R56 | The Antigravity / Claude Code Creative Engine

Imagine you had a creative AI assistant with access to the best image and video models in the world, trains itself on [music] elite prompting practices, masters every model, and generates content for you while you sleep. Well,

we just built a new skill that lets claude code, anti-gravity, codecs, or whatever agentic platform you use to connect and prompt any generative AI model in the world. So whether you're

after cinematic shots, UGC style videos, or product videography, you just describe [music] what you want, how many assets you need, and your agent handles everything for you. [music]

Generative AI models have gotten scary good recently. Nano Banana Pro and GPT

good recently. Nano Banana Pro and GPT Image 1.5 are top tier when it comes to generating realistic images, and Cling, VO, Sora, and Seance are best-in-class for video. But using them properly is

for video. But using them properly is where it starts to get painful. Because

first of all, for you to get results that actually match your vision, you usually need to understand complex prompts and prompting techniques. And if

you want a whole volume of content, you'll have to prompt these models manually over and over. Now, automation

tools like Wave and N8N are honestly great and they help quite a lot because they let you chain models together and batch content. But they do come with

batch content. But they do come with their own trade-offs as well. The first

one is around complexity. So workflows

like Weev become a complex spaghetti of nodes pretty quickly if you start using them for production. And the other challenge is that new model updates usually break your setup. So especially

in today's world when new AI models drop almost monthly each time you need to go back in N8N or WV and reconfigure your nodes just to keep up. And finally,

they're also not that flexible. So if

your automation workflow is already hardwired, say for vertical videos, you can just tell your workflow to give you a landscape video instead. Usually you

need to hard-code a configuration say in N8N to give you that flexibility. Well,

with the advent of agentic platforms, I think a lot of these challenges can now actually be solved. Complex prompts and techniques can live in the agents memory, which I'll show in a bit. So,

it's already trained on how you want it to prompt. And you can even train it to

to prompt. And you can even train it to be even better as you use them more. You

can also request multiple content pieces in one go without doing that manually and repeating your prompts every time.

And you also get flexibility. So if you need to use a different model, if you want to have 10 seconds instead of 8 seconds, or if you want a landscape video instead of a vertical video, you just need to ask. And so this is now

possible if you connect your agents to these generative models through this skill that I'm about to show you. So

I'll be giving you guys a demo while also walking you through the key unfair advantages of this system as we go. And

also just to mention for this lesson, I'm using clawed code in anti-gravity.

But if you want to use the default anti-gravity agents or even open AAI's codecs, which can also be used in anti-gravity, the process is pretty much the same. The first advantage is that

the same. The first advantage is that you assume the role more of a creative director instead of being the prompter.

So you're the one just reviewing and approving while your AI agent is handling the labor. So just to give one example, let's say we want to create multiple ads for this product that we want to advertise. All we need to do is to tag this specific photo, which is

this serum reference JPEG, and just ask Claude code, hey, can you create 10 ads with this product? use a mixture of GPT image 1.5 and Nano Banana Pro across

those 10 ads so that I can compare the performance of those models and before you actually do everything end to end.

Can you just show me your step-by-step process so that I can also show people what you are about to do? And I just made that prompt so that you can also understand the step-by-step process that it will follow. Because what it's about

to do is basically look at this image, set them up on Air Table so that I have proper logging of these ads and then it will actually craft the prompts for me including how much that will cost and just go ahead and generate the images

all by batch. So I'll just ask it to generate the images now. But obviously

if you want this to run end to end, sort of like a dark factory is the term that I've been hearing being used where basically the AI just runs everything for you end to end including creating the videos. That is also an option for

the videos. That is also an option for you. And just to show you real time what

you. And just to show you real time what it's doing. So it now created the 10 Air

it's doing. So it now created the 10 Air Table records for me which if you haven't used Air Table before it's basically a spreadsheet tool sort of like Google Sheets but the difference is you can preview images in here and it's also free to use and integrate into

Cloud Code or anti-gravity. And what it did for me is generate multiple unique image prompts that I can either set up cloud code to pause at this point so that I can do sort of a quality check on each of these prompts if that's

something that I want to do or in what we did here. I just ask it to create the images directly so that it's more automated. So now it's generating those

automated. So now it's generating those images. And when it's done, you can see

images. And when it's done, you can see that it generated those images for us without me even having to do anything.

So these ones look good because you can then feed this to a video model which we'll do in a bit for like UGC content.

And then you also have some images and content around product shots which you can use for your e-commerce product pages or even blow up into 4K resolution posters that you can use for your store.

And all of that has been handled by Claude Code. And a big reason why these

Claude Code. And a big reason why these image prompts are so well defined and are such good quality even without me reviewing it is because I train Claude code on a lot of best practices around prompting which I'll show how to do in a bit. And you can see I generated two

bit. And you can see I generated two images at a time just so that we have a higher probability that we'll get the image that we want. And that part is important because with these models, if you've ever worked with them, there's still a lot of components that are probabilistic. It's obviously much

probabilistic. It's obviously much better now versus how it was months ago.

But if you're working in production, it's usually often better to generate, let's say, two to three pieces of images per run, so that you can save a bit of time. And the reason why I'm okay to

time. And the reason why I'm okay to actually generate multiple volumes of images is because right now Cloud Code, the way it's generating these Nano Banana Pro images is that I have it connected to my Google AI Studio account, which in case you don't know,

for quite a long time now, Google gives you $300 in welcome credits for every Gmail account that you have. So, you're

able to use Nano Banana Pro in there and just connect cloud code to it directly, which I'll also show later how to do.

And it also gives you access, by the way, to VO3.1, which is one of the video models. So, practically, it allows you

models. So, practically, it allows you to do it for free. But now let's say we want to not just have these images as our ads but we also want them to be videos. Well with claude code or your

videos. Well with claude code or your aentic platform of choice you can simply just tell it that. So to give an example so these images look great. I think I want to turn into videos uh let's say

row 26 with that shot of the product on the marble board. And I just want you to animate these images so that they are more engaging. Include row 25 as well.

more engaging. Include row 25 as well.

And then our UGC's. I think the one that turned out really well is row 19. So

pick out let's see I think is image one better or image two. I think image one is better. So can you pick out row 19 um

is better. So can you pick out row 19 um and pick out generated image one and turn that into a UGC. So just a script based on what you know about the brand.

So you can literally see I'm just giving directions to Claude code there and just rambling about thoughts on what I think should happen. And you can see here

should happen. And you can see here because I didn't really give it instructions on what model to use. It's

asking me that right now. So because I wired it up to V3.1 and I also wired it up to cling 3.0 0. It has those options.

So, let's just use 3.1 for this one. Or

actually, let's do this. So, for the UGC ad, let's use Sor2 Pro instead. And for

the other two, which are more studio shots, let's use V3.1 for one of them and then cling 3.0 for the other. And

then I'll just type that out as a specific direction. And for the

specific direction. And for the variations per record, we can either do one or two. Let's go ahead and do two.

And that part, by the way, if you want your agent to just handle that automatically, you can also just brief Claude code to use a default model whenever you want. You can also see that I trained this claude code engine to always show me how much the cost of the

generations will be, which I think is important throughout this process. But

again, if you just want your agent to proceed with generating everything on autopilot, that's an option for you as well. You just need to tell your agent

well. You just need to tell your agent on how to do that. So now here you can see that it created the video prompts for us. And what is now saying is all

for us. And what is now saying is all three generation jobs are running in parallel. So, we'll just wait for that

parallel. So, we'll just wait for that until it comes. All right. So, now that our agent has finished with this, you can see that not only was it able to specifically pick out the images that we want to be animated, which it has done

correctly, it was also able to craft these prompts that is already pretty high quality. So, you can see it

high quality. So, you can see it provided the dialogue, provided the action, plus some camera specs around what we want, and also was able to use the model that we dictated for it. So,

if we were to preview the video that it did for us. Okay, so I have been using this Esme serum for like 2 weeks now and honestly my skin has never felt this smooth.

>> So that one is Sor 2 Pro and then for the more sort of cinematic studio product shots we use Cling 3.0 to animate this shot and VO3.1 to animate this one. [music]

this one. [music] And all of that work using different models, crafting the prompts and even calling on the models. All of that was handled by our AI agent, which is CL code in this case. And all we did was

give it direction. So you're literally the creative director. You are not the prompter in this whole equation. Now the

second advantage is that this agent system is self-learning and never obsolete. It can learn from your work

obsolete. It can learn from your work and it can also train itself on best practices. So as you work with them more

practices. So as you work with them more and you work with these agents on your setup, you can actually slowly teach them about the best practices to do it.

And just to show you that we can just ask cloud code, hey, can you share some of the best practices, principles that we worked on together to get you to a point where you can actually craft prompts in my style and using our

techniques? And what it will probably do

techniques? And what it will probably do is look at this documentation that we have around prompt best practices. And

this is a pretty long document. You can

see it's something like let's see 387 lines long that I didn't write myself. I

just had Claude code do this. But you

can see a bit of the principles around this whole document is that when it's creating UGC's, it's good to use these words in the prompt so that you get a more realistic looking character. It has

principles around being more specific, principles around skin details and some words to help with that so that it knows what to do. And then it has different prompt frameworks as well that I trained it on. So for influencers, it's using

it on. So for influencers, it's using this Bopa consistency framework. This

seal cam framework are good for cinematic or hero content and so on and so forth. And there's multiple ways by

so forth. And there's multiple ways by which you can train cla code in your way of working. Let's say you already have

of working. Let's say you already have like a database of prompts for your specific niche. Then you can just feed

specific niche. Then you can just feed that to claude code. In our case, for example, in the community, let's say we have these N8N workflows that we created in the past. What I can just do is copy all of these notes and copy them to my clipboard. I can just send them to

clipboard. I can just send them to Claude code and say, "Hey, can you take a look at this N8N workflow and gather any learnings from it on how you can prompt better? give me a download of the

prompt better? give me a download of the principles that you are able to take away so that I can pick and choose which ones I want added to your training data set. And so this is useful because each

set. And so this is useful because each of us probably has different use cases for AI. Some are into branded content,

for AI. Some are into branded content, some are probably creating it for faceless media and then maybe some are onto film making. But as you develop more the specific problems that are needed for your niche, then you can just

feed them to claude code and have it sort of absorb the best practices so that the longer that you use it, the better it becomes. So you can see here it actually provided me a few principles. Some of these I already

principles. Some of these I already trained it on. So let's say hey so some of these are already in your best practices but I think number two on text and product fidelity I want you to just add that to your best practice

references. That principle around

references. That principle around capping dialogue length and enforcing a more conversational tone. So whenever

you need a dialogue for a prompt or whenever we need a script for the video itself that is good. And number five around using a structured format like YAML is actually interesting. So let's

add those in. And so what it's now going to do is simply look at all of that text and add it to its prompt best practices markdown file, which is basically just a long text file that it can use as

reference. And the other advantage

reference. And the other advantage within this bucket is that it won't go obsolete because it's easy to update whenever new models drop. So past few days, Cance 2.0 Zero has been gathering a bit of steam and from what we know internally I think the API for this is

going to drop next Monday but the way to integrate these models so that cloud code has access to it is pretty much the same across the board and all you need to do is to go to the AI model aggregator of your choice key is one of

them file.ai is also quite popular as well as waves.ai and since cense 2.0 is not yet released let's say you want to just integrate cling 3.0 zero. For now,

you just go to that models page and then under the API section here, it used to be that you need to read through these yourself so that you understand how it works. But now with tools like cloud

works. But now with tools like cloud code and anti-gravity, you can just copy this page, which you can see here, it will copy to your clipboard a markdown or text file for large language models.

And if we just ask our agent, hey, can you integrate this API into our workspace? And then we just provide that

workspace? And then we just provide that documentation that we just copied, it's now going to do all of the heavy lifting in order to do that for us. And there

you go. it was able to absorb all of the different modes and parameters. So

whether you want to use standard or pro or whether you want to choose the duration of the video. Now you can just dictate that to cloud code and it will generate those videos for you depending on the specs that you want or you can just have cloud code decide those for

you. Now the third advantage is that you

you. Now the third advantage is that you can actually spawn multiple agents and multiple models simultaneously. And so

if you're working on big projects, if you've generated with these AI models, you know that a big component is the waiting time in order for you to review your output. So for example, let's say

your output. So for example, let's say we're creating directing for this lipstick brand and I want to create a couple of image ads and studio shots using this product, but I also want to feed Claude code this image so that it has context on the different variations

and product angles that I want. That is

something that you can do as well. So

for example, hey, can you create three image ads for this lipstick brand? It's

a premium brand, so make sure to use dark backgrounds for it. But I do want you to learn from the layout of this product shot 9x9 grid to pick out the ones that would most apply to this brand. And then if we just tag the right

brand. And then if we just tag the right references so that we can give Cloud Code more guidance there. What it's now going to do is create those images for us in the background. And while that's happening, let's say we need to create

video ads for this same product, but this time we want the video ads to feature a model just like this. Well,

what you can simply do is spin up a new agent through this plus icon and say, "Hey, can you create video ads for me?

Let's use VO3.1 as well as uh Nano Banana Pro 4Ds where essentially I like this style of the close-up ad but just have models a diverse set of models have

the lipstick being used. So we can just provide that reference for our model so that it knows what it needs to do and say we need to create new images for another product entirely. So let's say

this matcha product but we want this specific effect where you have like different products floating within space. So what we can just do, hey, can

space. So what we can just do, hey, can you create a few image ads for me featuring this matcha brand, but I do want this sort of floating objects effect where the different matcha

ingredients are floating around the AI character that we have in the middle.

Make sure to emphasize real ingredients as well as the fact that this is geared towards the Japanese target market. And

then we fire that off. And now three agents are now working for you to use those different models and create the content for you even if you step away from your computer. And now once those multiple agents have done their job, you

can see we have multiple images here of that lipstick image using Nano Banana Pro and also replicating that style that we wanted of a more close-up shot. This

one turned out quite well. You can see her pores in here. So the realism is really next level, especially when you use Nano Banana Pro. And so if you want to animate this into a talking head video, that's something that you can do as well. This one is a bit more

as well. This one is a bit more stylized, but you can see how preserved the product is, how good the skin textures are. And this one, I even

textures are. And this one, I even animated it to cling 3.0 true. same

agent. And this is the result that it came up with. [music]

And if you remember, we also requested for some studio shots. So these are the results that we got. And just to try it out, I also had that agent create and animate the videos for us.

And you can see this is really high quality, right? Cling 3.0 basically

quality, right? Cling 3.0 basically handled all of the rendering of this for us. And not only was it able to preserve

us. And not only was it able to preserve the look of it, it even preserved, if you look at our original product, that very subtle cut of the Guerilla brand in this lipstick. And it was able to do

this lipstick. And it was able to do that without me having to write any prompts. And finally, for our matcha

prompts. And finally, for our matcha example, essentially we had this product be placed in the center. and our agent decided the whole setting plus the actor and also the ingredients to put here. It

even created a second variation that features a male model. But since we asked for Japanese characters, it was also able to follow that instruction.

And similarly, when I asked for a video prompt, it was able to write that down along with using the framework that we taught it and it came up with videos like these. [music]

like these. [music] And the fourth advantage is that it's actually much easier to use and set up because you don't need any complex wiring, no node spaghettes in order to set this up. You can just describe what

you want and claude code will do your tasks for you. So if you compare it to another popular platform for creating content like we workflows are honestly quite complex to use. And the reason why

you can see there's like a spaghetti of nodes in here is because in tools like Weevi if you need to generate let's say multiple images. You need to have a

multiple images. You need to have a specific node for each and every one of them just like this. But if you're working with an AI agent you can just tell it to create like 10 images for you and it will do that automatically.

Another example is that previously to set up something like this in N8N a common pattern for running these models is that you have to set up some sort of polling mechanism wherein you pass in the request. Let's say this one is

the request. Let's say this one is creating a video with claim 3.0. You set

up a wait node in order to wait for that generation to finish and then you check if that video is already available because if not then this switch node will just pull and route this node back to processing in order to wait a bit

more. Now this is probably good for like

more. Now this is probably good for like simple workflows but what's great about working with AI agents like cloud code is that you can see that it is doing the waiting for me. So it knows that claiming 3.0 typically takes 2 to 4

minutes and it does all of that polling under the hood without me having to set anything up. And to add to that, if in

anything up. And to add to that, if in case you have multiple providers set up on your workspace like I do, you can see that it detected that wave speed actually resulted to a timeout after 10 minutes of generating a video. And so

what it's doing is use key AI instead automatically to see if it can generate the videos that way. The fifth benefit is that when you're using an AI agent system, you actually get more flexibility on models. Plus, it is also

less expensive overall. Let's take for example Higfield.ai. Now, Hicksfield is

example Higfield.ai. Now, Hicksfield is basically just a rapper on these models, but they require you to pay a monthly subscription and have actually been criticized quite a lot recently because of their sort of heavy-handed tactics when it comes to marketing their

platform. So, I won't go into too much

platform. So, I won't go into too much detail, but if you just Google Higsfield issues, then you'll find several news articles including Forbes talking about this as well. But basically, all these platforms do is just allow you to connect to these same models. And you

also have to type in the prompts yourself, which by the way, if you know where to look and with a lot of the platforms that we're using in this channel, since we don't take sponsorships from these platforms who approach us, so we can remain unbiased, you can see that KAI, for example, they

offer all of these same models at much cheaper prices. And another example I

cheaper prices. And another example I gave in the beginning is that you can actually take advantage of the Google Cloud $300 welcome credit to use Nano Banana Pro and VO. But the problem with Key is mainly on reliability issues.

They're sometimes quite slow. And so the two other providers that we usually use is FA AI as well as Wavespeed AI. I

think Wavespeed has slightly cheaper prices versus FAL. Depends on which model that you're using. But what's

great about these two platforms is that they give you so much more access to the models that are available beyond just the ones that are given by the big names like Google, OpenAI, and Clling. So if

you go to file AI's explore page, you can see they have probably more than 400 models at this point where they have like 3D models, they have different imagetovideo models. And then similarly

imagetovideo models. And then similarly for waves AI, they also have that same database of models that you can access.

And all of these by the way, if you just give the same documentation to cloud code, you can integrate it directly without you even having to set up anything. Now I think it is good to

anything. Now I think it is good to mention though that if you're using claude code, it is only available in cloud pro subscriptions. But the good thing about this subscription is it doesn't just give you access to cloud code. It also gives you access to

code. It also gives you access to co-work. It also lets you try Entropic's

co-work. It also lets you try Entropic's latest models like Opus. Or if you don't want to use Cloud Code and instead just use anti-gravity's default agent, you can actually try that for free. And if

you have a Gemini subscription to them, then apart from the other tools that Google gives you in that plan, it also gives you access to anti-gravity at more rate limits. And then thirdly, since a

rate limits. And then thirdly, since a lot of people use chat GPT, you can just use codeex. So that's something that you

use codeex. So that's something that you can also use from within anti-gravity.

And similar to the model of entropic and Google, codeex is also included in their chat GPT plus subscription. So if you're already subscribed to CHAGPT then just use codeex because that is the agentic platform that comes with your

subscription already. But the point

subscription already. But the point being if you're in the AI space most likely you may have a subscription already with chatbt in which case you can just use codeex Gemini pro in which case you can just use the anti-gravity default agents or if you have a

subscription to claude then you use cloud code. So in that way you can

cloud code. So in that way you can centralize all of your subscriptions into just one tool which is probably why a lot of people are saying that SAS platforms are losing their edge because a lot of the work that we're doing including these creative tasks are

slowly but surely being offloaded to agents like these. So there I think this is a great alternative to how things are being run now in the content AI generation space. So if you haven't

generation space. So if you haven't tried claude code or anti-gravity yet, this might be a good signal for you to test it out. And if you are setting it up for your own workspace, what you can do is to set up a framework similar to what I did using this specific

structure. And this is a very simple

structure. And this is a very simple create framework, which stands for essentially the folders that you need to organize in order for your claude code or anti-gravity to be able to work with your files better. And if I jump to my

system here, you can see I have the same thing set up where Claude, which stores all of the knowledge and commands and directives of your AI agent. So it has your API keys in this environ. And that

would also include this cloud.md which

is basically a text file that is several lines long and this contains all of the instructions that we gave cloud code so that it can effectively function as our creative AI agent. Now if you're using anti-gravity this will be a dot agent.

It functions very similarly but just think of this as the brain of your AI agent. The next one is going to be your

agent. The next one is going to be your references and these are all of the inputs that you are providing as references for your creative AI agent.

So let's say for example you have images of your brand. You may have references around the style that you want to create or recreate. You can even have multiple

or recreate. You can even have multiple product shots that you get from online so that cloud code can analyze these images and craft prompts that will recreate these images for you. And of

course this also contains your best practices for prompts which you can just continuously feed to cloud code so that it updates or self-updates this documentation for you. And then the last is the tools. And think of this as sort

of the workflows that your agent has access to. So if you've worked with nadn

access to. So if you've worked with nadn orate.com before then each of these py files which if you open them they are all just python code and you don't need to understand each of these what they basically do is take in some inputs

let's say some reference image or a specific command and depending on the complexity of your setup. What cloud

code does is just use this python tool to route let's say that this specific ad or creative you wanted to route to sora 2. This specific creative you wanted to

2. This specific creative you wanted to route to cling. These Python workflows are basically code that helps your AI agent route those tasks more effectively. And they're also the ones

effectively. And they're also the ones that update your Air Table. They're also

the ones that generate your video, do your video analysis, and basically anything that you need. And to make this easy, I shared this whole prompt right in the description below so that you can just give this to your AI agent, whether it's codeex, flood code, or

anti-gravity, and it will guide you through the setup. And also, if you're part of the RoboNuggets community, you can just download my whole setup here so that you can replicate what I have. And

when you import that to your agent, that instantly gives you all of these tools that I have here, including the connection to all the different image and video models that I featured on this channel, plus these prompt best practices that we will continuously

update as well. And lastly, just to emphasize, because these are all just Python files and markdown files, you can use it if you're using Claude Code, if you are using the native default anti-gravity agents, or if you're using OpenAI codeex, whichever subscription

that you have, you can work with it pretty much the same way. That's it for this one. I'll see you guys next time.

this one. I'll see you guys next time.

Thank you.