How to make a carousel that does not look AI-made

You can spot an AI-generated carousel without looking at the artwork. The tell is in the structure, and it is always the same: ten cards saying the same thing in ten ways, each with a paragraph too dense for the time anyone spends there.

The good news is that this is not a model limitation. It is the consequence of asking for “a carousel about X” and accepting the first result. Five rules fix it.

1. One card, one idea

The most important rule and the most violated. If a card has two sentences that could live apart, it is two cards — or one of the sentences is redundant.

The reason is physical: people swipe. Each card gets between one and three seconds. In that time you can read a title and a supporting line. You cannot read a paragraph, however good it is.

Quick test: cover everything except the title of each card and read the titles in sequence. If the sequence already tells the story, the carousel works. If you need the body text to understand, it will be swiped past unread.

2. The cover promises what the others deliver

The cover has one job: make the person swipe once. That is all. It does not need to summarise, it does not need to teach and it does not need to be beautiful — it needs to open a gap that the next card closes.

What works is a specific, verifiable promise: “Three things we stopped doing in the oven” beats “Tips for your bread”. The first says how many, says something changed and implies someone got it wrong before. The second says nothing.

And the promise has to be kept. A cover that promises three things and delivers two and a half is the fastest way to teach people not to swipe your next carousel.

3. Five cards, not ten

EverFeed accepts 3 to 10 cards and uses 5 by default. The default is not arbitrary.

Ten cards require ten good ideas. Almost nobody has ten good ideas about a specific subject each week — what shows up from the seventh onwards is usually repetition or padding. And padding is not neutral: it teaches people that your carousels have a bad tail, and they start abandoning earlier next time.

Five cards force the choice: cover, three ideas, closing. If good ideas are left over, they become next week carousel — which is a much better problem to have.

4. The text has to fit the artwork, not the artwork the text

Here is the mistake that generates the most rework. The model writes the text without knowing how much room exists, and you get a card with a fourteen-word title that fits nowhere.

The fix is to invert the order: the template defines the space, and the text is generated with the limit in hand. A carousel card template has a schema — title with a maximum length, supporting line with a maximum length, card number. The text the copy step produces respects those limits because they are part of the request.

The side benefit is large: when text fits by construction, the card needs no manual adjustment. And since a carousel card is typography on a background, it should come from a template anyway, not from an image model — which would get the words wrong.

5. The brand appears in the right place, which is not everywhere

The instinct is to put the logo on all ten cards. The result is a carousel that looks like an ad, and ads get swiped past.

What works in practice: the brand shows up discreetly and consistently — the primary colour on the card number, the heading font on all of them, a small logo on the cover and a larger one on the last card, next to the CTA. The identity is obvious without the carousel turning into a leaflet.

That is only possible when colour and font come from the company kit and are injected at render time. A model trying to draw your logo on ten cards will draw ten slightly different logos — which is worse than having no logo at all.

The last card, which almost everyone wastes

Whoever reached the last card spent twenty seconds with you. That is the most interested person in the entire feed, and the card they see is usually “thanks for reading” or the logo on its own.

The last card should have an action, and only one: go to the link, save the post, send a message, come to the shop. Pick one. Two competing actions produce zero actions.

About the caption

The carousel caption is not a summary of the cards — people who read the cards will not read the same text again, and people who did not read them will not be convinced by a summary.

The caption works as context: why this subject now, what happened that made you write it. It is where the brand voice has the most freedom, because there is no space or layout limit. It is also where the offer with price and link goes, if there is one, and the compliance disclaimer when the post type calls for it.

How this looks in the workflow

Put together, the process is this:

1. Pick the subject — or accept one of the topics proposed from your kit, which already tells you which persona, offer or differentiator it came from.
2. Ask for five cards, not ten.
3. Read only the titles in sequence. If the story does not close, fix that before looking at the artwork.
4. Lock the copy that came out well. A locked step survives any regeneration — you can work on the artwork without risking the copy.
5. Render the cards from the template, in your colours and your font.
6. Write the caption as context, not as a summary.
7. Check the last card. One action.

Step 4 saves the most time and is the most forgotten. Without it, every attempt to improve the artwork risks the copy that was already good — and that is exactly how a good carousel becomes a mediocre one after three regenerations.