How to Design Instagram Carousels with AI — Without a Designer

Most people learning how to design Instagram carousels with AI ask the model to make a nice-looking post and get something generic. The technique John Castaño demonstrated at Atlantis University’s AI workshop works differently: you extract a visual style into structured data first, install it as a permanent design system, and then generate unlimited on-brand carousels from any article. The setup takes one afternoon. Every carousel after that takes minutes.

The idea behind the method

The distinction that makes this work: you’re not asking AI to design something. You’re building a design system once, then feeding it content.

Most people use Claude, ChatGPT or Gemini for writing. Far fewer use them for design — and even fewer build a reusable system rather than requesting one-off outputs.

The workflow has three phases: find references → extract the style → build the project.

Phase 1 — Build a reference habit

You cannot extract a style you haven’t collected. So the first step is a habit, not a tool.

Spend 10 to 15 minutes scrolling Instagram — deliberately. Not for entertainment. You’re looking for one thing: posts whose typography, color palette and visual identity you’d want to work in the direction of.

“Not for looking at funny videos, because that’s not the idea. The idea is style, typography, color palette.”

Castaño recommends a Chrome extension for downloading carousel images in a single click, and organizing them into a dedicated inspiration folder on your computer.

⚠️ An important caveat the workshop didn’t cover. Downloading content from Instagram may conflict with the platform’s terms of service, and the images themselves are someone else’s copyrighted work. Use references to inform your own visual direction — palette, typographic feel, layout logic — not to reproduce another creator’s designs. Where possible, work from your own brand assets, licensed stock, or design systems you have rights to. The extraction technique below works identically on material you own.

Phase 2 — Extract the style into JSON

This is the step that separates the method from ordinary prompting.

Open Claude, upload your reference images, and use this prompt:

“Extract the style, design, colors and fonts from all these images. Ignore the written content. Send me the description in JSON format.”

Three things make it work:

• “Extract the style” — you want the visual system, not a copy.
• “Ignore the written content” — the text in the references is irrelevant; you only want the design language.
• “In JSON format” — this is the critical instruction. JSON is structured data. Instead of a vague paragraph describing “a warm, modern aesthetic,” you get design tokens: exact color values, font specifications, spacing rules, hierarchy.

That structured output is what makes the style reusable. Prose descriptions produce inconsistent results; tokens produce consistent ones.

One small but genuinely important detail from the demo. When copying the JSON, don’t select the text manually — use the copy button on the output block. Castaño made a point of it because he’d been burned by it: manual selection frequently misses the last lines, and truncated JSON breaks the whole system downstream. It looks trivial. It isn’t.

Phase 3 — Build the Claude Project

A Project in Claude is a workspace with persistent instructions. Whatever you put in those instructions applies to every conversation inside it — which is what turns a one-time extraction into a permanent capability.

Steps:

1. Go to Projects in the left sidebar, or navigate directly to claude.ai/projects.
2. Click + New Project in the upper right corner and name it (the demo used Atlantis University Design Project). A description is optional.
3. Open the Instructions panel.
4. Paste your JSON design tokens.
5. Above them, add the operating instructions.

The instruction set used in the demo, which you can adapt directly:

You are the [Your Brand] Instagram carousel design system. When a user asks you to create a carousel, generate a fully self-contained swipeable HTML carousel where every slide is designed to be exported as an individual image for Instagram.

Step 1: Collect brand details before generating any carousel.
Step 2: Apply the provided JSON design system to all slides.
Step 3: Ensure each slide includes branding elements and a progress bar.

Save the instructions. The system is now live.

Using it — turning an article into a carousel

Here’s where the setup pays off, and the workflow becomes almost trivially simple.

You find an article relevant to your audience — an industry announcement, a news piece, a report. You copy it. You open your project and type:

“Transform this article into an Instagram carousel.”

The output arrives as a complete, swipeable set of slides, styled according to your design tokens, with branding and a progress indicator on every slide.

The quality marker Castaño pointed to is worth understanding, because it’s what separates professional-looking carousels from amateur ones: the slides aren’t identical. In his example, the first slide was orange, the second black, the third white. Alternating background treatments within a consistent system is what makes a carousel read as designed rather than templated.

And iteration is conversational. “Change the title of the first slide.” “Change the call to action at the end — ask people to comment instead of follow.” “Write this version in Japanese.” No redesign required, because the system persists.

Frequently asked questions

Do I need a paid Claude plan?

No. Projects are available on the free plan, with a limit of five projects; paid plans (Pro, Max, Team, Enterprise) remove that cap and add enhanced project knowledge along with higher usage limits. Check claude.com/pricing for current details.

Why JSON instead of just describing the style?

Because structured data produces consistent results. A prose description gets reinterpreted differently every time; design tokens don’t.

Can I use my own brand assets instead of Instagram references?

Yes — and it’s the better approach. Upload your existing brand materials and extract the system from those. You avoid any rights questions entirely and get a design system that’s genuinely yours.

Can I do this in ChatGPT instead?

The style extraction works in any capable model. The persistent-instructions container differs by platform — Claude has Projects, ChatGPT has custom GPTs.

What if the output isn’t quite right?

Ask for changes conversationally. The project retains the design system, so revisions are cheap.

Can I generate in other languages?

Yes. Request the language you want and the design system applies unchanged.

Conclusion

The reason how to design Instagram carousels with AI trips people up isn’t the prompting — it’s that most people treat every post as a fresh request. The method above inverts that: build the system once, generate against it indefinitely.

And the pattern generalizes well beyond social media. Extract a structure into reusable form, install it as persistent instructions, then feed it content. That’s the same shape as almost every serious AI workflow.

The advice Castaño closed on was less about tools than about sequencing: structure first. Without a structure, you don’t have anything.

The Master in Artificial Intelligence at Atlantis University is built entirely on hands-on practice — you learn by building working systems, not by watching demonstrations.

👉 Book a one-on-one career development consultation

Become a student at Atlantis University