How to make an Instagram carousel with AI (and what it still gets wrong)

What AI actually solves

Two things, and they are the two that stop most carousels from existing: the blank page and the shape.

The blank page is obvious. Given a topic, a language model will produce a first draft in seconds — and a bad first draft you can fix beats a perfect one you never wrote.

The shape is the less obvious win and the bigger one. A carousel is not an essay cut into pieces; it is a sequence where each card has one job. Deciding how many cards a topic needs, what goes on each, and where the argument turns is exactly the kind of structural work models are good at — and exactly what people get wrong when they write carousels by hand.

What it does not solve

The thing that makes the post worth posting: what you know and nobody else does.

A model has read what the internet says about your subject. That is why its first draft always sounds like every other post on the topic — it is, quite literally, the average of them. The number from your own data, the objection your customers actually raise, the thing you tried that failed — none of that is in the training data, because it is yours.

Which gives you the honest division of labour: the model gets you a structure fast, and you spend the time you saved putting something in it that only you could put.

The workflow that works

Four steps, in this order. Skipping the third is what produces the carousels people scroll past.

• Give it a topic, not a title. "Three mistakes freelancers make when pricing" beats "write me a carousel about pricing" — the narrower the input, the less generic the output.
• Let it decide the card count. A topic with three points needs five cards; forcing it into ten produces filler, and filler is visible.
• Rewrite the first card and one middle card by hand. The first because it decides whether anyone reads the rest; the middle because that is where you put the thing the model could not know.
• Check the promise. If the cover says "5 mistakes", count them. This is the mistake AI makes most.

The three mistakes to check for

They repeat across models, so it is worth knowing them by name.

• The broken count. The cover promises seven and the carousel delivers five, because the cover and the closing card ate two slots. Count the items, not the cards.
• The confident invention. Statistics are where models are least reliable and most fluent — a made-up number reads exactly like a real one. If a figure matters enough to put on a card, it matters enough to check.
• The flat voice. Every sentence the same length, every card the same rhythm. It reads as competent and forgettable. Breaking two or three sentences by hand fixes most of it.

Does Instagram penalise AI content?

Not for being AI. Instagram ranks on what people do with a post — how long they stay, whether they swipe to the end, whether they save and send it. A carousel written by a model that people finish will outrank one written by hand that they abandon.

The catch is that generic content is what gets abandoned, and generic is the default output of any model. So the penalty is real, but it is indirect: not "you used AI", rather "you posted the average of the internet and it read like it".

What about the images

Generating every image is usually the wrong instinct. It costs money per image, it is slow, and image models are still unreliable with text — the one thing a carousel card is mostly made of.

The cover is where a picture earns its place, because it is what stops the scroll. For the cards themselves, type and colour on a clean background outperform decoration almost every time: the card has about a second to be read, and an illustration competing with the headline spends that second.

Where this leaves you

Use the model for the draft and the structure. Spend the time it gives you back on the first card and on whatever in the post could only have come from you. Check the count, check the numbers, break the rhythm.

That is most of it. SwipeSwipe does the draft-and-structure half of this and hands you an editor for the other half — but the split above holds whatever tool you use, including none.