Making a LinkedIn Carousel With AI

Most advice about using AI for carousels skips the boring part: knowing which half of the job to hand over. If you ask a chat assistant for “a LinkedIn carousel about pricing”, you get ten slides of confident, generic filler. If you ask it to argue one specific point in eight steps and then cut every slide to nine words, you get something worth designing. The difference is not the model. It is where you point it.

This page walks through a workflow that holds up in practice: outline and write with AI, design in a real tool, then publish. It also says plainly where AI gets in the way, because that part usually gets left out.

What is AI actually good at when making a carousel?

Three things, and they happen to be the three that take the longest by hand.

Sequencing an argument. A carousel is not a list of facts — it is one idea broken into steps that each earn the swipe to the next. That is a structural problem, and language models are good at structure. Give one a claim and a rough pile of supporting points, and it will order them into a sequence far faster than you will on a notepad. It is also good at spotting the step you skipped, which is usually the reason a draft feels like it jumps.

Cutting words. This is the single biggest quality lever on a carousel and the thing people hate doing. A slide with thirty words is a slide nobody reads on a phone. Ask an AI to rewrite a slide in under twelve words without losing the meaning, and it will do it, repeatedly, without getting precious about the sentence it just wrote. Do the same edit yourself and you will defend your own phrasing.

Generating alternatives. Ten openings for slide one. Three ways to frame the same middle section. Four different endings. You are not asking the AI to pick the best one — you are looking at ten options in thirty seconds instead of writing three in twenty minutes. The judgement stays yours; the raw material gets cheap.

Underneath all three is the real reason to bother: AI removes the blank page. Editing something mediocre is far easier than starting from nothing, and a carousel has eight or nine blank pages in it.

Where does AI fall short?

Design judgement. This is the honest limit. An AI can tell you a headline should be large and the body text should not compete with it — that is written down in a thousand places it has read. What it cannot do is look at your actual slide and see that the heading is one word too long so it wraps badly, that your logo is sitting exactly where the swipe indicator lands, or that slide four looks fine on a laptop and turns to soup on a phone. Describing a layout in words is not the same as seeing it. Even models that can look at an image will happily tell you a slide is “clean and professional” when it is neither.

Facts and numbers. Ask for a statistic to open a carousel and you will very likely be handed one that sounds exactly right and does not exist. This is the fastest way to damage your own credibility in public, because a carousel is a screenshot-friendly format and a wrong number outlives the post.

Your specific experience. The parts of a carousel that make people stop scrolling are the parts only you have: what happened on your project, the mistake you actually made, the client objection you keep hearing. AI cannot invent those, and when it tries, the result reads like every other post in the feed. The pattern that works is to give the AI your raw specifics and ask it to structure them — not to ask it for the specifics.

Sounding like you. Left alone, AI writes in a particular register: tidy, balanced, faintly promotional, fond of three-part lists. It is readable and it is forgettable. You fix that by feeding it your own writing as a sample, or by keeping a saved description of your voice that you reuse every time.

What does the workflow look like, step by step?

Five stages. The AI does the heavy lifting in stages one to three, almost nothing in stage four, and helps again in stage five.

Stage 1 — pick the one idea yourself

Write the claim in one sentence before you open a chat window. “Most teams pick a project management tool before they have agreed how they actually work.” That is a carousel. “Project management tips” is not. If you cannot write the sentence, the AI will cheerfully fill the gap with something vague, and you will spend the rest of the workflow polishing a post that was never about anything.

Stage 2 — outline with AI

Give it four things: the claim, who it is for, roughly how many slides, and any raw material you already have. Ask for one line per slide and nothing else. Something along the lines of: Here is my claim, my audience, and six things I have learned about it. Give me an eight-slide outline, one short line per slide, that builds this argument in order. No design suggestions, no emoji, no filler slides.

Then push back once, before you write a single full sentence. Ask which slide is weakest and why. Ask what a sceptical reader would object to, and whether that objection is answered anywhere in the sequence. That second pass usually improves the outline more than the first pass produced it.

Stage 3 — tighten the words

Now go slide by slide. The instruction that works is a hard limit: rewrite each slide in twelve words or fewer, keep the meaning, no cleverness. Then read the whole set aloud in order, with nothing else on screen. If it still makes sense read straight through, the sequence works. If you have to reconstruct the argument in your head as you go, a step is missing.

This is also where you catch the AI-isms: the “in today’s landscape” opener, the neatly symmetrical phrasing, the closing slide that says nothing at all. Cut them.

Stage 4 — design in a real tool

Take the finished text into a design tool and lay it out yourself. This stage is manual on purpose. Keep one idea per slide, keep the text large, and check it at phone size before you export — that is where most carousels quietly fall apart.

Export as a PDF if you want LinkedIn’s swipeable document view, or as separate image files if you want a multi-image post. Those are two genuinely different post types on LinkedIn, and it is worth deciding which one you are making before you export rather than after.

Stage 5 — caption and publish

The caption is a separate piece of writing, not a summary of the slides, and AI drafts it well from the outline you already have. Then publish. You can upload in LinkedIn’s own composer, or — if you are already sitting in a chat window — publish straight from there.

That last option is what a connector does. Kalovio is an MCP server (Model Context Protocol — an open standard that lets an AI assistant use outside tools) that connects your own LinkedIn account to Claude or ChatGPT. You upload the exported PDF or the images to it, then ask the assistant to publish the post with them. It returns a preview and waits for your confirmation before anything goes live, so the assistant is never posting on its own.

Should I let AI design the slides for me?

You can ask an image model to produce carousel slides, and for a purely decorative background or an illustration that is fine. For slides that carry text, it is usually a poor trade. Generated text inside images tends to come out subtly wrong, you cannot fix a single word without regenerating the whole image, and very little stays consistent across nine slides — which is precisely what a carousel needs most.

The more reliable version of “AI designs it” is a template that you or someone else built once, filled with AI-written text. You get the speed without giving up control of the layout, and slide three still looks like it belongs beside slide two.

How do I stop the result sounding like AI wrote it?

Four habits, in rough order of how much they help:

• Feed it your own writing. Paste two or three of your past posts and ask it to match the rhythm and vocabulary. This does more than any instruction about “tone of voice” ever will.
• Save a voice description and reuse it. A short written description of how you write — sentence length, words you avoid, how blunt you are — applied to every draft keeps things consistent between sessions. Kalovio stores one as a voice profile so you are not re-explaining yourself each time; a saved prompt or a note file does the same job.
• Put one specific thing on every slide. A real number from your own work, a real objection, a real tool name. Generic slides are exactly where AI writing shows.
• Rewrite the last slide yourself. Closing slides are where models default hardest to filler, and the last slide is the one that decides whether anyone does anything.

Is using AI for carousels actually worth it?

For the outline and the copy, yes. It turns an hour of staring at a blank page into fifteen minutes of editing, and editing is the easier task by a wide margin. For the design, not really — a template and your own eyes beat a prompt. And for anything factual, treat every AI output as a claim to check rather than a fact to publish.

Split that way, AI takes the carousel work you dislike and leaves you the part where your judgement is the whole of the value.