Big Idea

The argument in one line.

A single Claude skill backed by fal.ai can replicate every feature of a $100/month AI video platform for a few cents per generation, and self-corrects model errors without any manual intervention.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You pay monthly for Higgsfield, Runway, or a similar AI image/video SaaS and mostly use it for image gen and image-to-video.
You use Claude Desktop and are open to installing a ZIP-based skill to extend its capabilities.
You want access to best-in-class models (Flux, Cling 3, SeedDance 2, Topaz) without a subscription commitment.

SKIP IF…

You generate media at high volume — pay-per-use costs scale quickly and may exceed a flat subscription.
You need a no-setup GUI; this requires a fal.ai account, an API key, and basic Claude skill installation.

TL;DR

The full version, fast.

Vic cancelled his Higgsfield subscription and built a Claude skill called MediaGen that calls fal.ai APIs directly. The skill rewrites vague prompts into model-optimized versions, generates images for around 8 cents, animates them into video for 34 cents per 3 seconds using Cling 3, and upscales the result to 2K using Topaz for 24 cents versus a $400/year Topaz subscription. The skill also handles its own errors: when Cling 3 rejected a 10MB image, Claude compressed it and retried automatically. Installation is a free Skool community download.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:52

01 · Hook + cost pain

Higgsfield UI, pricing tiers ($19-$99+), and BANKRUPT graphic frame the problem. Vic states his credentials and teases the free skill.

00:52 – 02:43

02 · Demo 1 — text-to-image

MediaGen skill triggered in Claude Desktop. Claude rewrites a vague Superman dog prompt into a model-optimized version and generates a 4K image via Nano Banana Pro for ~8 cents.

02:43 – 04:06

03 · Demo 2 — reference image swap

Cristiano Ronaldo and Trump rapper-aesthetic photos provided as references. Claude uses Nano Banana Pro edit to transplant the look onto Ronaldo.

04:06 – 06:15

04 · Image-to-video + fal.ai tour

Cling 3 animation at 34 cents/3sec. fal.ai marketplace overview. Error handling demo: Claude auto-compresses a 10MB image and retries without user input.

06:15 – 07:54

05 · Topaz upscaling — 24 cents vs $400

Claude upscales the Cling 3 video to 2K via Topaz on fal.ai for 24 cents. Contrasted with the $400/year standalone Topaz subscription.

07:54 – 09:41

06 · Installation + CTA

Join free Skool community, go to AI Automations classroom, download ZIP, drop in Claude, load fal.ai credits, create API key. Security scan tip included.

Atomic Insights

Lines worth screenshotting.

A $100/month AI video subscription can be replaced with a pay-per-use API and a Claude skill for most casual creator use cases.
Claude rewrites vague natural-language prompts into model-optimized versions before sending them to the image generation API.
fal.ai gives access to Flux, Cling 3, SeedDance 2, and Topaz upscaling under one API key, and the same key works as new models launch.
Cling 3 charges per second of video output, so specifying an exact duration in your prompt directly controls cost.
When a model returned a file-size error, the skill auto-compressed the image and retried without any user input.
Topaz video upscaling via fal.ai costs 24 cents per use versus a $400/year standalone subscription.
Specifying the exact model name in your prompt is how you control routing inside fal.ai — the skill passes it through.
Never paste an API key into the Claude chat window — ask Claude where the ENV file is and put it there yourself.
Reference-image swaps work by providing both source face and pose reference as file paths, not just chat attachments.
A Claude skill distributed as a ZIP file can be security-scanned by Claude itself before installation.

Takeaway

Pay per use beats paying per seat.

WHAT TO LEARN

The same underlying models that power expensive AI subscriptions are available directly via API, and a well-built wrapper can handle prompt optimization, error recovery, and output routing automatically.

Most AI media SaaS platforms are wrappers around APIs you can call directly — find the underlying model and the pricing math changes dramatically.
Separating prompt capture from prompt execution with a rewrite step in between makes AI outputs more consistent without requiring users to learn prompt engineering.
Pay-per-use pricing only wins for casual use; anyone generating dozens of assets daily should model out whether a flat subscription is cheaper at their actual volume.
Agentic error handling — where the tool detects a failure, adjusts parameters, and retries — eliminates a whole category of friction that breaks most AI workflows.
Distributing a capability as an installable skill with an ENV-file API key is a low-friction way to share a workflow without building a product.

Glossary

Terms worth knowing.

Claude skill: A packaged set of instructions dropped into Claude Desktop (usually as a ZIP) that extends what Claude can do — in this case, triggering API calls to external media generation services.
fal.ai: A pay-per-use API platform that aggregates leading AI media models under a single API key with no subscription required.
Nano Banana Pro: A Flux-based image generation and editing model on fal.ai, used for both text-to-image generation and reference-image editing.
Cling 3: An image-to-video model on fal.ai that converts a still image into a short animated video clip, priced per second of output.
SeedDance 2: An alternative image-to-video model on fal.ai; generally higher quality but more expensive than Cling 3.
Topaz Video: A video upscaling model that increases resolution in existing footage; available as a $400/yr standalone app or per-use via fal.ai.
Higgsfield: A subscription-based AI video platform offering image gen, video models, and editing tools on plans from $19 to $99+ per month.

Resources

Things they pointed at.

00:00productHiggsfield ↗

05:23productfal.ai ↗

05:11productCling 3 ↗

05:11productSeedDance 2 ↗

07:20productTopaz Video ↗

01:28toolWhisperFlow

08:23linkAI Systems for Human Creators (Skool) ↗

Quotables

Lines you could clip.

01:08

“Claude goes in and writes my prompt for me in a way that it knows the image generation model will understand the best.”

Captures the core skill mechanic in one sentence, no setup needed→ TikTok hook↗ Tweet quote

06:40

“Claude recognized that, and it on its own compressed the image to a 1.4 megabyte size, and now it's rerunning the image completely on its own.”

Concrete agentic self-correction example, no jargon needed→ IG reel cold open↗ Tweet quote

07:47

“Topaz subscription for the year is like $400... yeah, bro. Literally $400 for Topaz, and I just used it for 24 cents.”

Price contrast lands instantly, no context needed→ TikTok hook↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

analogystory

00:00I rebuilt Higgs Field using Claude Code, and this is saving me personally over a $100 a month. And I was using Higgs Field actively for about six months.

00:11And really, I was just using image generation and then using some of the video models to turn those images into videos, and that's pretty much it. So now instead of paying a $100 a month or more, I'm using one API key that ties into Claude code that generates whatever I want, and I only pay for what I generate.

00:31So I actually created a Claude skill to do this. And if you stick around until the end of this video, I will give you that Claude skill completely for free. If you don't know who I am, my name is Vic.

00:40Creating content is my business. I've gotten well over a 100,000,000 views. I've made millions of dollars as a professional creator, and this channel is all about providing AI systems for human creators.

00:52So if that sounds like something you wanna be a part of, make sure you subscribe. Alright. So I got the Claude desktop app open, and I'm going to trigger my media skill.

01:03And the media gen skill can do a bunch of different things. First off, it can generate an image based just off of a description. Now the way that I programmed it is pretty interesting.

01:14So I'm able to just kinda tell Claude my general idea, and then Claude goes in and writes my prompt for me in a way that it knows the image generation model will understand the best. So let's try an example here.

01:29Generate me an image of a dog that is dressed like Superman and is also playing some sort of game console, maybe like a Nintendo Switch, make it really cool.

01:41I wanna use Nano Banana two because that's the best image generation model on the market, and I want it to be a square image and absolute max quality that you can provide. Alright. So that came out of WhisperFlow.

01:53We'll hit enter, and now Claude is cooking in the back end. And here it is rewriting my prompt for me.

01:59It says a photorealistic studio portrait of a golden retriever wearing a tailored Superman costume, the iconic red cape draped over its shoulders. So it's like literally take took my garbage prompt and made it an amazing prompt for me.

02:13And now it's running Nano Banana Pro one to one ratio at max four k resolution. ImageGen is cheap. So to generate this image, it's probably gonna cost me, like, 8¢, maybe a little less, which is way less obviously than what you would have paid if you had a Higgs Field subscription.

02:27And another really cool thing about MediaGen is that it will create a MediaGen folder in your documents, and then it will actually save everything locally for you.

02:36Okay. We got an image back, and it gave us our file path. So I'll just open that up.

02:41And here's our image of a dog in a superhero outfit playing a Nintendo Switch.

02:48And this is this is dialed. I I really like it. It came out super good.

02:52But we can also provide images to Claude as reference, and it knows how to handle these images in a way where it doesn't destroy them. So sometimes if you give a picture of yourself, you'll come out as, like, another person. This knows how to handle reference images.

03:06I've programmed it like that. I work with it a lot. So what I'm gonna try to pull off is I'm gonna take this image of Trump where he's, like, got grills on and stuff, and I'm gonna make that Cristiano Ronaldo.

03:16I want Cristiano Ronaldo with all the the grills and stuff like that. So let's see if we can pull that off here. I'll drop both images here, but I'll also have to get the file path for each of them.

03:24So I've or else it won't know where to download from, and let's give it a prompt. Okay. So the first image is Cristiano Ronaldo.

03:30The second image is Donald Trump, and Donald Trump has, like, grills and a watch and, you know, has, like, this kind of rapper thing going on about him. I basically wanna replace Trump with Ronaldo, and I want him doing the same pose.

03:44I want him having the same jewelry. I want him still looking like a rapper with the black background. Everything about the Trump image should be the same, but it should be Ronaldo instead.

03:52So, again, if you do give it an image as reference, give the file path so it knows where to source the file from, download the file, and provide it to Nano Banana. Nano Banana Pro max quality one by one ratio.

04:04So I'm always gonna tell the skill what model I wanna use to make sure it uses the right model, but it'll probably suggest the best one. When we get into video models, this matters a little more because the cost differences are pretty big.

04:16But when we're talking about images, it's pennies, so it's not a big deal. So it's using the Nano Banana Pro edit model, which allows Cloud to actually edit the images that we provided.

04:27So it has to change models, but still Nano Banana Pro. Okay. Result is coming in.

04:33And now we have the goat, Ronaldo, grilled up. He's also still wearing the suit, Trump suit, which is pretty funny.

04:40But, I mean, that looks pretty good. Okay. So Claude knows that I typically convert images to videos.

04:45That's what I like to do in some of my content. So it's saying, do you want to animate it into a video? So I'm gonna tell it for sure.

04:52Animate this into a video. Use cling three.

04:55No audio. Max quality. And only do three seconds.

04:58So the video models actually charge per second, so I would be pretty precise about how long you want these videos to be because the difference between, like, a five second video and a ten second video is a few dollars sometimes. I like to use Cling three because it's the best cheapest model.

05:14You could also use SeedDance two, but SeedDance two is definitely way more expensive. Now the really dope thing about this is that we're actually using fowl.ai as our API call, which always has the best model.

05:27So even when a new model comes out like cling four or whatever in the future, we'll be able to use it without actually changing the skill because all we gotta do is prompt it for that model. So this is fowl.ai, and you will need to go and make an account for this to work because you'll need an API key.

05:43They don't have any free credits available, so you do actually have to pay per use. But if you have, like, $10 or $20 on here, this will last you a long time if you're just generating images and videos here and there. But you can see here they have the best models out.

05:58They have seed dance two, nano banana two, cling three, happy horse, which is newer, pix verse six. They have nano banana edit, like, they have flux. They they have everything you need.

06:09So Foul AI definitely is amazing at bypassing Higgs Field. Alright. So now we know the secret sauce.

06:15Let's tell Claude to animate this video and see what the video looks like. So a really cool thing is that it will actually tell you how much it's going to cost before you run it. So it's telling me 34¢ for three seconds, one video, and it's also giving me the prompt before it runs it, and it's asking me to confirm.

06:33So I will confirm. Okay. So we did actually get an error, and that's because the specific model, cling three, that we're using can't handle image files over 10 megabytes.

06:43So Claude recognized that, and it on its own compressed the image to a 1.4 megabyte size, and now it's rerunning the image completely on its own. And that's just a limitation on cling three, which Claude handled on its own. And here is the video.

06:59Let's take a look.

07:02So just a three second animated video. It looks crazy. Just shaking head.

07:07No. Like, no, bro. You don't get it.

07:09Okay. So Kling output this at ten eighty. But what if I wanna upscale to four k?

07:15Well, I can actually do it right in this skill using the Topaz upscale video model, which is also built right into file. Alright.

07:23So I'm just gonna tell Claude upscale this to a two k resolution using Topaz Should be built right into the skill. If it strips the audio, that's fine.

07:33I don't need the audio, but if it keeps it, whatever, that's okay too. So now that's moving. So this is gonna cost us 24¢ to upscale, which is really not a lot.

07:42Topaz subscription for the year is like $400 or something. Yeah, bro.

07:47Literally $400 for Topaz, and I just used it for 24¢. Come on.

07:55And here is the upscaled video. Like, you can see, like, the quality is actually drastically better here.

08:02The detail in his hair, the detail in his face, it definitely looks like a much better render. And this comes in handy, especially if you download some, like, low quality video off of YouTube or something and you need upscaled, drop it right into Claude and say, upscale this.

08:17Like, it doesn't need to work with just your generations. It can work with any video at all.

08:22So installation on this is so easy. All you need to do is come over and join my free school community.

08:29Click classrooms up at the top and go to AI automations, and you will find this video and the file to download. It's a ZIP file. You just literally drop the ZIP file in Claude and tell it to download it.

08:41It's a skill. It will know exactly what to do it. If you don't trust it, then just ask Claude to scan it for security first.

08:47I always do that with anything I put into Claude to download. I say, scan this for security, and then I install it so you can do that if you want. And then you'll need to load up some money on FAL AI, get your API key, which is very simple to find.

08:59You literally just click your image up at the top and go to API keys and create a new API key. Claude will tell you exactly where to put that key. It's probably in an ENV file.

09:08Don't give it to Claude. Get Claude to tell you where to put it, and then it will just work, and you can do whatever you want with this skill completely for free. All that I ask is that you join my school community, subscribe to the channel, turn on notification bells because I am dropping more dope skills just like this to take your creator business to the next level.

09:28My name is Vic. Thank you so much for watching, and I will see you in the next one.

The Hook

The bait, then the rug-pull.

Vic opens on a live Higgsfield session — then cuts to the pricing page and a BANKRUPT overlay. The provocation is immediate: he is paying over a hundred dollars a month for something he rebuilt in Claude for pennies.

Frameworks

Named ideas worth stealing.

01:08model

Prompt rewrite chain

User gives a vague natural-language description; the skill has Claude rewrite it into a model-optimized prompt before sending to the image generation API.

Steal forAny AI workflow where user prompts are inconsistent — add a rewrite step before hitting the API

00:00concept

Pay-per-use stack replacement

Replace a flat monthly subscription with a single pay-per-use API key that routes to the same underlying models. Cost scales with actual usage instead of seat count.

Steal forAny subscription SaaS that is an API wrapper — find the underlying models and call them directly

CTA Breakdown

How they asked for the click.

VERBAL ASK

08:23link

“Come over and join my free school community. Click classrooms up at the top and go to AI automations, and you will find this video and the file to download.”

Soft, no urgency. Gated behind free Skool signup. Skill is the lead magnet — he asks for a subscribe and notification bell in return, not money.

MENTIONED ON CAMERA