Article · ChatGPT · Scheduled tasks
I use scheduled tasks in ChatGPT not as a reminder, but as a tiny newsroom that puts together AI news for me every day.
Simply enough. I asked ChatGPT to prepare a news digest for me every morning, roughly between 10:00 and 11:00. The topics were broad: media, technology, careers, culture, books, Russia, emigration — but all through the lens of artificial intelligence.
The first version of the task was more of a research job. I didn’t want just a list of links. I asked it to:
So at first ChatGPT worked for me as an analyst: it read the news, selected, shortened, explained and helped me see the bigger picture.
But this news had a second audience — the readers of my Telegram channel, where it comes out under the tag #новостиИИ (AI news). So over time a digest for myself had to become a text for other people.
Then came the most useful part — tuning through feedback.
I looked at the daily issues and said what interested me and what didn’t.
It turned out I’m hardly interested in data centres, state support for IT projects, investment for its own sake or technical infrastructure. If there’s no clear effect on people, I skip it.
Once, for example, ChatGPT missed a story about Yandex’s driverless taxi test run across four Moscow districts. I showed it a morning round-up from another Telegram channel that I liked and said: this is the kind of news I need. It took nothing from it and put together a different one of its own. Yet that round-up had plenty of stories, some even more interesting, and all of them were publicly available. The example alone didn’t work, so I had to name sources outright: we added Russian applied-AI sources and named the specialised Telegram channels I follow myself.
The same happened with science news. I sent it the story of Kevin Murphy’s Model Discovery Agent — a system that doesn’t just analyse data but decides for itself which next experiment will yield the most information. After that we widened the criteria: the digest now has to include not only everyday uses of AI but also strong stories about how AI is changing the scientific method itself.
And so every day — with small remarks:
After significant edits I asked ChatGPT to remember them and update the scheduled task itself. Otherwise the edit would have stayed in a single issue.
Another layer of tuning is style.
The first digests sounded too much like a neural network: lots of generic conclusions like “AI is becoming part of our everyday life” or “this is not only a question of technology but also of privacy”.
Bit by bit I explained that I wanted a different language — closer to the Russian business press: Kommersant, Vedomosti, Kommersant FM.
That means:
Here is what it looked like in headlines:
“AI helps employees save time”
“31% of working Americans save one to two hours a week with AI”.
“OpenAI expanded its ChatGPT ad test”
“Ads in ChatGPT arrive in five more countries”.
Then I asked ChatGPT to stop handing me its internal editorial work and write a finished Telegram post straight away.
Before that the digest could contain separate blocks like:
During tuning they were useful: I could see how ChatGPT thought about the material, which patterns it noticed and what it considered important.
But once the selection logic was clear, I asked it to drop everything internal. Now every morning I get a finished text that I can publish almost immediately.
For the finished post we also tuned the order of stories. First the strongest one, with a clear human effect and numbers. Then mass consumer services, then media, the market and other weaker news. Advertising, investment and corporate reshuffles usually go to the bottom.
We also set up the technical details separately:
The tuning didn’t end with one big prompt. The main part took about a month, I think. And to be honest, I’m still editing: for a perfectionist it’s an endless process.
Gradually the task turned from a simple “find AI news every day” into a fairly detailed editorial system: with its own sources, selection criteria, hierarchy of topics, style, structure, fact-checking rules and even a sense of which stories I personally find interesting.
And that, to me, is the most interesting thing about scheduled tasks.
The perfect prompt isn’t written in one go.
I set a task, got results, edited the system itself on live examples — and after about a month ChatGPT was carrying out not an abstract instruction every day, but a workflow tuned quite precisely to me.
The topic can be anything, not just AI. The order is the same.
Tell ChatGPT in a chat what to do every day and at what time — it will offer to turn it into a scheduled task. All tasks then live in the Scheduled section.
“Why it matters”, “Trend”, “Post idea” — they show how the model selects material and where it goes wrong.
What’s unnecessary, what’s missing. Show an example — a story or someone else’s round-up you like. But one example isn’t enough: if the right news isn’t turning up, name the sources outright.
Name the outlet whose language to follow and rewrite a couple of headlines yourself. It works better than any adjectives.
Once the selection logic has settled, ask for a finished post without the internal blocks.
After a significant remark, ask: “Remember this and update the scheduled task.” Otherwise the edit stays in one issue.
Every day at [time], prepare a Telegram post for me with news about [topic] from the last 24 hours. What interests me: [3–7 topics]. What to leave out: [topics I don’t need]. Sources: [outlets and channels]. Link to the specific article, as a hyperlink on the outlet’s name. Check the date of every source. Don’t repeat stories from the last three days without a new twist. Order: first the strongest story with a clear effect on people and numbers. Advertising, investment and reshuffles go to the bottom. Style — like [outlet]: fact first, numbers, a concrete verb, no generic conclusions or filler. Don’t start a headline with a brand if you can start with the substance. Don’t repeat the headline in the lead. While we are tuning, add “Why it matters” to each story, and “Trends” and “Post ideas” at the end. When I say “drop the kitchen”, send only the finished post. Post format: up to 3,900 characters. Between stories — [divider]. At the end — one tag [#tag]. Emoji — kept to a minimum.
Replace the square brackets with your own. You’ll drop the internal blocks at step 5 with one sentence, and the task text itself will change along with your edits.