Big Idea

The argument in one line.

A single self-built interface wired to a handful of raw model APIs can replace an all-in-one AI creative subscription, because the paid platform's real product is just prompt refinement, a model catalog, and a billing wrapper around infrastructure anyone can call directly.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A solopreneur or small business generating a handful of AI images or videos a month who's paying for an all-in-one subscription they barely use.
Someone comfortable directing an AI coding assistant like Claude Code, Codex, or Cursor to build and maintain a small internal tool.
Anyone who wants to see, mechanically, what a $40-200/month AI creative subscription is actually doing behind its interface.

SKIP IF…

A creative team that needs a polished, ready-made, multi-user tool today and has no interest in maintaining custom software.
You've never used an AI coding assistant and have no interest in starting.

TL;DR

The full version, fast.

Bench Studio is a local interface the creator built that routes to 37 image and video model APIs, refines prompts per model, tracks exact per-generation cost, and exposes the same tools to Claude Code and Codex through MCP. He demos turning a product photo into a synced-dialogue UGC ad with Veo 3.1, then reverse-engineers what a subscription platform like Higgsfield is actually doing: accepting reference media, meta-prompting a rough request for the chosen model, routing to a model library, and exposing only that model's parameters. He lays out two paths to rebuild it yourself: use an aggregator like fal.ai or kie.ai for one API across many models, or feed a provider's own docs to an AI coding assistant and call each model directly, trading convenience for lower markup and full cost control.

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 – 00:56

01 · The creative studio replacing another subscription

Cold open naming Higgsfield directly; the creator shows video, product shot, website, slide deck, and report output all from one interface he built and owns end to end, including cost breakdown and MCP access.

00:56 – 01:49

02 · Turning a product image into a UGC ad

Finds a demoable product photo (a pizza oven), drags it into Bench Studio, and sets it up as a UGC ad using existing off-the-shelf applications and APIs.

01:49 – 02:18

03 · Model-specific prompt refinement

Types a rough description of the desired UGC actor and reaction, then runs 'refine prompt' so AI rewrites it specifically for the selected model before generating.

02:18 – 03:41

04 · The generated Veo video and exact cost ledger

Veo 3.1 Fast generates the UGC clip; results grid shows every past generation with its exact billed cost logged next to it, giving full spend transparency.

03:41 – 04:24

05 · Using Bench Studio through Claude Code and Codex

The same generation tools are called from inside a Claude Code chat via MCP: inspecting model inputs and generating media from a plain-English request, no separate app needed.

04:24 – 04:54

06 · Websites, PDFs, reports, and other creative outputs

The same routing pattern extends past images and video to 3D websites, well-formatted PDF reports, and slide decks generated inline.

04:54 – 05:24

07 · What Higgsfield is doing behind the interface

Pulls up Higgsfield's own site and starts reverse-engineering it: for a solopreneur making a handful of assets, the full subscription toolkit is more than needed.

05:24 – 06:19

08 · Reference inputs and prompt optimization

Breaks down step one and two of any aggregator: what media a model accepts as input, and how a vague user prompt gets meta-prompted for that specific model.

06:19 – 07:26

09 · Model libraries, parameters, and routing

Covers step three and four: the model library aggregators maintain for you, and how each model exposes a different, non-interchangeable set of parameters (duration, resolution, bitrate, start/end frame).

07:26 – 07:46

10 · The two ways to recreate the workflow

States the fork: use a model aggregator, or go straight to the source and integrate each model provider's API directly.

07:46 – 09:17

11 · Using fal.ai as a model aggregator

Demos fal.ai's dashboard: fund an account, grab an API key, and tell an AI coding assistant to build a UI that syncs to the provider's current model list and parameters.

09:17 – 10:05

12 · Connecting directly to model-provider APIs

Names the direct-source models (Veo, Omni, MiniMax, Seedance, Kling) and shows pulling up Google's own Gemini API documentation for Veo 3.1.

10:05 – 11:23

13 · Teaching AI how each model should be prompted

Shows feeding a provider's documentation page, viewed as markdown, to an AI coding assistant so it learns exactly how to structure API calls and refine prompts for that model.

11:23 – 12:05

14 · Mapping the full interface with AI

Returns to Bench Studio to show that every preset button (like Unboxing) is a pre-written prompt template with its own mapped parameters underneath.

12:05 – 12:26

15 · Transparent pay-as-you-go pricing

Aggregators forecast the exact cost of a generation before you click the button, replacing the guesswork of a flat monthly subscription.

12:26 – 13:01

16 · Free Bench Studio repo, skill, and setup guide

Closes by pointing to the open-sourced repo, MCP setup instructions, and skill, plus a paid community for going deeper.

Atomic Insights

Lines worth screenshotting.

A subscription AI creative platform is functionally four parts: reference-media intake, meta-prompting, a model library, and a parameter set, none of which requires the subscription itself.
Meta-prompting means a cheap workhorse model rewrites a vague user prompt into one optimized for the specific model chosen, using that model's own prompting documentation.
There are two ways to rebuild a model-aggregator platform yourself: route through an aggregator like fal.ai or kie.ai for one API across many models, or call each model provider's API directly for lower markup and more control.
Model aggregators like fal.ai and kie.ai continuously crawl and catalog every new model's parameters so you don't have to maintain that library yourself.
Every serious AI model ships a system card or prompting guide from its provider, and feeding that documentation to an AI coding assistant teaches it how to prompt that specific model correctly.
An AI coding assistant can read a provider's API documentation page saved as markdown and write the integration code, so unfamiliarity with APIs doesn't have to block a non-technical builder.
The two hardest parts of building a model-routing interface are making the AI tailor the UI's inputs to whatever model is currently selected, and giving every generation a transparent, per-request cost estimate before you click generate.
Every input a model can take, such as a start frame, end frame, reference image, bitrate, or resolution, is model-specific, so a UI built to route between models has to change its visible fields when the model changes.
Connecting a custom tool to Claude Code and Codex through MCP lets you generate video, images, websites, PDFs, and slide decks from a plain-English request inside a coding assistant instead of a separate app.
Pay-as-you-go through direct APIs or aggregators avoids the classic subscription trap of forgetting you're paying $40-80 a month for a tool you use to make ten or fifteen images.
A creative team that needs convenience and doesn't want to maintain tooling is still better served by a turnkey subscription; a solopreneur generating a handful of assets a month is who actually benefits from owning the stack.

Takeaway

How to reverse-engineer any AI subscription platform and own it instead

OWN YOUR STACK

Every AI creative subscription is really four swappable parts (media intake, meta-prompting, a model library, and parameters), and either an aggregator API or a direct provider API plus an AI coding assistant is enough to rebuild it yourself.

01The creative studio replacing another subscription

Subscription AI creative platforms bundle prompt refinement, a model catalog, and billing behind one interface, which is convenience you're paying a markup for.
Owning the interface means owning the routing logic, the prompt intelligence, and a transparent cost breakdown instead of being locked to one vendor's credits and roadmap.

02Turning a product image into a UGC ad

A single product photo dragged into a custom interface, plus a plain-language description of the desired UGC actor and reaction, is enough input to generate a synced-dialogue ad clip.
The workflow reuses existing consumer applications and public APIs rather than custom-built generation technology; the value is in the orchestration, not new models.

03Model-specific prompt refinement

A 'refine prompt' step runs the user's rough request through AI a second time to sharpen it specifically for the model that's about to generate it.

04The generated Veo video and exact cost ledger

Every generation logs its exact billed cost next to the result, so cost tracking doesn't require checking a separate billing dashboard.
The underlying model doing the generation is swappable; the interface itself is model-agnostic by design.

05Using Bench Studio through Claude Code and Codex

The same generation tools exposed in the web UI are callable from inside Claude Code or Codex through MCP, so a plain-English coding-assistant request can trigger a full UGC video generation without opening a separate app.
AI can handle roughly 99% of building this kind of tool end to end; the remaining human work is testing that it actually behaves as intended.

06Websites, PDFs, reports, and other creative outputs

The same routing pattern extends past images and video to 3D websites, PDF reports, and slide decks, generated inline from a terminal, desktop app, or web UI.

07What Higgsfield is doing behind the interface

Reverse-engineering a competitor's product into its component steps, media intake, meta-prompting, model routing, and parameter exposure, is how you decide whether to build or keep renting.

08Reference inputs and prompt optimization

What a model accepts as input, such as an image, video, or start/end frame, is entirely model-specific, so the UI has to adapt its own input fields per model rather than offering one fixed form.

09Model libraries, parameters, and routing

Maintaining a current library of every video and image model and its parameters used to be a standalone engineering burden; aggregator APIs now do that maintenance for you.
Parameters like duration, resolution, and bitrate are gated per model, and switching models can make a parameter like bitrate disappear entirely because it no longer applies.

10The two ways to recreate the workflow

Recreating an aggregator platform comes down to two paths: route through an existing aggregator, or integrate directly with each model provider.

11Using fal.ai as a model aggregator

Aggregators like fal.ai continuously crawl provider sites to keep their model list, parameters, and pricing current, and hand you a single account and API key for all of it.
The workflow to stand one up yourself: fund the aggregator account, grab the API key, and tell an AI coding assistant to build a UI that syncs to that provider's model list.

12Connecting directly to model-provider APIs

Going straight to a provider is more setup work than using an aggregator but avoids the aggregator's markup entirely.
A provider's own API documentation page, saved as markdown and handed to an AI coding assistant, is enough for it to write the integration code, even for a non-technical builder.

13Teaching AI how each model should be prompted

Every serious AI model publishes prompting guidance from its own provider; feeding that to an AI coding assistant teaches it how to prompt that specific model well.
Once the assistant has read a model's prompting guide, a short vague prompt gets automatically expanded and optimized for that specific provider.

14Mapping the full interface with AI

Every button and preset in a generation UI is really a pre-written, glorified prompt template with its own parameter set mapped underneath it.

15Transparent pay-as-you-go pricing

Aggregator platforms will forecast the exact cost of a generation before you click the button, so pay-as-you-go pricing removes the guesswork subscriptions hide.

16Free Bench Studio repo, skill, and setup guide

The creator open-sourced the interface, the Claude Code and Codex skill, and a build guide, meaning the entire workflow described is reproducible without paying anyone for it.

Glossary

Terms worth knowing.

MCP (Model Context Protocol): A standard that lets an AI coding assistant like Claude Code or Codex call an external tool or API directly, as if it were a built-in function.
Model aggregator: A service such as fal.ai or kie.ai that maintains one API and dashboard giving access to many different AI image and video models from different providers.
Meta-prompting: Using AI to rewrite a user's vague prompt into a more detailed, model-specific prompt before sending it to the actual generation model.
Model library: A catalog of the specific AI models a platform can route a generation request to, each with its own supported inputs and parameters.
UGC ad: A user-generated-content-style advertisement shot to look like an ordinary person's authentic phone video rather than professional marketing.

Resources

Things they pointed at.

07:40toolfal.ai ↗

08:04toolkie.ai ↗

09:50linkGoogle Gemini API - Veo 3.1 video generation docs ↗

12:26toolBench Studio public repo ↗

12:26productBench Studio ownership kit (free) ↗

03:13productEarly AI-dopters community ↗

Quotables

Lines you could clip.

00:00

“Before you pay a company like Higgs Field yet another dollar, you'll wanna see this.”

cold-open pattern interrupt naming the competitor directly→ TikTok hook↗ Tweet quote

00:31

“I own the interface, the routing, the prompt intelligence, and most importantly, a completely transparent cost breakdown.”

thesis statement delivered as a punchy ownership list→ IG reel cold open↗ Tweet quote

12:09

“If you're the type of person that would accidentally forget about a subscription and spend $40, $80 a month just to make ten to fifteen images, you could save tons of money over time just by doing a pay as you go model.”

concrete, relatable subscription-guilt number with a clear alternative→ newsletter pull-quote↗ 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.

Before you pay a company like Higgs Field yet another dollar, you'll wanna see this. These videos, product shots, websites, slide decks, and reports all came through one creative studio I built for a fraction of another monthly creative subscription.

Through one interface, I can access up to 37 cutting edge models and use both photo and video models as I please. And on top of that, I can access all of the same services through things like connectors and MCPs in things like Cloud Code and Codecs.

I own the interface, the routing, the prompt intelligence, and most importantly, a completely transparent cost breakdown.

And this means that I'm no longer locked by another provider's credit system, features, or roadmap. And if you watch this video till the end, I'll give you the skills and the clone of my creative studio so you can set it up and own your own product.

So if you're ready to run your own content factory out of the box, then let's jump in. So to start things off, I'm gonna show you how simply we can create something like a user generated content clip just using a sample image and dropping it into this AI generated UI using existing applications and APIs that you can take off the shelf.

So if we go into a brand new tab and we write cool products and we click on images and we scroll down for something that is demoable. Okay.

So let's say we take this pizza oven here and we save this pizza oven on our computer. Let's go to the next step, which is actually dragging and dropping it. And just like any other platform, we can click on this plus button right here.

We can drag and drop the little file and we can set it to UGC add and we'll say, want someone who's super enthusiastic about using this pizza oven and says how much time it saves them and how great the crust of the pizza comes out once they put it in.

So we will put that there. We'll click on, let's say, man in his thirties. We'll just click on refine prompt.

This will use AI to make this prompt better for this specific model. Then we'll click on generate, and then voila, you'll see right here, it is using VEO 3.1 fast.

But like I said in the intro, this could be any one of these 37 models. And the best part is I had AI use an existing API that I'll show you in a bit that can do all this dirty work for you, and all you have to do is tell AI to create an interface that always syncs to this API. And just like that, you get a clip like this.

Look at that crap. This saves me so much time. Now was that perfect?

Absolutely not. We could refine the prompt. We could add some more sample images.

We could tell to really focus in on the person, then the pizza oven. All that stuff is doable. But it's the idea that we have full transparency on exactly how much this cost.

We have our ledger on the top right hand side that can always keep track of exactly what we're spending and why. And more importantly, you have the flexibility to keep changing this as you need to focus on your use case of choice. And by the way, I'm going through everything at a very theoretical level just to keep things as approachable and as accessible to as many people as possible.

But if you wanna be able to learn how to build a platform like this from the ground up, how to map out infrastructure, how to use the right tools for the right use case at the right time, and go infinitely deeper to make yourself more confident with things like AI, then you wanna check out the first thing down below for my early AI adopters community.

Typically, every single video that you see on YouTube has a pairing with a deeper dive in the community with exclusive resources, takes, and access to coaches can help you go from zero to one. So if you wanna learn how to become more independent and build your own tooling, then check out the first thing down below, and maybe I'll see you inside.

And whether you use Claude, Codex, Gemini, or whatever other language model, you can create your own local MCP that hooks up to this UI. And, you can have AI do 99% of this.

All you have to do is test it. And then you could say something like, use Bench Studio to create a beautiful UGC video of a woman demoing a very interesting product I'd never come across in my day to day shopping. That would be a scroll stopper on Instagram.

And like you would with other products, it will go and see what inputs it needs for this particular model. My default one is video, but you can switch it to anything like a C Dance, like a Mini Max, etcetera. And then after a bit, you will get this inline generation.

So you can actually create your work within a terminal, within a desktop app, or in the UI of choice. On top of that, you can create three d websites on the go, documents like PDFs, and really high caliber documents like this one where we can open this up. You can see it's well formatted, it's well structured, and it doesn't look like the typical AI slop.

And like I said, once I walk you through the mental model of how I built this, I'm gonna be open sourcing all of this including the MCP and the skill that you can use out of the box so you can take this, run with it, and make it your own. Now if you go on a platform like Higgs Field, you will have all kinds of different tools.

And for the average Joe, you don't need all this stuff. So if you're a creative team, a subscription like this might make sense. It might be infinitely more convenient especially on a team basis.

But if you're just a solopreneur, a small business owner, and you wanna be able to generate a handful of creative images, but do so in an easy to use and an easy to observe environment, you don't need all of this.

So let me just show you behind the scenes what is happening when we do something like create video. Since this is a wrapper, every single element here we can map to something that we can reverse engineer. So first and foremost, when it comes to uploading media, like you saw me do on my end, this just depends on what the model takes as input.

So some image models, video models can take an actual image or even a video as input to use as the inspiration for what will be generated. So that's one part there. Number two, if we take the prompt, how this works is you as the user can enter a very vague prompt.

Behind the scenes, they use some AI to enhance said prompt either in general or for that specific model that you chose. So in this case, this is basically meta prompting, getting AI to write the prompt for AI comma with additional information about what model you're using and what are the best ways to prompt said model.

So in this case, it's using c dance two point o and it's creating a prompt for that. Now next up, we have the model. And back two, three years ago, it would be a huge headache to maintain an entire library of every single model and constantly update it for the latest and greatest.

But now, like I'll show you in a bit, we have aggregators that are using APIs that you can tap into, not pay as much of a markup as you would for something like a Hicksfield or competitor of them, and then be able to take advantage of the exact same infrastructure. So in this case, we just have a model library, in which case we have multiple providers that you can do this with.

Then we have the core parameters, and the parameters are really here. So how long the video is that is really subject to that model. Not every model can do eight seconds, ten seconds, twelve seconds, fully contingent on that specific model as well as the resolution and the bit rate.

So those are specific to the model as well. Not all of them will have the exact same configurations. So if I go to something like a cling, right, all of a sudden, bit rate disappeared because that is no longer a variable or a parameter that is relevant to this kind of model.

And for this model, it seems like it can take a start frame and an end frame as an input. So the different inputs will change like we saw here. And the hardest part is just making sure that AI can really tailor your own UI, your own clone to have all the inputs needed for that specific model at the right time.

If you wanna be able to recreate this experience to work for you, you have two core paths. Path number one is you use what are called model aggregators, and that's basically what something like a Higgs field is doing.

They're doing all the legwork to pull the latest models, map them, map all the parameters that they take, the things that are optional, the things that are mandatory, and they make it easy for you to use. So very turnkey.

There are other services and some services are called fal.ai and let's say key.ai, both of which not affiliated with whatsoever.

They do the dirty work as well where you can hook up to those APIs and they do all the legwork to constantly crawl all of these existing websites to pull the latest names and the latest models and the latest parameters. So if you go to something like file.ai, you'll see on trending, you have all of the core brand new models that have come for images.

You can check for video as well. You have all the labs here, Kling, MiniMax, LTX, Cdance two point five, and they're always on top of the latest and greatest.

So all you have to do is create an account, port into it, add, let's say, $10.20 bucks, and then you take your API key, you feed that to AI, you tell it to create a UI that syncs with the list of models that are coming from this provider. And then when you send a request, you can just ask AI to say, I wanna have a box where I ask for a simple thing, and that thing gets carried over, translated into however I'm supposed to send that specific request to that specific vendor.

Now option two is actually infinitely more simple and potentially cheaper where you go straight to the source. And what this means in plain English English is if you know you wanna use a model like Google Vayo or Google Omni or you wanna use a different model like a MiniMax or SeedDance or Cling, all of these vendors here all have individual API documentations.

And even if you're nontechnical, even if the word API is daunting on its own, you don't have to worry about it. So you can go to Google and put something like Google VEO 3.1 fast documentation. You can grab the first one that looks legit.

So this one, this will go through all of the different parameters and everything that would be needed for this particular API. So you can click on video generation and then take this entire page, view it as markdown, take this whole thing, feed it to Codex, Cloud Code, Gemini even, have it go through all the documentation, and it will figure out exactly how to structure API calls.

All you have to do is give it the idea and you can even tell it. I want you to build a system where you use an additional API key or an additional API credit to go and take my vague prompt and make it something adjusted and optimized for this existing model. Now how would it know to refine it?

Well, with every model, just like you would with any model that comes out of the box, if you go to video understanding, dialogues, text generation, there are always tips and tricks from the model provider themselves as to how you should manipulate and prompt their specific model.

So if we go to something that's a little more familiar, let's say we go to an Opus five model card. Now this is not a video model.

Just to show you the concept, every model in the world has some form of a system card. And in these system cards, they break down how best to prompt said model.

If you have AI, go do all the legwork and find all the documentation per provider assuming you want to use multiple of them. It will figure out exactly how to best prompt it. Once it knows that, then you can send a very vague prompt like this, a stunning drone view or something even more simple than that, and it will take it and refine it and optimize it for this specific provider.

So with that overall understanding, you can really start to map out every single part of this user interface from these buttons that pretty much have a glorified prompt on each one of them. And then when you click on something like unboxing, these are the different parameters that are associated with something like an unboxing video.

And when we click on something like refined prompt, all that's happening behind the scenes is we're just having a cheap workhorse model, let's say a Gemini three flash that takes that prompt, looks at their reference documentation for this specific, in this case, image model, and takes your vague prompt and just adjusts it for that specific model.

And the best part is model aggregators like fowl.ai will also give you an estimate for the generation. So before you even click on a button, you can always forecast exactly how much you're gonna spend. So if you're the type of person that would accidentally forget about a subscription and spend $40.50, $80 a month just to make 10 to 15 images, you could save tons of money over time just by doing a pay as you go model.

And that's pretty much it. So like I said, I'm gonna be open sourcing this entire repo along with some skills and a full guide on exactly how I built this step by step that I'm gonna make available to you completely for free down in the second link. So hopefully, you can take all these resources and the transcript of this video and finally unshackle yourself from any of the providers if you feel like you're not getting the best value for your dollars put in.

And as always, if you wanna go infinitely deeper on things like codex, Claude code, building your own tooling, and becoming an AI beast, then check out the first link down below, and maybe I'll see you inside my community. For the rest of you, if you found this helpful, if you learned something, I would be infinitely appreciative of a like on the video, a comment if you so choose, and I'll see you in the next one.

The Hook

The bait, then the rug-pull.

Before you pay for another all-in-one AI creative subscription, the creator wants you to see what's actually running underneath one: a routable model library, a prompt-refinement step, and a billing wrapper, all of which he rebuilt himself in a tool he calls Bench Studio.

Frameworks

Named ideas worth stealing.

04:54list

The Four-Part Aggregator Anatomy

Accept reference media
Meta-prompt for the chosen model
Route to a model library
Expose only that model's parameters

The reverse-engineered structure of what any model-aggregator platform, Higgsfield included, is doing behind its interface.

Steal forreverse-engineering any SaaS wrapper around a set of AI APIs before deciding whether to build or keep renting

07:26model

Two Paths to Recreate a Model-Aggregator Workflow

Path 1: model aggregators (fal.ai, kie.ai) - one API, many models, small markup
Path 2: direct provider APIs (Veo, Omni, MiniMax, Seedance, Kling) - lowest cost, most setup work

The two build-your-own-stack options and the tradeoff between convenience and cost/control.

Steal fordeciding whether to build a custom AI tool through an aggregator or straight to the model providers

CTA Breakdown

How they asked for the click.

VERBAL ASK

12:26link

“I'm gonna be open sourcing this entire repo along with some skills and a full guide on exactly how I built this step by step that I'm gonna make available to you completely for free down in the second link.”

Two-tier ask: a mid-video pitch (3:13) for the paid Early AI-dopters community, then an outro CTA (12:26) for the free open-sourced Bench Studio repo, skill, and setup guide, both linked in the description.

MENTIONED ON CAMERA

07:40toolfal.ai ↗

08:04toolkie.ai ↗

09:50linkGoogle Gemini API - Veo 3.1 video generation docs ↗

12:26toolBench Studio public repo ↗

12:26productBench Studio ownership kit (free) ↗

03:13productEarly AI-dopters community ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.

OTHER LINKSAlso linked in the description.

Storyboard