress draft about ten minutes after the request; Luckyfairy approved each preview before the WordPress and Naver posts went live; @yamashita_pivot's Dot booked a 9 PM newsletter send only after he replied "go with this". For recurring posts, 超级峰's Dot runs three Xiaohongshu slots a day, and Aruru Seisakusho's first plan had seven X posts at 8:10 each morning. Saved repeat tasks are listed under Scheduled, where you can pause one or change its time (Getting started with your dot); ChatGPT Dots features explains schedules next to calls, apps and custom rules.

Will it publish something I haven't approved?

Not if a rule says so. Custom rules can set an action to Ask before taking action or Hand off to you (Dots privacy FAQs). The reports below show why the brief matters too: Kurt Bell's Dot checked Wix's subscriber-notification settings before publishing old posts, while Diana asked her Dot to "publish" a review page and got a private page on ChatGPT Sites instead of her WordPress site. Name the site, and say whether you want a draft, a private page or a public one.

Is the first draft good enough to post?

Often not. bytenote's first rewrite read like a template, Pat Simmons's first newsletter-growth summary was generic, and 陳盟升's Dot replaced carefully sharpened site copy with generic text. Plan one review round and write your corrections back into the brief; "What people changed after the first draft" near the end collects what worked.

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). Source files usually come from Google Drive. Fernanda Kipper also connected Gmail and her calendar before her Dot edited her Reels and found a 40-day-old unanswered email.

Real examples, by job

One recording into many pieces

Stephanie Nyarko, Dot "Remy" (YouTube, 6 October 2026). Stephanie gave Remy the job over a phone call and hung up: take the transcript of her YouTube video and turn it into a repurposing pack with clips, LinkedIn posts, and timestamps. The Dot kept working after the call ended. Her onboarding brief set one rule, "you prepare, I decide", and she checks its quotes and timestamps before she uses any of them.

Manuel Suarez, @mrmanuelsuarez (Instagram, 2 October 2026). Manuel asked his Dot to find his whole webinar in his files and the Zoom recording with screen shares, then cut an Instagram reel with burned-in captions on one topic from it. The Dot made the reel while his car drove him home and sent it back through the ChatGPT app on his phone. He downloaded it and posted that reel.

@yulhee___, Dot "닷율" (Threads, Korean, 30 September 2026). After a first phone call with the Dot, he handed over the call recording for editing, subtitles, motion graphics, and music, and asked it to upload the result to YouTube. It did not come out right the first time: alignment was off, his voice was cut, and the motion graphics were thin. He kept watching and sending feedback until the 87-second Short was done, and he says that loop is simply how he works with AI.

Lena, @lena_podzorova, Dot "Busya" (Threads, Russian, 1 October 2026). Lena had been using another assistant for a month and switched to test a Dot. Her post shows what Busya built for her: a five-slide carousel and a 46-second Reel draft cut from a 2-minute-23-second source video.

Ben Angel (YouTube, 3 October 2026). In his solo-founder walkthrough, Ben shows a Dot turning a video transcript into social content and keeping launch materials consistent with each other, among seven business jobs. He also calls the Dot to talk through an assignment and lets it continue after he hangs up.

伊藤貴將 (Itopan), Dot "シバット" (note, Japanese, 3 October 2026). Itopan went the other way, from writing to sound and picture. The Dot took one of his earlier articles and made an MP3 of about 12 minutes 43 seconds in his own cloned ElevenLabs voice, then used that audio for a video with a waveform that moves with his voice, for posting on X. The waveform took three tries: the first version was too fine, the second too small, so he asked for bigger bars and the Dot rebuilt the whole layout around big bars that clearly move with the speech. All the feedback went into the same conversation. The note itself opens with a reading in his cloned voice for anyone short on time.

Copyable ask:

Here is one recording: [file or folder]. Read the transcript and build one package for [audience] on [platforms]: five clip moments with start and end times and why each works, three LinkedIn posts, and a timestamp list for the description. Use these two samples of my writing for voice: [links]. Quote me exactly and mark anything that needs a fact check. Save everything as drafts in [folder] and send me the list. Don't post, upload, or schedule anything.

Blogs and newsletters

Manami, @AIformum (YouTube, Japanese, 2 October 2026). Manami had her Dot draft a blog article from the audio of her voice broadcast, and she hands it work from her phone while she is out with her kids.

Yuto Nonoguchi, @nonoboxer (Threads, Japanese, 3 October 2026). Yuto woke up to find his Dot had finished long, simple admin tasks and blog article drafts overnight. He says this amount of long-running work was not possible for him in agent mode, and he logged it as his first data point.

@yamashita_pivot (Threads, Japanese, 2 October 2026). He had the Dot check the body of his email newsletter, then replied "go with this", and the Dot booked the send for 9 PM that night.

Pat Simmons, Dot "László" (blog, 30 September 2026). The Dot didn't know about his newsletter from his ChatGPT history, so Pat filled it in and asked it to study how The Hustle, The Rundown, and Morning Brew grew. It wrote a 30-page PDF. The first summary was generic; when he pushed for specific tactics, several were good, and the Dot now keeps the goal in memory and sends him ideas.

안준영, @techknowedge.junan (Threads, Korean, 1 October 2026). Right after DevDay he put his Dot in charge of coordinating his work, and the list includes Naver Blog and Cafe posts, Instagram and Threads content created and published, and product listings on Smart Store, alongside paper revisions and app debugging. He still has to correct misunderstandings and ask again sometimes, but plans to keep working this way.

Kurt Bell, @softypapa (Instagram, 1 October 2026). His Dot was moving old blog posts from WordPress to Wix. Away from his desk, Kurt realized that publishing them all could flood his subscribers with notification emails. He rushed back and found the Dot already checking Wix's subscriber-notification settings before publishing anything.

Luckyfairy, Dot "미루" (HaruLabs blog, Korean, 3 October 2026). Luckyfairy handed the HaruLabs blog to the Dot and wrote up the first days: the first job was a full post on a Gemini Live screen-sharing problem, from the draft and images to publishing in WordPress, all on the Dot's cloud computer so the work could continue while Luckyfairy was out. The image upload stalled for about 25 minutes, and the post went live after midnight. The follow-up post covers the next two days: five posts across WordPress and Naver Blog, with the Dot re-checking official release dates and API prices, making a separate cover image for each channel, and converting a PNG that was over the 2 MiB upload limit to lossless WebP. Luckyfairy sent real iPhone screenshots for the parts only a person had tested, asked twice for a badly drawn hand to be redrawn, approved each post after the preview, and had the Dot request indexing in Search Console. Signing in to Google on the Dot's computer from a phone was the clumsiest step.

A Libecity member, ID Uub9Jo4b (Libecity knowledge library, Japanese, 4 October 2026). The author can't write code or HTML, so they asked ChatGPT how to set it up and then had the Dot do the work. The Dot was given the blog's rules, and now one instruction produces the body text, SEO elements, and a featured image in the blog's direction, then saves the post as a WordPress draft. From instruction to draft takes about ten minutes.

@jinsei_saisei (Threads, Japanese, 5 October 2026). An essayist on Threads, @jinsei_saisei talks ideas through with the Dot in English, and the Dot writes the finished pieces in Japanese for readers, so English practice and preparing posts happen in the same conversation.

Shohei Nonaka (note, Japanese, 6 October 2026). An executive coach who works with founders, Nonaka wrote up the first week with a Dot and had the Dot put that very article into note as a draft, with a thumbnail it made. Before Dots, Nonaka wrote in a chat and then called a skill in ChatGPT Work to upload it. In the same conversation Nonaka had it turn notes from a lunch conversation into a document for a future project, write an introduction to another company's service from research, and make illustrations, lyrics and cover images. Most of it was spoken during a 20-minute commute, in whatever order the ideas came. Nonaka checks three things separately: whether the text was written, whether it actually landed in the draft, and whether it was saved.

Namiki Kani (umibe.art, Japanese, 3 October 2026). A marketing designer and DX consultant, Kani handed a Dot the Knowledge articles for the site 海辺の部屋, built on Wix, to see whether finding material, keeping to one standard and restarting after a stop could run without a person pushing. On the first run the Dot reported about 49 minutes, three conversation threads processed, five existing articles updated and no actions needed from Kani, but no new articles, which were the point. Another run turned old conversation logs into an article and published it to Wix, and Kani found the structure broken on reading it. Writing rules saved in the project weren't always applied. What changed things was writing four boundaries into the brief: the outcome wanted, the sources to use, the condition for moving to the next step, and who acts when the work stops. After that, Kani rated the next five new articles publishable, and the Dot reported a further five finished without asking for extra approvals.

Alistair Barr, Business Insider (10 October 2026). Every issue of Barr's AI Insider newsletter means copying a Google Doc into a complicated web form, a job his editor usually does. At a Dots training, product lead Alexander Embiricos told him a Dot could likely do it from written instructions, but that his employer would need to approve that kind of access (Business Insider). Once the company gave the go-ahead, Barr connected several work apps, had the Dot read the tool's how-to guide, and gave it an old newsletter as a Google Doc for a test run. The Dot asked for his password for the tool through a pop-up and walked him through Okta two-factor sign-in, for which he read out a six-digit code from his phone. About five minutes later it sent a link to the test draft. His editor called it "95% good": the photos, a chart and all the links were in the right places. Barr expects it to save them several hours a week. The site he built in the same test is case 25 on build apps and sites.

선꾼 (@jotbugg) (X, Japanese, 14 October 2026). With little day-job access to GPT and spare tokens to burn, jotbugg had the Dot create about 100 personas, then used ChatGPT Work to write a novel those virtual readers scored. Work and the Dot became a virtual publishing house: drafts went back through reader feedback, then through another evaluation pass.

Dayu (@another_dayu) (X, Chinese, 14 October 2026). Dayu built a personal-interest newsletter with ChatGPT Dots from more than 330 of their own blog posts. The first issue already matched their taste; they said they would visit a place it recommended after returning to the UK.

Copyable ask:

Draft a blog post from this voice note: [file]. Audience: [who]. Keep my examples and opinions, cut the filler, and add subheadings. Give me three title options and a 150-word newsletter version. List any claim you couldn't check. Save both as drafts in [folder or CMS draft]. Don't publish or schedule until I reply "approved".

Social posts on a schedule

Nana Iwanaga, Aruru Seisakusho, Dot "Mr.Aruru" (note, Japanese, 2 October 2026). The small company gave its Dot the job of marketing a new book and an AI tutoring service on its blog, X, and Instagram, with a weekly report. Before starting, the Dot asked whether to publish automatically or after review, then checked the company's current posts in its own browser and returned one file with a plan and a first week of drafts: seven X posts at 8:10 each morning, two blog posts of about 3,100 characters on Tuesday and Friday at 18:30, and three five-slide Instagram carousels with roughly 330-character captions. It noticed that one post format the company had asked for had been discontinued and proposed another, and it listed the posting times as candidates because posts already scheduled by another tool weren't visible to it. A person signed in to each service by taking over the Dot's computer; the Dot then checked it could reach each posting screen and reported that nothing had been published or scheduled. The company's rules before auto-publishing: new posts only, no deleting, editing, replying, DMs, or follows, no client names, ask when unsure, list each day's posts with their URLs, and start only after approval. A Monday 8:00 report now covers what went out, the visible likes and comments, and what to change next week. Later the company added a 30-to-45-second Reel series on Tuesday and Friday at 10:30, starting with one prototype to approve first.

Steph Mui, @stephthefounder (Instagram, 30 September 2026). Steph's DevDay recap carousel covered Dots, the new model, Pages and Space, and ChatGPT in Slack, and the caption notes that the entire post was compiled and edited by Steph's Dot during a flight that day.

Viktor Malinovskyi, Dot "Byte" (LinkedIn, early October 2026). Viktor used his Dot to work on LinkedIn with him: it analysed his own posts, browsed the feed with him, and drafted a comment under someone else's post. Because they were looking at the same examples, he could stop and ask "why does this work?" before anything went out.

超级峰 (@Mileson07), a Xiaohongshu account (X Article, in Chinese, 9 October 2026). 超级峰 set up a Dot to run a Xiaohongshu image-post account on a daily schedule, one post per meme character, explaining who the character is and which work it comes from. The rules and records sit in four Feishu Bitable tables (topics and publishing, content standards, post reviews, comments). The Dot reads and writes them through Feishu's official CLI and reads each record back before it counts as saved. A separate restock task only researches, de-duplicates and adds candidates, and never posts: it keeps at least nine checked topics in stock, tops up to twelve when it falls below that, and adds no more than five per run. The publishing task has three slots, 9:30, 14:30 and 19:30 Beijing time, with one post per slot and at most three a day. A post counts against the day as soon as it is submitted, and a missed slot is not made up later. Each image is made with Codex image generation at 3:4, then checked for the right character, correct Chinese text and the ratio. The Dot then uses Computer Use on its cloud computer to upload the image, paste the text, pick five topics, check the preview and submit once; if the page hangs, it looks for the post in the dashboard instead of clicking again. After that it confirms the post is publicly visible and writes the result back to Feishu. Reviews are scheduled two, three, five and seven days and one calendar month after a post goes public. Replies to comments are only drafted, and sending them needs a separate OK. 超级峰 reports one post's views at 242, 295 and 311 on 4, 6 and 7 October, and no platform warning in five days, which 超级峰 calls a short observation of one account. One run did break: one evening and the next morning the cloud computer was missing the Feishu user credential, so the Dot couldn't read the rules or past posts to check for duplicates, and publishing stopped. Once access was back, it checked the existing records before deciding where to pick up. A captcha or an expired login also makes it stop and hand over.

ひーさん, Dot "バジ" (note, Japanese, 7 October 2026). After 2.5 days with a Dot named バジ, ひーさん published a numbers-first review. The Dot's room logged 30 task units; on the screen she checked, 26 were done and 2 were still in progress. These are work units, not thirty finished deliverables. On 6–7 October she confirmed production that reached publish or handoff states: 16 SNS images and 4 captions (Instagram: one published, one scheduled for 19:30 on the 7th; one X image still unposted); 9 ad creatives (3 concepts × 3 sizes) plus an 18-page research report; and three video drafts of the same piece cut from 14 clips (~143 seconds total), which she took back to finish herself after the BGM and on-screen credits did not fit. She had the Dot build a status board called "バジの部屋" so she could see in-progress, waiting-on-her, and done work with artifact URLs, but it sometimes showed no tasks in progress while work was running, and the display went stale until she asked for a refresh. Her prep tips for anyone starting: flesh out that status board, set clear approval points (for example show one sample first), show liked designs and say what she likes about them, and ask the Dot to review past work for her taste. She pays ¥16,800 a month and says she has not yet shown that fee recovered.

Copyable ask:

Plan one week of posts for [X, Instagram, blog] about [topic or offer], using [links to our site and recent posts]. Check what we already publish and which post formats still exist. Give me every draft in one file with a proposed time for each, and treat the times as suggestions if you can't see posts already scheduled elsewhere. Rules: new posts only; never delete, edit, reply, DM, or follow; no client names; ask me when unsure. Don't publish or schedule anything until I approve the file, then list each day's posts with their URLs.

Scripts, series, and campaign plans

Rafael Cohen, @rafaelcohenai (Instagram, Hebrew, 4 October 2026). Rafael gave his Dot a research summary and got back video ideas, a script, and a shot list. His video shows both his first request and the refinement that led to the final result.

AI at Work, Dot "Dash" (YouTube, 2 October 2026). She onboarded her Dot on camera and gave it real work: a 50-to-60-second video script, a 2-minute MC opening, research and an outline for her newsletter, and a campaign plan for a workshop launch on Maven. Dash worked several threads at once and delivered a 40-page plan with promo dates, tracking URLs, discount codes, landing-page copy, and LinkedIn copy. She estimates the plan alone would have taken her a week.

Milad Alemi, Dot "Diz" (Instagram, 30 September 2026; sponsored, #ChatGPT_Partner). Milad's ongoing job for his Dot is to keep his weekly phone-tips series researched and ready to film. The screen recordings show it researching and testing new iPhone tips, reading his Instagram comments for features people are missing, checking device compatibility, and producing downloadable PDF and Word filming packs.

Sabba Quidwai, EdD, @askmsq (Instagram, #openaidots). An AI strategy adviser, Quidwai used to spend hours every week digging through saved articles, messy notes and quote snippets just to outline one podcast episode. Now she drops her favourite articles into her Dot and records quick voice memos on her key takeaways, and the Dot sorts them in the background. Every Thursday it pulls all of it together, pitches a core theme for the episode, and uses a podcast plugin to build a script outline that is ready to record. The storytelling is still hers; what went away was the organising.

Carousels, thumbnails, and short edits

Nicky Saunders, @thisisnickys (Instagram, 5 October 2026). In her screen recordings, Nicky prompts her Dot by text and by voice call and gets back a multi-slide Instagram carousel, a social post, and a "MEET DOTS" thumbnail.

Fernanda Kipper, @kipperdev, Dot "Javito" (Instagram, Portuguese, 1 October 2026). With her computer, Gmail, Google Drive, and calendar connected, Javito edited her Reels, filtered sales leads from LinkedIn, and found a 40-day-old unanswered email and drafted the follow-up.

象先志, TMTPost (钛媒体, in Chinese, 1 October 2026). The writer gave the Dot an old article and asked it to first study how Xiaohongshu image posts are put together, then pick a piece from the content library and produce the copy and illustrations. The Dot opened Xiaohongshu, ran into a login pop-up, and worked from comparable posts it could find; the writer found that research thin, partly because the Dot had no earlier memory or prompts for this kind of job. Mid-task the writer switched topic just by sending a message, and the Dot marked it Read and carried on with the new brief. It delivered seven 3:4 images in one blue, black and yellow palette with readable Chinese text, copy ready to paste, a Word publishing pack and a zip file. A second test asked it to change only image four, replacing a general introduction with a concrete task flow, with the replacement paragraph supplied. Codex downloaded the old and new files and compared them: the other six images had identical checksums, and the copy and Word file had changed only at that spot. The writer also closed the Dot's page and came back later to find the job still running. The slow parts were a call that took at least 20 seconds to connect and a cloud computer that lagged when the writer took control.

Copyable ask:

From [source: video, research notes, or last week's post], make a 5-slide Instagram carousel and one thumbnail at [size]. Slide 1 is the hook, slides 2–4 give one point each, and slide 5 is the call to action. Use [brand colors, fonts, logo file]. Send me previews here and save the files to [folder]. Don't post.

Your website: links, SEO fixes, and copy

Fusion Events (Instagram, 5 October 2026). On a phone call, the owner asked the Dot to audit the backlinks pointing to Fusion Events, find sites worth approaching, and work out how to approach them. What came back was an audit to review, one broken backlink, and a correction email drafted with the owner's signature and left unsent for approval. The post's advice is to have the Dot separate confirmed findings from gaps in the data, and to check the source page yourself before sending anything.

aoma (blog, Japanese, 7 October 2026). aoma, an engineer, asked the Dot to research a topic and add links from existing blog posts to another page. The Dot worked out where each link fit in the existing text, and aoma reviewed the changes before they went live. It also put blog traffic and app sign-ups into a spreadsheet, which needed the source and period spelled out so the two sets of numbers didn't mix. A scheduled X post was set up but hadn't run yet when aoma wrote the post.

Luckyfairy, Dot "미루" (HaruLabs blog, Korean, 3 October 2026). In a third post, Luckyfairy had the Dot check the WordPress site from the Mac. The read-only check took 3 minutes 16 seconds and found a theme widget error, a cache file one version behind its plugin, and PHP notices. After a manual backup, the Dot applied the theme vendor's temporary fix for the widget. Updating the cache file failed because the hosting account couldn't overwrite it, so with Luckyfairy's approval the Dot explained the problem to Cloudways support chat, and the agent replaced that one file. It also fixed an ads.txt file that returned 404. Luckyfairy handled the sign-ins and noted what was left unverified.

陳盟升, @darks0603 (Threads, Chinese, 2 October 2026). 陳盟升 gave the same job to a self-built agent and to a Dot: read the AI-agent study site 陳盟升 runs and find contradictions in the copy. The self-built agent, running a 35B open-source model, flagged the conflicts and confusing parts. The Dot took much longer, listed a long set of changes, and rebuilt the whole site, replacing carefully sharpened copy with generic AI text. The verdict: the self-built agent did better because it knows its owner, not because it is smarter.

Copyable ask:

Audit [site URL] read-only: broken internal and external links, backlinks you can confirm, pages missing titles or descriptions, and copy that contradicts itself. Separate what you confirmed, with the page and evidence, from what you couldn't check. Don't rewrite pages or change settings. Give me a list of suggested fixes, in priority order, and draft one correction email for the most important broken backlink without sending it.

A Dot as coordinator: Codex writes, Claude reviews

AI整理パパ, @seiri_papa (note and X, Japanese, 13 October 2026). AI整理パパ runs note articles, X posts and affiliate work with AI, and gave his Dot the role of "COO" to take over the coordinating: the COO hands production to Codex, Claude reviews, and he only checks at the end and decides on publishing himself. Signing up for Pro didn't make the work run on its own. He couldn't tell what the COO was doing without asking each time, so he needed a dashboard showing the owner, progress, deliverables and questions for him; and writing "Codex" in the owner field didn't mean the job had actually reached Codex. So he tested one small job first: a revised note draft, related X post ideas and a table matching each fact to its source, returned privately instead of posted. The COO read the returned text instead of stopping at the "done" notice, then passed the full draft through the browser to Claude, where he was signed in, labelled with a version name and date, and checked that Claude's reply covered that version before sending the notes back to Codex. One rule backfired: a cap of two reviews per article meant a version with his own additions went straight to him without Claude seeing it. He removed the cap, and now every new full version goes to Claude before he reads it. Two things were still open when he wrote it up (X thread): when the COO asks him to confirm something, it's hard to see where to look in the document, and work doesn't yet start on its own from task notifications.

Copyable ask (from AI整理パパ's setup):

Before you mark any piece done, open what came back and list what's in it: the draft, [the extra items], and anything still unverified. Every time the text changes, including after my own edits, send the full latest version to [reviewer] with a version name and date. Tell me which version the review covers, and when you need me to confirm something, point me to the exact section. Don't publish or post anything.

Going public: what happened at the publish step

Drafting is the easy half of a content pipeline. In the reports from bloggers, newsletter writers and small businesses, most of the surprises came at the step where a draft becomes something other people can see, not in the writing.

Where the job stopped before anything went out

In almost every report, the last step before publishing was either a draft left waiting or an explicit yes from the owner. The Libecity member's setup ends at a WordPress draft, about ten minutes after the request. @yamashita_pivot read the newsletter body first, and the Dot only booked the 9 PM send after he replied "go with this". Luckyfairy approved each preview before the WordPress and Naver posts went live, and approved the Search Console index requests separately. Fusion Events got a correction email drafted with the owner's signature and left unsent. Shohei Nonaka's Dot found the right email from a screenshot and prepared a reply, which stopped at the send confirmation and hadn't gone out at the time of writing; a Canva edit was only partly done, and an app Nonaka asked for wasn't finished. Namiki Kani's Dot had five new texts checked and then stopped at uploading images to Wix, and Kani had to ask more than once whether anything had been published.

One report shows why "publish" needs a clear meaning in the brief. Diana at AI Agents Library asked her Dot, Sparky, to build and publish a standalone review page. It did, but on ChatGPT Sites: the address ended in chatgpt.site, not her WordPress site, and the page was private by default, which the Dot told her without being asked. The design was good, but it didn't follow her brand. If you mean your own site, name it, and say whether you want a draft, a private page or a public one.

Side effects nobody asked about

Publishing has consequences outside the post itself. Kurt Bell's Dot was moving old posts from WordPress to Wix, and before publishing any of them it stopped to check Wix's subscriber-notification settings, so his subscribers wouldn't get one email per old post. Kurt hadn't mentioned that risk; he was rushing back to the computer to stop it himself.

Luckyfairy's posts ran into the plumbing. An image upload stalled for about 25 minutes, and the Dot checked which files had already gone up before retrying, so nothing was uploaded twice. A cover image in PNG was over the site's 2 MiB upload limit, so it was converted to lossless WebP. On Naver, an image landed in the wrong place and had to be undone and moved to the top of the post. These are small problems, but each one would have stopped an unattended run.

Checking the live page, not the draft

The most careful check in these reports came after publishing. In Luckyfairy's runs the Dot opened the post without being signed in, confirmed that all five images showed, compared the published title and body against the final draft, and checked each channel's cover. Search Console first reported the new WordPress addresses as unknown to Google, so the Dot ran a live test before requesting indexing. Luckyfairy also noted what wasn't checked: the post on a real phone, and whether Google actually indexed it.

Fixing a site went the same way. In Luckyfairy's third post, the read-only first check took 3 minutes 16 seconds, and a manual backup came before any change. When replacing one cache file failed because the account couldn't overwrite files owned by the app, the Dot didn't reset permissions across the site. With approval, it asked the host's support team to replace that single file, then confirmed the versions matched byte for byte and that the home page still returned HTTP 200. aoma, likewise, reviewed every link change before it went live.

Where the first draft was weak

Two patterns show up in the research behind the content. Pat Simmons found the Dot's first newsletter-growth summary generic; the useful part came after he asked for specific tactics. Fusion Events asked for confirmed findings to be kept apart from gaps in the data, and Lee Shand, whose Dot is called Purple, got a daily note that labelled itself a partial day, covering until about eleven in the morning. That same Dot couldn't reliably open some Substack pages in its cloud browser, but could read them through Lee's signed-in Chrome on his own computer. If your sources sit on sites that block cloud browsers, see the cloud computer page.

字节笔记本 / bytenote (blog, Chinese, 2 October 2026). The author gave his Dot source material, an original draft and his requirements and asked it to rework an article. The first version came back fast but read like a template. He then sent a screenshot of an AI-detection result, expecting a rewrite. Instead the Dot explained the screenshot: which sentence patterns looked machine-written, which structures were too formulaic, and a list of suggested fixes. It only started rewriting after he added one word, "rewrite". Meanwhile its proactive suggestions about other connections and tasks kept adding decisions while the first job was still open. He stopped using it for now, and the test he proposes is how long you can leave a job before you need to look at it again.

Copyable ask (from bytenote's report):

Rework [article or draft] using [source material] and these requirements: [requirements]. If I send you a screenshot, a detector score, or a comment about the draft, treat it as a request to redo the draft, not to explain it: rewrite first, then add at most three lines on what you changed. Hold any unrelated suggestions until this draft is finished.

Copyable prompts from these reports

Each prompt below combines at least two of the reports above and names the problem it is meant to avoid. Replace the parts in brackets.

Draft only, in my CMS (from the Libecity member, @yamashita_pivot and Luckyfairy; avoids a post or newsletter going out before you've read it):

Write [the post] using my blog rules in [doc or memory], and save it as a draft in [WordPress / Wix / Substack] on [my site address], not on a new site. Don't publish, schedule, or send anything to subscribers. Send me the preview link and a list of the facts, prices and dates you used, with a source for each. I'll reply "publish" or "schedule for [time and time zone]" when it's ready.

Before you publish more than one post (from Kurt Bell and Luckyfairy; avoids an email blast to subscribers and duplicate uploads):

Before you publish or import more than one post, check and tell me: whether subscribers get an email or notification for each post, whether posts are shared to social accounts automatically, and whether any of these posts already exist on the site. Wait for my answer before publishing anything. If an upload stops partway, check what is already there before trying again.

Images that will actually upload (from Luckyfairy and Diana's ChatGPT Sites test; avoids failed uploads, misplaced images and off-brand visuals):

For each image, check the site's upload size limit and file type before uploading. If a file is too large, convert it without visible quality loss (for example lossless WebP) and tell me. Use my brand colours and the reference images in [folder]. Show me every image, including hands, faces and any text in it, before it goes into the post, and tell me where in the post each one goes.

Check the live page (from Luckyfairy and Lee Shand; avoids a post that looks fine in the editor but is broken in public):

After it's published, open the page while signed out. Confirm the title, body and every image match the approved draft, and list anything that differs. Then tell me what you could not check, such as the page on a real phone or whether Google has indexed it. Ask me before requesting indexing in Search Console.

Fix one thing on my site, safely (from Luckyfairy's site repair, aoma and Fusion Events; avoids a repair that changes more than it should):

Check [site] read-only first, and give me the list of problems with evidence, keeping what you confirmed separate from what you couldn't check. Don't change anything yet. When I pick one fix, make a backup first and change only what that fix needs. If you hit a permission problem, stop and tell me rather than changing permissions. Ask before contacting the host's support team, and after the fix, show me the page still loads.

For how approvals work in general, see Approvals in practice. For codes and sign-ins during a publishing run, see where first jobs stall at sign-ins. For ads and outreach beyond content, see small business, and if a run stops with an error, start at Dots not showing or not working.

What people changed after the first draft

Most of these pipelines got good on the second pass, not the first. Rafael's video shows the refinement that produced his final script. @yulhee___ went through several rounds on alignment, cut audio, and motion before the Short was done. Pat Simmons only got useful newsletter tactics after he asked for specifics instead of a summary, and 안준영 still re-asks when the Dot misreads a task. Luckyfairy checked every image before publishing and caught the same drawing problem twice. When 陳盟升 asked for contradictions in the site copy, the Dot rewrote the whole site instead, so ask for a list of suggestions when you want to keep your own words, and say so in the brief. AI整理パパ dropped a two-reviews-per-article cap after a version with his own additions skipped review; send the latest full version for review every time instead of counting rounds. Plan for one review round and write your corrections back into the brief, so the next package starts from them.

Small-business marketing beyond content, such as ads and outreach, is on small business. Watching outside sources for topics belongs on research watch, and turning a long call into drafts is covered on voice calls. When you want the video itself made, from edits and launch videos to Blender 3D and avatar work, see video editing & animation.

Related

Copy this task brief

38 real ChatGPT Dots content examples: recordings turned into clips and posts, blog posts published to WordPress, a week of scheduled social posts, and site link and SEO fixes, plus copyable asks.
Download Markdown ↓

Sources & further reading