How Topicly Writes a Discovery: Two Steps and an AI Review

A Topicly discovery is written by an OpenAI model in two separate steps. First comes a title, a lock-screen notification and a short "In brief" summary; the longer "full story" is written only if you tap to read it. Drafts that miss their length target are regenerated, and on the Advanced and Professional plans a second AI from a different provider reviews each draft before it is saved.

That is the short version. Below is the longer one, including the parts we are still not fully satisfied with.

#What happens before you get the notification?

Everything starts from the settings on one of your topics. Take a realistic example: you follow health, with the focus set to sleep, depth on Balanced, reading time at 3 minutes and delivery at 22:15.

About 30 minutes before that time, Topicly starts preparing. The request follows that topic's settings, and the article is written in your language. Your personal preferences are not sent into the prompt, and calls to OpenAI are made with storage turned off.

If the depth is set to All ("A thoughtful mix of every depth"), the app picks one depth for that day's slot, so Monday might be Light and Thursday Deep.

#Step one: the part you see first

The first step produces three things at once.

• A title. Short and specific, like "Your brain after lights out."
• A notification. The title is capped at 32 characters and the body at 120. The point is a notification you can read at a glance.
• The "In brief" summary. Between 90 and 150 words: one takeaway, followed by two or three short sections.

The notification limits are tighter than they look. Thirty-two characters is about five words. It forces the model to commit to one idea, which we think makes for better titles anyway.

On a busy day, "In brief" can be the whole discovery. You read it at the bus stop, you know one new thing, you are done.

#Step two: the full story, only on request

Below the summary sits a button: "Read the full story". The long article does not exist until you press it.

We split it this way for a plain reason. Most of what an AI app generates goes unread, and writing 1,000 words nobody looks at is wasteful.

Starting the quiz at the end also opens the full story.

#How long is a discovery, exactly?

You choose a reading time per topic, and each one maps to a word target:

Here is the rule we care most about: if a draft falls outside the range, it is regenerated. It is never padded to hit the number, and it is never trimmed by cutting off the end. Padding produces the filler you have seen in a thousand SEO articles. Cutting produces pieces that stop mid-argument. A full rewrite costs more, but the result reads like it was meant to be that length.

Depth works alongside length rather than replacing it. A 1-minute Deep piece and a 5-minute Light piece are both possible, and they feel quite different. There is more on choosing between them in Light, Balanced or Deep.

#Where does news come from?

Every discovery gets one of two labels: "Timeless fact" or "In the news".

Timeless pieces draw on general knowledge, the kind of thing that was true last year and will be next year. News pieces are grounded in articles that a moderator has added in the last 48 hours. There is no automatic crawler pulling in headlines. That makes news coverage narrower than a news app, and more deliberate. The difference between the two is covered in timeless facts vs news.

Some discoveries also include a picture. When they do, it is a public-domain or Creative Commons image from Wikimedia Commons, shown with its attribution.

#Who checks the draft?

This depends on the plan.

On Advanced and Professional, the draft from OpenAI goes to independent reviewers (Anthropic's Claude or Google's Gemini) before it is stored. If a reviewer rejects it, the draft is rewritten. Using models from different providers is deliberate. Two models trained differently are less likely to share the same blind spot than one model checking its own work.

Now the limit, stated plainly. This is AI review. The reviewers are language models reading a draft; they do not look up sources and compare. They are good at catching a confused explanation or a claim that sounds off. They can also both miss the same error. So we do not call any discovery fact-checked, and the reader does not show citations.

That is why every discovery carries the same line at the bottom: "This discovery was generated by AI and may contain errors. Check original sources for important decisions." We would rather you read that sentence a hundred times than trust a wrong number once. Articles also avoid personalised medical, legal or investment advice. A health discovery can explain how sleep cycles work; it will not tell you what to take for insomnia.

#What gets written after the article?

Two smaller pieces are generated from the article itself.

"Keep pulling this thread" lists three to five related concepts. Tapping one runs the same process on that concept and gives you a new discovery. From a piece about sleep you might jump to circadian rhythm, then to jet lag. How many of those you can open per day depends on your plan.

Then there is the quiz, "Three quick questions". Each is multiple choice with three to five options, drawn only from that article and written in its language. You see "That is right" or "Not quite" with a short explanation, and at the end "You answered {n} of 3 correctly". Because the questions come from the text you just read, they test what the article said, which is not always the same as what is true. If the article got something wrong, the quiz inherits the mistake. We would rather say that than pretend otherwise.

#Why go to all this trouble?

Because generating text is cheap and generating something worth three minutes of someone's evening is not. Each step above exists to push against a specific failure: the notification cap against vague titles, regeneration against filler, lazy full stories against waste, cross-provider review against a single model's quirks, and the disclaimer against misplaced trust.

None of it makes a discovery a replacement for a book or a doctor. It makes it a decent short read, most days. For the bigger picture of what the app is, see what is Topicly, or the step-by-step walkthrough on how it works.

A good way to judge the process for yourself: pick a topic you already know well and read a week of discoveries on it. You will spot what the AI does well and where it slips faster than any description could show you. Topicly is coming soon to iPhone and Android; the download page is where the links will appear.