How To Make Instagram Carousels With Claude That Do Not Look Like Everyone Else's

It has never been easier to generate an Instagram carousel with AI, and that is exactly the problem. Open the app, ask for a carousel, and you get the same rounded cards, the same purple gradient, the same stock sans-serif that every other account posted this week. The tool is not the limitation. The missing input is.

The default carousel prompt is some version of "make me a cool Instagram carousel." That instruction contains no visual information at all. The word "cool" does not describe a color, a typeface, a layout, or a margin. So the model returns the statistical center of every design it has ever seen, which is the safest and most common look available.

Claude works far better when it has a reference. Not a mood word, an actual image of an actual design with actual decisions in it. Once you hand over a reference and ask for the underlying rules to be extracted, the output stops being an average and starts being a specific thing. The rest of this guide is one long version of that single move.

Go to Pinterest and search something broad like graphic design. Scroll until something stops you. That is the whole method. You are not looking for a carousel template, you are looking for a visual world: a palette, a type treatment, a feeling. In the video the pick was a soft gradient look, chosen deliberately because the previous carousels had been retro, and the goal was a different world this time.

What makes a reference work well as a source:

• It has an obvious, nameable palette. Two or three colors doing most of the work, not fourteen.
• The type has a point of view. A real display face, a real hierarchy, not default system text.
• There is visible layout logic. You can see where things sit and why, so the rules can be extracted.
• It survives being small. If you cannot read it at thumbnail size, it will not survive a phone feed.
• It is not already a carousel. Posters, album art, and editorial spreads make better sources because you are borrowing a look, not copying a post.

One honest note about taste. What works for one account will not work for another, and there is no correct answer to pick here. The only wrong move is picking nothing and describing a vibe instead. Screenshot the reference and move on.

Drop the screenshot into a regular Claude chat. Not a coding session, not anything special, the ordinary chat window in the desktop app. Then ask for two specific outputs from that image: a design system rendered as an HTML file, and a design.md file. That single instruction is the whole step, and it runs in a couple of minutes.

A design system is everything that makes up your look, written down as rules rather than vibes. In practice the returned file covers:

• The full color palette, usually with both light and dark variants, not just the three colors visible in the screenshot.
• A type hierarchy: what a headline looks like, what a subhead looks like, what body text looks like, and how they relate.
• Composition and layout rules, including how elements sit inside a slide and how much space sits between them.
• Spacing values, so slide four does not silently drift away from slide one.
• Component treatments: buttons, image handling, background construction, any repeating element.
• Generated background assets, such as the gradient fields, so you are not recreating them by hand each time.

The HTML file is for you. Open it in a browser and you can see the entire system laid out visually, which is how you sanity check whether the extraction actually captured what you liked about the reference. If the HTML page does not feel like the screenshot, redo this step before going any further, because every slide you generate later inherits from it.

This is the fork people get wrong, and it is worth being precise about. You asked for two files and they do two different jobs.

The HTML design system is the visual reference. It is the thing a human looks at. It is beautiful, it is long, and Claude does not need it to work.

The design.md file is the same system translated into language a model reads well. It decodes the main decisions behind the design and states them plainly: these are the colors, this is the type, this is the spacing, this is how a slide is composed. The design.md is the file that actually drives your outputs. It is the one that has to live in the folder you work from.

So which do you keep? Both, and for different reasons. The HTML is how you evaluate and share the look. The design.md is how you enforce it. If you only keep one, keep the design.md, because a system you cannot apply consistently is decoration.

Make a folder somewhere obvious, on your desktop is fine, and name it after the project. Open that folder in Claude Code. Unzip your downloaded design system into it and give the markdown file a clear name so there is no ambiguity about what it is.

This is the mechanical part that quietly decides whether any of this works. Claude Code reads the folder it is opened in. A design system sitting in your downloads folder is invisible. The same file sitting in the working folder gets found, read, and applied without you mentioning it every time. The folder is the context.

Before building anything real, run one throwaway test. Ask for a single image on any topic, generated from the design system in the folder. Watch what happens: it should locate the folder, find the markdown file, read it, translate it into an image prompt, and generate. In the video that test produced a cinematic editorial photograph of a humanoid robot emerging from a near-black void lit by volumetric atmosphere. Nobody wrote that prompt by hand. It came out of the design system.

Claude needs a way to make images. The setup used here is Higgsfield connected over MCP, which stands for model context protocol and is simply the standard way to give Claude access to an outside service. You connect once and every image and video model on the platform becomes reachable from your terminal instead of from a browser tab.

1. 1Go to higgsfield.ai/mcp and copy the connector URL shown there.
2. 2In Claude, click the plus at the bottom of the input, open connectors, then choose manage connectors.
3. 3Click add custom connector, give it a name, and paste in the URL.
4. 4Click add, then sign in with your account and approve access when the browser prompts you.
5. 5Confirm it worked by opening connectors again and checking that it appears in the list with its permissions toggled on.

One setting is worth changing straight away. The first time a generation runs you will be asked to approve it, and you can choose to always allow. Approving every single image by hand turns a two-minute carousel into a twenty-minute one, and you are approving the same action over and over. Set it once.

There are endless opinions about what makes a carousel perform. The structure used here is deliberately simple, and it holds up because each slide has exactly one job:

1. 1Cover: lead with the hook. This is the only slide most people will ever see.
2. 2Pain: name the problem your audience is actually having, directly, in their words.
3. 3Steps: the numbered sequence that solves it. One step per slide, no exceptions.
4. 4Result: what life looks like once the steps are done.
5. 5Call to action: join this, get this, comment this word. One ask, clearly stated.

The tempting mistake is to treat the middle slides as an essay and cram three points onto one card. Resist it. A carousel is a sequence of single ideas, and the swipe is the punctuation. If a slide contains two ideas, it is two slides that have not been separated yet.

Here is the part that generalizes far past carousels, and it is worth sitting with because it explains why the design system step matters more than the copywriting step.

A feed is a sameness machine. Everything in it is competing on the same square, at the same size, in the same two seconds. When a hundred accounts all use the default AI look, that look becomes invisible through repetition. It is not ugly. It is familiar, and familiar is the thing that gets scrolled past, because a thumb makes its decision before a single word is read.

So the first job of a cover slide is not to communicate. It is to be visually unexpected in a feed that has trained the viewer to expect one thing. That is why sourcing your look from posters and editorial design rather than from other carousels is the higher-leverage move. You are importing visual language the feed has not worn out yet.

The second job follows immediately: once the thumb has stopped, the hook has to be readable in one glance. Practically, that means a short line, one idea, at a size that survives a small screen. The design system handles the stopping. The hook handles the staying.

The third piece is the one people never plan for. Consistency across posts compounds. When ten posts share a system, the eleventh gets recognized before it is read, and recognition buys you attention that a stranger's post has to earn from scratch. That is the real return on the fifteen minutes you spent extracting a design system from a screenshot.

Doing all of the above once is a carousel. Doing it every week by hand is a chore. The fix is to turn the sequence into a skill, which sounds technical and is not. A skill is just a written-down set of steps that Claude follows the same way every time, so the output stays consistent instead of drifting with your mood and your phrasing.

The skill you want here should encode four things: the design system it reads from, the five-slide spine, the image generation step, and the caption written at the end. Then the only input you ever supply is a topic.

Two operational details that will save you an annoying half hour:

• A newly created skill does not appear in the session that created it. Restart Claude or open a new session, then type a slash and the skill name to find it.
• A well-built skill will interview you the first time it runs. Expect it to ask who the carousel is for, what your offer is, and what tone of voice to use.

That interview is the highest-value moment in the whole setup, and most people answer it and move on. Instead, answer the questions once and then ask for the skill to be updated with those answers. It never asks again. In the video the answers were plain: the audience is non-technical founders new to Claude, the call to action points at the community, and the tone was left to best judgment. Fold your answers back into the skill and you have converted a questionnaire into a default.

Each of these is cheap to prevent and expensive to discover after you have posted.

• Describing a style in adjectives instead of supplying a reference image. This is the root cause of nearly every generic output.
• Sourcing your reference from other carousels, which imports the exact sameness you are trying to escape.
• Leaving the design.md outside the folder you opened, so it is never read and the look silently reverts to default.
• Two ideas on one slide, which kills the rhythm that makes a carousel worth swiping.
• A hook that only reads at full size. Check every cover at thumbnail scale before you post it.
• Approving every generation by hand, which makes the process feel slow enough that you stop doing it.
• Answering the skill's intake questions fresh every single run instead of folding the answers into the skill.
• Changing the visual system every few posts, which throws away the recognition you were building.

This is the shortest honest path from reading to posting. It takes an evening the first time and a few minutes every time after.

1. 1Search Pinterest for graphic design, scroll until something stops you, and screenshot it.
2. 2In a normal Claude chat, drop in the screenshot and ask for a design system as an HTML file plus a design.md file. Download both.
3. 3Open the HTML in a browser and confirm it feels like your screenshot. If it does not, adjust and regenerate before continuing.
4. 4Make a project folder, unzip the system into it, rename the markdown file clearly, and open that folder in Claude Code.
5. 5Connect your image generator over MCP and set generation to always allow.
6. 6Run one throwaway image request against the design system to confirm the file is being found and read.
7. 7Have the sequence written up as a skill covering the design system, the five-slide spine, image generation, and the caption.
8. 8Restart Claude, call the skill, answer its intake questions once, and immediately have those answers folded back into the skill.
9. 9Give it a real topic. Use something you have already written, such as a lesson or a post, so the content is proven before the design is applied.

That last point is worth underlining. The carousel in the video was built from an existing lesson, pasted in as raw context. Starting from writing you have already done removes the hardest variable, because you are only testing the visual system rather than testing the idea and the system at the same time.

The carousel is the smallest thing this unlocks. A design system extracted from one screenshot is not carousel-specific. It is a description of your visual brand, and anything visual can be built against it: a landing page, an email, a newsletter template, a thumbnail, a slide deck. The site built alongside the video in this walkthrough runs on exactly the same system as the carousels, which is why they read as one brand rather than two efforts.

There is a quieter benefit as well. Once the look lives in a file rather than in your head, it stops depending on you being in a good mood or remembering what you did last month. A written system is what turns taste into output you can repeat, and repetition is what a personal brand is made of.

Everything above is runnable by hand today with a Pinterest screenshot, a folder, and a connector. Running it by hand once is how you learn which parts carry the weight, which is why the whole method is written out here in full.