Building a AI-Powered Carousel Studio 🎨

So there's gonna be numbered stages, right?

Each one has one job.

So here on this one we've got research, then you're gonna do a script, then you're gonna produce the video.

For example, this is just a workflow, right?

This is a recipe.

So in that research you have context, which is a markdown file that just explains how things work, where they are, where to go to get what you need.

the output of this stage will become the input for the next stage.

So everything that came out of the research will then go in as the input into the script phase.

Right?

Do the research.

That research looks good.

Now we're going to go write a script.

So it will use the audience, the rules and the patterns, the references to then go through the process of drafting, what we need and then show that to me as the script to review it when the script is done.

These are what agents do and that's why we build these as folders, not agents.

So they're built to stand the test of time.

New models come along and we can apply those models, right?

So the script is now written, the output of that script, guess what?

That's going to become the input for the next stage, which is production.

So now we go into production, we've got the script that was the output.

We can look at that final script, the instructions.

It's going to be a reel at 9 by 16.

Here are the colors we want to use.

These are references.

So we've got instructions, context, and then output.

And the output is going to be the work that the model, the agent that's connected to this folder, like Claude.

This is the work that they do.

And so that comes as the output in this stage, it's production, so it's producing the final output.

Okay, so a lot of that is probably hard to follow.

So let's literally look at a folder on my computer.

This folder is called Carousel Studio.

This is a workspace is what I would call it because it does a task.

So making carousels.

So let's look at this first folder we have here is a config.

Config means, and when you see md, that just means text markdown.

So check this out.

Design spec.

This is the design spec for Laimen AI.

This has my brand MD that I've created.

I'm happy to show you guys how you make those.

This tells the size of the canvas, the colors, the type, signature elements that we're going to use.

Rules.

The important thing, I didn't write all of this.

I worked with Claude to develop this over time.

you know what my colors is, what my brand is.

I can talk to it, give it references.

And when we build this, Claude builds these folders, Claude builds this structure, and it writes all of this in this organized fashion.

So it looks complicated.

But I work a lot with the LLM to build the process.

And the LLM or AI builds a repeatable process using scripts and Python that does the exact same thing every time.

LLM's AI does not do the exact same thing every time.

So you want to use it to build the frames, build the scaffold, how you do things, but use AI, in that process, as little as possible.

Hopefully that makes sense.

So here's the design spec formula.

Here's the formula that we use.

Hook, stakes, value receipts, call to action.

Here's a template that shows the header, the top that gets filled in as a reference.

And then this shows our rules.

These things are banned.

This is our tone.

All you do is tell Claude, these are the things that I want.

These are the things that I believe.

These are my rules.

And in a conversation, you can then say, okay, let's build this.

The assets.

These are things like my logo, the word mark in black and white in case it needs to go dark or light.

some reference hooks.

Always good to show ideas, if you have them.

I don't have any on this because it's new.

And then we get into the building, the carousel.

When you open this folder, it opens at the top level, Carousel Studio.

The first thing it reads is Claude md.

This is a map.

So Claude knows, wow, this is what's in here.

This is how you navigate.

It tells cloud how it moves through here, the conventions and then the skills or tools it has.

So remember, folders, agents are just Folders with, instructions, tools and data.

And so that's kind of this Cloud MD says, hey, Claude, this is what we're working with.

And it has everything it needs to get going.

There's also a context MD in every folder.

And that's like, where am I?

What am I doing?

Claude will see this.

And it has what's called a router.

This is a.

It looks weird.

This is markdown, but honestly, this is.

If you put this into a Word Doc or Obsidian or wherever, it'll show you this as a table.

in my.

In.

In Google, in a word, Google Doc, this would be a table.

that's how this, what this formatting does.

And these are just headers.

So anyway, it says, when you need to start a new carousel, go to stages 01 brief.

And you'll also need setup questionnaire.

So the first thing it's going to do is say, okay, opened, I'm ready to start.

Where's that questionnaire?

It told me it was under setup.

Let's go.

Okay, Mark, what's the carousel about?

Who should this.

It's just doing what the folders tell it.

Hopefully this, this spells that out a little bit.

Okay, questionnaire gets answered and we're ready to go into the stages.

It says at that point now, when you're ready, I open these up.

Here, are the stages.

Brief, copy, hook, slides, export, references.

Down at the bottom is a whole nother folder.

Okay, guys, this maps directly to the example I showed you with the production, the brief, research script and production.

This is just concrete example to make a carousel.

It is walking through it being Claude or whichever agent.

I can build this with an Agents MD instead of a Cloud MD that would work in Codex.

Another trick, I can build an Agents MD that says, hey, you are gonna.

When you see this, when you open this, you're gonna go look at that Claude MD file.

just called an adapter.

So this is it.

Hopefully this helps make sense.

You want to run through it?

Let's run through it now.

All right, so this is Claude.

You see here we've gone mount the folder.

You say, open a folder.

I already have Carousel Studio, but if you click Open Folder, it'll just go into your finder here I would go to, users Mark Garza and then my Content folder.

And then in Content I have all these things.

I could go all the way down into that I want open that So there we are.

That's.

It's called mounting the folder that's mounting the project.

So this is where we're working.

So let's say, I want to make a new carousel.

Let's see if we can trigger it to do the interview.

Because I was gonna say what I wanted, but watch.

So it's Skill Loaded.

Workspace has its own site.

Let me read it.

This is how Claude goes and looks and sees what to do.

The folder is prompting It's prompting Claude.

give me these.

Whatever Shape topic, what's it about?

Audience receipt, call to action and the source material.

All right, so let's do this.

Let's say I want this carousel to be about Jev.

That's spelled J E V.

That is a new model that was just released this week.

this is going to be to people that are really new or just trying to understand the basics of AI and how to build folders to structure, them for agents.

But let's make this just focused on Jev and why we need to know about Jev for building automation.

Jev has only been around for a couple of days to the public, maybe a day or two.

So find some numbers on that that would prove why it's valuable.

Do your own research and then let's skip a call to action for this one.

For now, it's just going to be education.

So as you see, it's done the research and now it is writing the brief.

Because if we look here, the first step is brief.

We will get to that through that stage.

And when we got a brief, we say, great.

That becomes the output there, becomes the input for the next stage.

let's see what we And here we go.

So it just finished.

That was actually pretty quick.

here's the claim, here are the receipts, from the research.

Responds very quickly.

Vercel swapped a ChatGPT safety check demand crashed the API briefly after.

Well, good.

I got in and I got access.

so, okay, prove.

Sounds good.

One other trick you'll get.

Claude will try to Guess what you wanted to say next.

And a lot of times it's right.

I just hit tab enter.

All right, so pretty quickly there, you, said you saw, say, approved.

It wrote, and then boom.

That was actually probably less than a minute, maybe a minute and a half.

Here is the context.

It said, here's the hook.

Developers crashed it on day one.

Anyway, it doesn't write.

It decides it doesn't write.

That's what's really cool about this model.

We'll do another video on that too.

So here it lays it out for me.

Wrote me that brief.

Here's the caption for the, for the post.

All right, give me edits or say locked.

No, I mean, that sounds good.

needs.

So let's just say locked So I didn't do anything.

As you can see here.

It built.

It added this, Jev explained on a carousel run that's in the instructions.

And then it's got a folder for the exports.

When it does them, the images it makes.

there they are.

And then the slides when we're done.

So these images, what it's done here, it'll show us here in just a second.

But part of the instructions we have is show us three options.

Do we like this?

Do we like this?

Or do we like this?

with a stamp, but it'll show us here in a second.

See, it's building these, concepts.

Boom.

Are on brand.

But the tray concept stands out.

one.

Here you can see the alternative, the alternative images.

but it went with this one.

It looks good.

That's awesome.

So I said, we're good to go.

And then it had the pattern from what it built.

The contract for stage four, which is slides, builds the rest of them two through eight with a Python generator, which means it's done with Python.

It's called a deterministic process here because it's using code, running Python and building it out.

It's not using the LLM to figure this out.

time.

It's the same every time we run this process.

It does the same things through code.

It's codified.

So that's an important part about why we do this.

All right.

And then Claude did it.

I'm not doing anything except answering the questions or approving the output.

And it continues to build.

So if we go into the Jev explained folder, here's our slides.

This is slide one.

Came out five days ago.

It can't write a sentence, so it's been five days.

Developers crashed it on day one.

Anyway, Jev only picks from a list.

You give it.

And again, this is all using my logo, my standards, my.

If you go to Laimen AI, my website, this is it.

This is my brand standards.

And it's building up that seconds.

fine.

Fine with that.

It fell outside the box.

could, could edit that.

but for this demo purpose is fine.

But it's not perfect.

I get that.

Ten questions.

I would probably stack that you.

I would just go into Claude and say, a few of the slides have text that overlaps or outside of a container or a box.

Let's just look at it for formatting and cloud will know.

so there they are.

Unbelievable.

And so those are the export.

Once we finish, it'll go down to slides down here.

So it just kept running.

And in here we've got each of these build fully out.

What it does is it builds them as HTML and then exports them.

It's pretty powerful.

So There they are.

These are ready to export and they're now built at 1080 by 1440 in posting order.

So I can drag those wherever I need them.

Hopefully that helps.

I know it's a lot, but hopefully that helps.

horror.

If you have any questions about this, I just want to give a shot, see if I could build something a little bit longer to explain this to some of you that have been asking me for more.