§ 01 · The Hook
The bait, then the rug-pull.
The hook inverts the expected question — not what Claude can do, but what it cannot — and then immediately answers it by showing a capability most people assume is off-limits: generating polished, brand-consistent Instagram carousels without any image model at all.
Where the time goes.
00:00 – 00:32
01 · Cold open — the inverted question
Host flips the usual AI framing: the interesting question in 2026 is what Claude cannot do. Introduces the carousel workflow as the example.
00:33 – 00:53
02 · Inspiration — existing carousels that perform
Shows his Practically AI Instagram page and infographic carousels that have been getting traction, which became the design reference for this workflow.
00:54 – 01:20
03 · Step 1 — Extract design system to JSON
Upload reference carousel images to Claude, run the style-extraction prompt (ignore content, capture colors/fonts/layout as JSON). The output is structured design data, not training.
01:21 – 02:35
04 · Step 2 — Set up a Claude Project
Create a new Claude Project, paste the JSON design system into project instructions, and add the carousel generation system prompt with brand-gating logic.
02:36 – 03:15
05 · Why code, not images
Explains the core mechanism: Claude writes HTML with real font names rather than rendering an image, producing accurate typography. The slides are code, exported as images.
03:16 – 03:55
06 · Step 3 — Generate carousels from content
Demo: paste a newsletter article, Claude asks for brand/account, confirms color palette, generates seven slides with progress bar. Looks clean, no errors.
03:56 – 04:13
07 · Step 4 — Download as PNG
Download each slide individually. Troubleshoot tip: if download fails, ask Claude to convert to PNG first.
04:14 – 04:38
08 · When to use Canva connector instead
If you already have a Canva template system, connect Canva inside Claude and use that. Do not rebuild what already works.
04:39 – 05:07
09 · Closing — do not overengineer with AI
The final lesson: capability does not equal obligation. Ask whether you should use AI for something, not just whether you can.
Visual structure at a glance.
Named ideas worth stealing.
00:54
model
Design System Extraction to Project Injection
- Extract visual DNA from reference images as JSON
- Store JSON in Claude Project instructions
- Generate brand-consistent output on every prompt
A two-step setup that front-loads all branding work so ongoing generation requires no design prompting.
Steal for Any repeatable content format. Works for email templates, landing page sections, slide decks.
Lines you could clip.
04:39
"If there is something you could do with AI, does it mean you should?"
00:00
"The question we should be asking is not what things you can do with Claude, but what is something Claude cannot do?"
Things they pointed at.
04:14toolCanva connector inside Claude
How they asked for the click.
04:51
subscribe
"Give a thumbs up if this video was useful, you learned something new, you feel inspired on things you can do with Claude."
Soft and conversational. Bundled with newsletter link in description. No aggressive pitch.
Word for word.
HOOK opening / re-engagementCTA the pitch analogy
00:00HOOKIn March 2026, the question we should be asking is not what things you can do with Claude, but what is something Claude cannot do? Hi. My name is Kushank.
00:07HOOKWelcome back to the channel. I'm on a mission to teach 1,000,000,000 people there for free because it should be a human right. And we just discovered that even though Claude doesn't even have an image model, but you can actually create these carousels for Instagram right inside Claude.
00:19HOOKSo even though this video is a quick tutorial on how we can actually do the same, but at the same time, I want you to use this video as an inspiration of all the things you can start exploring you could do with Claude. It's probably not the only way, but this is the workload that we used. So on my Instagram and TikTok page, Practically AI, we've been posting bunch of these, uh, quick little infographics on guides that people might need for Claude.
00:40Claude is a very hot topic. These guys are in demand, and, uh, like, these guides have performed well.
00:45So we wanted to, like, use this as our inspiration for us to create more assets just so that we know what sort of branding and design we're gonna go for. So what we did was, uh, we went over to Claude. So we uploaded all these images as reference files and just use this prompt.
00:58Extract the style, design, color, fonts, etcetera from all these images, ignore the written content, send me the description in a JSON format. So, basically, we're not really training yet on these designs just yet. We're just extracting the vibes all these designs have in common.
01:11Now when you run this prompt, you're gonna get a result that looks like some code. But it's basically a JSON format, which is a way to structure all the data that is easier for a computer to read. So now what we're gonna do this is copy this output and then create a new project.
01:24So what we're trying to do is we got Claude to understand the vibes from those designs, and now we're just gonna, like, create a project inside Claude. So every time we have to create a new carousel for Instagram, Claude will know exactly what vibes that we need for that carousel. So we can always get consistent results with basically no prompting.
01:41So if you haven't done this before, on the left side, you will see an options for projects. This is very similar to ChatGPT projects for all the people who are migrating over. And here, just describe what you're working on and what you are trying to achieve.
01:53Again, this is optional. You don't have to get into the details here. Now once you're in the project, then go to your instructions.
01:58And here, copy paste that JSON prompt that we generated. And here, we start with simple instructions on what we want this project to do. For example, you are an Instagram carousel design system.
02:07When 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 posting. Step one, collect brand details before generating any carousel. Now after this, pretty much add instructions for Clog to just use the design system that we generated in our first step.
02:25And now here we have couple extra steps, uh, and this is only because I have multiple accounts and I wanna make sure that Clot is asking us which account we're creating it for. So that way it's easy for us to create it for multiple accounts just using the same project. Now this is your mini Clot project that is trained to create those Instagram carousels for you.
02:41Now you must be wondering, how is Clot able to create these carousels if it doesn't have any image generator? So actually, what we're trying to do here is to not create an image at all. We're basically asking Clot, which is good at coding, to create the code for all those different slides, which we can download as individual images.
02:56Now, here what we can do is I can take over my newsletter content and I can just ask Claude to transform this article into an Instagram carousel. Now based on instructions, it's asking us for the brands that we wanna create it for and it's double checking the color palette we wanna use for this post and responses to just use Claude's palette.
03:11Now it just shares the palette that we have for Claude, gives us a slight breakdown, and right here, we have the code generated for all those slides. As you can see, we have seven different carousel posts that are now generated.
03:25We can toggle through. It even has like a progress bar which is like a good strategy for Instagram carousel posts that you could use and it looks very clean. All the text looks really good.
03:35I don't see any errors whatsoever. In fact, I would say it's even better than actually using like image generator just because it's actually using the right fonts inside instead of creating a font. So if you can just look into the code here, you will be able to see all the fonts that's pulling to create this carousel post instead of just generating an image here, which is what makes it even more powerful.
03:54We have all these images for carousel generated right inside our Cloud project from a very simple prompt. And now I can just go here and I can download all my slides. By the way, if you run into an error that it's not downloading for you, just ask Cloud to turn all that into a PNG file, then you should be able to download each one of the slides just from download over here.
04:11Even though it's super cool that you can generate these carousels just within Cloud, I do wanna say that if you already have a template say you work with inside Canva that, uh, you wanna use for all your carousel posts, the better strategy would be to connect the Canva connector inside Claude and just do it that way instead of trying to create everything from scratch every single time.
04:28CTAAgain, the reason I wanted to do this video is because anyone who doesn't have that templates or their design system set up already, this could be a good way for them to start right away. But, hey, like I said, you know, these days, like, with so much happening in AI, everyone likes to over engineer everything. I would say is that before you start doing everything with AI, just take a step back and just ask yourself, if there is something you could do with AI, does it mean you should?
04:49CTABecause right now with the whole hype around AI, we tend to over engineer things just because it's cool where we shouldn't. And on that note, give a thumbs up if this video was useful, you learned something new, you feel inspired on things you can do with Claude, and leave a comment if there's a video topic you want me to cover next.
— full transcript
§ 05 · For Joe
Claude writes carousels better than image generators do.
WHAT TO LEARN
Because Claude generates code with real font references instead of rendering pixels, its carousels are more accurate and more repeatable than anything an image model produces.
- Storing a brand's design system as a JSON token set in Claude's project instructions eliminates manual styling on every future generation — the brand travels with the prompt.
- Style extraction works by instructing Claude to ignore all written content and return only visual properties: color values, font families, layout approach, and spacing conventions.
- HTML-generated slides use actual font names from Google Fonts or system stacks, which means typography is crisp and legible — image generators approximate fonts and often distort them.
- A brand-gating step (asking which account before generating) lets one Claude Project serve multiple brands without cross-contamination.
- Downloading a generated slide as PNG is a two-step fallback: if the direct download fails, ask Claude to convert the slide to PNG.
- The Canva connector inside Claude is the better path when a template already exists — building a new code-based system from scratch only makes sense when no template does.
- Capability does not equal obligation: the most durable lesson in the video is to ask whether a task should use AI before asking how it could.
