The $3,000/Day Solo AI Business With Astra + Upwork
Greg Isenberg builds a one-person creative agency where Astra plans and repairs the work, 50+ Higgsfield models produce it, and he just approves the output and gets paid.
Capable creative models, an orchestrator model in Astra, and one unified production API (Higgsfield) matured at the same time, making it possible for a single person to run a semi-autonomous creative service firm at 60%+ net margins.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
A solo operator or small team with real production or AI skills who wants a productized service business, not another SaaS idea.
Someone already freelancing on Upwork, Fiverr, or Contra who wants to raise margin instead of chasing more billable hours.
A creative agency owner curious about what an orchestrator model like Astra can take off their plate.
Anyone weighing which AI model to use for a given image or video job and wants a concrete decision framework.
SKIP IF…
You want a hands-off, fully autonomous business; the video is explicit that a human still approves every delivery.
You're looking for a SaaS or app idea rather than a productized service you personally run and approve.
TL;DR
The full version, fast.
Greg Isenberg argues that three things matured at once, capable creative models, OpenAI's Astra as an orchestrator, and the Higgsfield API that puts 50+ models behind one key, making solo AI-native service firms possible now. A traditional creative agency nets 10-15% because people leave and take their lessons with them; an AI-native firm can reach 60%+ net margins because Astra plans and repairs the work while a human just approves it and owns the client relationship. He demos Studio Operator, the dashboard he built with six sequential prompts, walks through which model fits which job, and closes with three startup ideas and four filters for picking one: existing demand, an obvious unit, repeat work, and a checkable definition of correct.
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.
Greg states his thesis: creative models plus Astra are about to produce a wave of one-person AI-native service firms, and he'll show the prompts and a working build.
01:46 – 03:35
02 · 3 Shifts, All at Once
Three things matured together: creative models got genuinely good, Astra became capable enough to orchestrate a business, and Higgsfield unified dozens of models behind one API.
03:35 – 04:22
03 · The AI-Native Service Firm
Defines the model: the customer asks, Astra runs the job, creative models do the specialist work, and the human owns the contract, taste, and rights.
04:22 – 05:16
04 · Demand on Upwork and Fiverr
Marketplaces already show validated demand for product videos, localized campaigns, and listing images; the advice is to pick one narrow deliverable and teach the system to reproduce it.
05:16 – 11:14
05 · Human Agency vs. AI-Native Firm
Compares a traditional agency's 10-15% net margin against an AI-native firm's projected 60%+, and argues the AI-native version keeps its lessons in a compounding loop instead of losing them to employee turnover.
11:14 – 20:08
06 · Studio Operator Tour
Live tour of the dashboard: active jobs, pipeline value, a real $10 client brief priced down to the cent, autonomy limits, client memory, and the final-approval step before delivery.
20:08 – 26:42
07 · The Six Build Prompts
Walks through the six prompts used to build Studio Operator in order: firm shell, brief intake, production layer, model router, QA and repair, and controls and approval gates.
26:42 – 32:42
08 · Model Primer by Job Type
A model-by-model primer: which of the 50+ Higgsfield models to use for people, products, text/layout, repairs, long scenes, hero shots, speech, motion copying, and finishing.
32:42 – 33:52
09 · Idea 1: Always-On Ad Studio
A weekly creative-pack retainer for one e-commerce SKU, sourced from the Meta Ad Library and sold as a paid sprint that grows into a monthly package.
33:52 – 41:00
10 · Idea 2: Franchise Localization
Turning one approved national campaign into dozens of checkable local versions for franchise brands with 20-200 locations.
41:00 – 43:15
11 · Idea 3: Industrial Catalog Sales Assets
Turning a manufacturer's dense spec sheets and outdated product photos into demo videos, training clips, and sales graphics for one SKU at a time.
43:15 – 45:28
12 · Four Filters for Picking an Idea
Four filters for choosing which AI-native firm to build: existing budget, an obvious deliverable unit, repeated work, and a concrete definition of correct.
45:28 – 46:11
13 · The Self Improvement Loop
Every job leaves a lesson about inputs, models, margin, and requirements, and in an AI-native firm those lessons stay in the system instead of leaving with an employee.
46:11 – 46:59
14 · Closing Thoughts
Greg reiterates that creative AI has reached the point where you can build a real company around it, and hopes the video sparks an idea.
46:59 – 48:05
15 · How to Start
Concrete starting steps: one narrow package, five customers, write down every missing input and mistake, then turn those lessons into intake, routing, QA, and approval rules.
Atomic Insights
Lines worth screenshotting.
Three technologies matured in the same window: creative models good enough to use, Astra good enough to orchestrate, and one API (Higgsfield) that puts 50+ models behind a single key.
A typical human creative agency nets 10-15% because hiring, turnover, and ongoing marketing eat the margin.
An AI-native service firm can reach 60%+ net margins because Astra plus a small human layer replaces most of the payroll a human agency carries.
Marketplaces like Upwork and Fiverr are a sales channel a traditional agency rarely uses, and they hand an AI-native firm validated demand for free.
On one sample $10 job, Higgsfield production cost about a penny, leaving a $9.99 expected profit before the marketplace's channel fee.
Upwork and Fiverr typically take a 5-15% channel fee, and it has to be built into the price before a job is ever accepted.
Studio Operator was built with six prompts given one at a time: firm shell, brief intake, production layer, model router, QA and repair, then controls and approval gates.
Astra is deliberately used for judgment calls only; normal code handles the price and margin math so the numbers stay reliable.
Before picking a creative model, ask four questions: image, video, or speech; new asset or repair; cheap exploration or final version; and what absolutely cannot change.
Soul is for people and editorial images, Seedream/Flux for products, Ideogram/Recraft for text and layout, Qwen for small repairs, Seedance for longer scenes, Kling for hero shots, Wan/Minimax for speech, and Topaz for finishing.
A weekly ad-creative retainer for one SKU can start as a $750-$1,500 paid sprint and convert into recurring monthly revenue once a hook proves itself.
Franchise brands with 20-200 locations feel the pain of manual localization every month but are still small enough for a solo operator to reach the marketing team.
Industrial manufacturers selling five-figure equipment with dense spec-sheet PDFs and product photos last touched in 2009 are an underserved market for AI-generated sales assets.
The four filters for a good AI-native service idea: does someone already pay for it, is the deliverable unit obvious, does the work repeat, and is there a concrete definition of correct.
In a human agency, institutional knowledge leaves with the people who learned it; in an AI-native firm, every job's lessons stay in the system and compound.
A firm with a creative surface and a boring, checkable operational core is the strongest kind of AI-native business to build.
Takeaway
Three technologies matured together, so build the firm one layer at a time.
WHAT TO LEARN
An AI-native service firm works because Astra plans and repairs the work while a human keeps the contract, taste, and rights, and the win only shows up when margin, model choice, and idea selection are all treated as separate, checkable decisions.
023 Shifts, All at Once
Creative AI crossed a real quality bar; models like Kling, Flux, and Seedream now produce work far beyond the 'AI Will Smith eating spaghetti' era that used to define what AI video looked like.
Astra can read a messy client request, plan the work, choose a specialist model, track the budget, inspect what came back, and repair a failure, which is what makes it usable as an orchestrator instead of just a chat window.
The Higgsfield API puts more than 50 creative models behind one key, so a solo operator no longer needs a dozen separate accounts and API integrations to run a production pipeline.
03The AI-Native Service Firm
An AI-native service firm still needs a human for the contract, the taste, and the rights, even though AI does the specialist creative work.
The pitch is a high-margin firm one person can run without hiring a 50-to-100-person team.
04Demand on Upwork and Fiverr
Upwork, Fiverr, Contra, and agency websites already show what buyers pay for repeatedly, product videos, localized campaigns, Amazon listing images, so there's no need to guess demand.
Upwork jobs alone run from about $10 to $50,000, and the advice is to pick one narrow deliverable, package it, and teach the system to reproduce it.
05Human Agency vs. AI-Native Firm
A traditional creative agency typically nets only 10-15% because finding and paying people competitively is hard, people leave, and marketing for new clients is a constant cost.
An AI-native firm keeps word-of-mouth and organic social but adds marketplaces as a third channel it can profitably serve, taking on $100-$25,000 jobs it couldn't afford to staff for.
Net margins in the 60%+ range are plausible when gross margins on the work itself run 70-95%, because a small human layer replaces most of the payroll a human agency carries.
The AI-native model creates a true compounding loop: lessons stay in the system instead of walking out the door with an employee who leaves.
06Studio Operator Tour
Studio Operator's dashboard tracks active jobs, pipeline value, estimated gross profit, and how many projects need a human decision, functioning as a cockpit rather than a single-purpose tool.
On a sample $10 aperitif brief, Astra estimated the Higgsfield production cost at about one cent for the image and roughly $0.70 for a five-second video, so the expected margin is visible before a job is ever started.
Every marketplace job needs its channel fee (Upwork/Fiverr typically charge 5-15%) built into the price before it's accepted, or the real margin comes in lower than expected.
Autonomy settings cap spend per job, cost per repair, and maximum attempts per step, and can exclude entire model families, which is how a solo operator keeps a semi-autonomous system from running away with the budget.
07The Six Build Prompts
The firm was built one layer at a time, in order: the operating system shell, brief intake, the production layer, the model router, QA and repair, then controls and approval gates, and each prompt only added its layer.
Astra is deliberately split from plain code: Astra handles judgment calls like accept/review/reject and quality checks, while normal code handles price and margin math, which keeps the numbers reliable.
A single service template swap, say from a launch video to a real estate listing pack, produces what feels like a different firm without changing the interface, because only the deliverable, recipe, and rules change.
An audit log records every decision, generation, cost, and repair, so a bad delivery can always be traced back to where it went wrong.
08Model Primer by Job Type
Before picking a model, ask four questions: is this a still image, video, or speech; is it a new asset or a repair; is it cheap exploration or the final version; and what absolutely cannot change (the product, the person, the words, the movement)?
Soul is the pick for people and editorial-quality images but is weak on packaging, logos, and small legal copy; Seedream and Flux hold a product consistent across scenes but need a label and color check against the original.
Ideogram and Recraft are for anything where real words and layout matter, like posters or campaign graphics; Qwen is a repair tool for fixing one detail, not for developing a whole creative direction.
Seedance suits longer cinematic sequences with native audio, Kling is for the hero shot once the direction is already planned, and Wan/Minimax handle speech but need a QA pass on lip sync and logos.
09Idea 1: Always-On Ad Studio
The pitch is a weekly creative pack for one SKU: the brand supplies the product page, photography, and brand rules, and every Friday it gets new hooks, product scenes, and video ready for a media buyer to test.
The prospecting method is the Meta Ad Library: find brands clearly spending on ads but running the same hook, background, and one or two formats over and over, then mine their five-star and one-star reviews for new angles.
Sell the first engagement as a paid creative sprint for one SKU (roughly $750-$1,500), then move a brand that gets results into a recurring monthly package.
10Idea 2: Franchise Localization
Franchises approve one national campaign, then need dozens of local versions with different addresses, prices, offers, and sometimes languages, which is repetitive, checkable work an AI-native firm is well suited for.
The target is brands with 20-200 locations: big enough to feel the pain every month but small enough that the marketing team is still reachable, found via franchise association directories and LinkedIn.
QA has to check every address, price, offer, phone number, and disclaimer against the client's spreadsheet before anything ships, since that's the part a franchise client will actually catch.
Start with a paid five-location pilot, then expand by location, campaign, or a monthly localization desk, and sell through agencies and multi-unit operators who can hand over multiple franchise systems at once.
11Idea 3: Industrial Catalog Sales Assets
Thousands of manufacturers sell expensive products with dense spec-sheet PDFs and product photos that haven't been updated since 2009; the service turns one SKU into a demo video, application video, trade show loop, training clip, and sales graphics.
Astra reads the spec sheets, CAD files, manuals, and approved claims to build a visual plan, while the rulebook checks every dimension, feature, and safety claim against the source material before delivery.
The best leads are five-figure products with a detailed spec sheet, a page with only three dim photos, and no video; sources include ThomasNet and trade show exhibitor lists.
12Four Filters for Picking an Idea
Filter one: does someone already pay for this work? An existing budget is easier to capture than a brand-new category you have to explain.
Filter two: is the deliverable unit obvious, one SKU pack, one localized campaign, one product video, so the customer can buy it easily and the system can be taught to reproduce it.
Filter three: does the work repeat, since repetition is what earns a subscription instead of one-time revenue and lets the workflow keep improving.
Filter four: is there a concrete definition of correct, product name, SKU number, claims, formats, required scenes, since creative taste is subjective but those facts aren't.
13The Self Improvement Loop
Every job leaves a lesson: which missing input predicts revisions, which model keeps the important detail, which route stays inside margin, and which good-looking output still fails the client's actual requirements.
In a human agency those lessons leave with the people who learned them; in an AI-native firm the lessons stay in the system and compound into better intake, routing, QA, and approval rules over time.
15How to Start
Start with one narrow package and five customers rather than trying to launch the whole firm at once.
Stay close to the actual work: write down every missing input, mistake, revision, and place the margin disappeared.
Turn those observations into rules, intake questions, routing logic, QA checks, and approval gates, so you're getting paid while the software is still being built.
Glossary
Terms worth knowing.
Astra
OpenAI's model used here as an orchestrator: it reads a messy client brief, plans the work, chooses specialist models, tracks budget, inspects results, and repairs failures.
Higgsfield
An API/platform that puts 50+ creative models (Flux, Kling, Seedream, and others) behind one key, with shared pricing, queuing, and job status.
AI-native service firm
A creative agency where specialist AI models do the production work and a human handles the contract, taste, and rights, instead of hiring staff to do the work.
Semi-autonomous
The video's term for a firm where AI executes the work end to end but a human still approves the output before it's delivered to the client.
Studio Operator
The dashboard MVP Greg Isenberg built to run job intake, production, QA, and delivery for an AI-native creative firm.
Model router
The layer of the system that decides which specialist creative model should handle a given production step, and can explain why.
Channel fee
The 5-15% cut marketplaces like Upwork or Fiverr take from a job's price, which has to be priced into the job before it's accepted.
Kling Motion Control
A mode that applies a reference motion or performance to a new subject or character, useful for repeatable formats like fitness or dance content.
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.
17px
metaphor
The number of one -person businesses built with GPT -Astra and creative models are about to explode. That little combo of Astra and creative models is barely talked about, so today I'm going to show you a glimpse of that future and give away everything you need to build one of these businesses before everyone else does. These businesses are called AI -native service firms, and companies are already paying $500 to $5 ,000 or more on sites like Upwork or Fiverr, I'm going to show you how you can find those jobs and fulfill them using Astra from OpenAI and creative models like Flux and Kling.
It's basically a semi -autonomous AI service firm. I'm going to walk you through how you can build it, the exact prompts you need. I'm going to give them away and I'm going to show you something I built myself.
By the end, you're going to understand what an AI native service firm is and why firms like Sequoia think there's a trillion up for grabs and how to start one yourself. At the end, I'm going to give you three startup ideas that you can copy and just go grab. This experiment is sponsored by Hicksfield.
They funded the episode and supplied the API credits used to make it possible. So thank you. I wanted to see what you can build when one API connects you to a range of creative models.
It's really cool. And that's the thing I'm going to show you. Let's get into it.
So why do I think that there's going to be this explosion of one -person businesses powered by models like Astra and creative models has to do with three things have become good enough at the exact same time. And that's why the explosion is going to happen. The first is creative models.
It wasn't that long ago that we were all sharing that video of Will Smith eating spaghetti that AI generated one. Everyone was laughing at it because AI couldn't do creative models, right? But we've gotten to the point now, you know, as of recording this September 2026, that, you know, these models are pretty insane.
I mean, you've got models like Kling, Flux, Seadream, and actually you've got dozens of these models that are just so good. So we're just no longer in that spaghetti phase anymore. The second is Astra.
I mean, when Astra came out, it just, I don't know if you've used it or not, but it's gotten to this point where you can build a operating system and actually it'll help run it. It could read a messy client request, figure out what's missing. It could plan the work, choose the right specialist model, and keep track of the budget, and inspect what came back, and repair a failure.
It can do all these things, and it just really feels like if you're a one -person business, a model at the caliber of Astra is just so good. Of course, you need to supervise Astra and models like it, but it does really feel like this next... layer, next level of AI models.
And that's really cool. The third is now you have all these creative models and they live in one production layer. You know, what do I mean by that?
You know, in the past, you know, you had to open a dozen accounts, learn a dozen APIs, move files between tabs, and try to remember which model made which asset. But now, you know, last week Higgs field came out with their API and you can just have all the models, all the pricing, all the requests, all the status, all the outputs behind one API.
So when you combine all those pieces, you can actually build something that looks less like a AI tool and more like this small service firm that's actually fulfilled by AI. And that's kind of the dream, right? I know a lot of us want to build, you know, small.
quote unquote, businesses, high margin businesses, businesses that we don't need to hire 50, 100 people. And you can kind of do that when you've got this combination of these three things. So, you know, when you can combine these three things, you can have something like a customer asked for a video, Astra is going to run the job.
the creative models perform the specialized work, and one person can handle the contract and the taste and the rights and all that sort of stuff. The other thing that I don't see a lot of people talking about is you can go to places like Upwork, Fiverr, Contra, or even just agency websites and just look at what people buy repeatedly, like product videos or localized campaigns or Amazon listing images.
And if you pick one narrow deliverable, you package it up and teach the system, this operating system, how to produce it again and again, you don't really need to guess what people want to pay for anymore. You just have to give it the brief, give it the budget, give it the customer language, understand how the system works and build a system that works.
And you're pumping out a machine that's creating good quality content. And that's why this opportunity is so large. So let me show you how you can actually build one, what it looks like end to end.
Let's go check it out. Let's talk about the difference between a human agency versus a AI native service firm. What is the difference?
What does it unlock? Where's the opportunities? And how do I think about it?
So today we're going to be using the example of a creative agency. So creative agency. Why creative agency is I've built them in the past, so I understand them pretty well.
If you are building a creative agency, just a traditional creative agency with human beings, the way that you're going to get customers is going to be word of mouth. It's also probably going to be just like organic social, building a brand on social media. And that's basically how you're going to get leads.
And those leads, you are going to get people to do the work. And the people that you're going to get to do the work is people. People do the work.
Human beings do the work. And when they do the work, you, my friend, are going to get paid. How much are you going to get paid?
It's not crazy that the net margins at the end of the year are in the 10 % to 15 % net margins. Why is it so low? It is so low because finding people, paying them competitively is tough.
People leave. You have to replace them. And it's also hard to do marketing over here to get customers into the door on an ongoing basis.
Now, what does an AI -native firm look like? Well, the AI -native firm unlocks a sales channel that was otherwise just out of reach. So an AI native creative firm can still do word of mouth, still can do organic social, but it has one more thing it could do.
It could go to marketplaces like Upwork, Fiverr, things like that online and ingest projects and needs that people have and go and fulfill them. These were traditionally lower cost work. things that are $100, $500, $1 ,000, $5 ,000, $10 ,000, $25 ,000.
But you can do them because the margins are going to be a lot higher when you're creating an AI -native creation shop. Because who is doing the work here? Well, you can have something like Astra creating the operating system.
of how you actually get things done and orchestrating. You can have something like Higgs field, which manages one API for all the creative models to actually create the work. And together, plus a small amount of people, could even just be one person, you do the work.
And the person approves the work. It's not a fully autonomous company. It's a semi -autonomous company.
A lot of people talking about how autonomous companies are here. They're not really here, but they're close. So it's semi -autonomous.
I don't think it's crazy that this bumps up to 60 plus percent net margins. Because the work that you can do here on a gross margin level could be 70, 80, 90, 95 percent. And that's what we see today with some of the things that I built.
And, you know, I think that this is just an example, just at a high level, you know, and there's obviously a lot of differences between human agency and AI native service firm. But one is sales channels that have been out of reach are now in reach. Price points, let's talk about what's different in human versus AI native.
So we have sales channels. now unlocked, new ones. You have a true loop of a business.
What do I mean by that? I mean that if you're a creative agency and you learn over time and learn over time and you gain these lessons, the problem is you have people doing the work. So those people leave to other places.
So you lose those lessons. If you create a loop over here, you're constantly learning and you're getting better and you're able to create. creative that's better and better and better and better at higher margins, higher output, less mistakes.
So because this is truly digital first, AI first, there's that loop there. You do have a true loop of a business and easier to manage. Now, I don't think creative agencies are going away.
You'll still have... people, human -created agencies doing stuff, I think that over time they're going to adopt a lot of what's happening here in AI native land. But I think this is just an example of how to think about, okay, what does a new way of doing business look like?
And what are the outputs? that come from it. What are the results that come from it?
Why should I be doing it? More margins, easier to manage, potentially higher exit multiples because it's more software oriented.
Human agency versus AI native service firm. So this is Studio Operator and it's basically your cockpit for running AI creative service work. You can see all the active jobs you have.
Okay, 25. Pipeline value is $3 ,400. Estimated growth profit is $3 ,200.
And there's 17 projects that need my attention. And the end state of this is you're going to have all these inputs for jobs that are coming in from places like Upwork, Fiverr, direct leads. You're going to go and...
go and search for these opportunities. You're going to put them in here and then you're going to have AI actually complete the work. That's what this is.
If you go to Upwork, there's literally hundreds of thousands of jobs or thousands of jobs of people willing to pay between $10 and $100 ,000 for different projects. And you can go and just copy and paste those into here and go and do those projects. with something like Studio Operator.
If this episode gets 5 ,000 likes and comments, we've had many videos on the channel, luckily get to that number, I'm gonna give away Studio Operator. I'm just gonna have it so that any one of you can go and duplicate it and actually run one of these businesses. You'll just have to copy and paste and you can go and build it.
So let's see the comments and likes roll in and I can't wait to give it away. So let me show you how. this works.
Here I'm in the QA phase with this. This is an interesting one. So say someone wants to launch a new blood orange aperitif.
You can see a direct lead over here. So it's not coming from Fiverr or someone like that. But someone's like, hey, I want this.
I go and create the client brief. If it was from Fiverr or Upwork, I would just upload the link here and Astra would go and extract all that data and create the client brief, create the reference assets, and create the client notes.
It goes and reviews all that brief and basically says, okay, what are the deliverables going to be? It's going to be in PNG. We want it to be 3x4 ratio.
Are there any questions for the client? This is basically productizing the operations of the actual job to be done.
You can even say like, hey, I don't want to spend more than five cents as a production budget. And as you can see in this product, minimum gross margin and gross margin and profitability is really, really important to me for this because I'm looking to create creative that's going to get me 70, 80, 90 % plus margin. You can see here the client was willing to pay $10.
uh i have astra actually give me a estimate based on what higgs field is going to actually go and create uh one cent here so my expected profit is gonna be 9 .99 you know if i can you know do that i want to do that all day and then you can see uh here's the workflow so it went and looked into my model router So what's really cool is Astra built a model router with Higgs field that basically says like, oh, hey, you know, this model is good for this.
This model is good for that. And it goes and creates the best possible possible stuff here. You know, you can even propose a visible production route over there.
And this is what it looks like. You know, to me, this looks really just breathtaking. Honestly, this is really cool, interesting, actually pretty scroll stopping.
And it did a good job. And it did a good job because of the model routing. It did a good job because of the brief.
It did a good job because of just the workflow that we've been able to do. And then there's a place for revisions as well. And if I like this, it comes to me on my main feed and it's ready and it's in QA.
I can just approve delivery and I'm done with it. Good to go. And I didn't add this to send the files to your client manually, but in the future I can just automatically have Astra send that.
How insane is that, right? Like you're grabbing the validated demand that exists on the internet. You're putting it in here.
You're able to actually create high quality creative that looks good, that feels good. You have that human layer too. You know, for some people who want to go completely autonomous.
Go for it. I think we're going to get there over the next 12, 18, 24 months. But right now, I do think having a human being go and looking at some of the stuff is important.
I'm going to show you just a couple more things here.
Just because the output is insane. Again, the aperitif. A five -second video.
You can see here the client willing to pay $10. about a 70 cent estimation of what it would cost and it's important to have that because like you know before you hit go you don't want to you know if if the expected margin is 1 .6 instead of 91 .6 you might not click it right um it also includes uh included like a channel fee uh because you know fiverr i think charges five to fifteen percent or or upwork charges five to fifteen percent these marketplaces you know, charge contingency channel fee.
So you were going to want to include that there. You can see the human checkpoints over here and the activity. This just shows me like what has happened, like all here's all the generations.
And I can just go and as a human being, you know, just look at this and be like, okay, yeah, this makes sense. This makes sense. Or this doesn't make sense.
So overall, you know, just a, just a really, That's why I said it was a glimpse into the future because something like this or more and more things like this are going to exist. And I'm just surprised that more people aren't doing this.
Other things I can show you, like there's this autonomy section, which I think is really important. I might not want to spend more than $10 per job or $100 per job. Or one of the reasons why agencies really get killed on margin is repairing things.
So you might want to say $0 .50 per repair or maximum attempts per step. You can have certain model families that you like and don't like. Hey, maybe you don't want to use any Chinese models.
So you're like, I don't want to use Quinn or something like that. or C -Dance, you decide and you can actually have all the client communication stuff here and the client account memory, right? So like, oh, we have a customer called Moro Coffee, here's their name, here's their logos, their colors, their fonts, what they like, rejected styles, like all their stuff will be included here just like a real agency.
So it's a... It's a real agency, except the stuff gets done by AI. Astra is like the orchestrator with Higgs field.
And the output is really good. It's shockingly good. It's got memory.
And the hardest part then becomes going and searching for what are the interesting...
projects on places like Fiverr and Upwork or creating your own brand and getting people to come to you. Maybe you start an X account or an Instagram account and you say, I do logos for $10. And maybe the growth strategy is you redesign other people's or famous logos.
And then you say, oh, hey, and I'll do a logo for you for $10. And you can just Venmo me. or whatever, right?
Like things like that and getting a lot more direct leads might be better, right? Because you don't have to pay the channel fee and you're building more enterprise value. So I think this is really interesting.
I honestly was shocked with how good it was. And there you have it. This is a tour of Studio Operator.
Let me give away the six prompts to actually go and build this using Astra and the creative models. So I built this product with six prompts and each prompt added one layer of the firm. I'll explain what I asked for and I'm going to show you what it created instead of just reading the instructions line by line.
I'll just give you like the high level of what I asked for. And I'm giving away all these prompts in the link in the description. You can go and grab that if you want to create something like this on your own.
So the first prompt, the build the firm prompt, it basically created the shell of the business. So I asked Astra to build an actual operating system instead of another chat window. So I wanted service templates.
I wanted a job inbox. I wanted production stages. I wanted client records.
I wanted deliverables, costs, margins, approvals, revisions, and delivery. As someone who's run agencies before, I knew I needed all those things. Every job needed a clear state and a next action.
So a semi -autonomous firm needs somewhere to store what happened, what was approved, what it spent, and what still needed a person. So that would be the foundation, and that was prompt number one. So the second prompt teaches Astra how to turn messy demand into a job.
so that it can actually run the job and we can get paid. So it reads an unstructured marketplace brief, like if it's coming from Upwork or Fiverr or one of those, and it extracts the deliverables, it extracts the formats, exact copy, the deadline, the supplied assets, the missing inputs, the risk, and the likely revision problems.
which is really important because that's going to affect margin. It drafts the smallest useful set of client questions and it recommends whether the job should be accepted, reviewed, or rejected. And Astra is going to handle the judgment here and the normal code is going to handle the price and the margin calculations.
Once we have that, that's when we can actually connect it to Higgs field to get the creative models, right? So the third prompt connected the Higgs field API as the production layer. One key opens a catalog of more than 50 images, videos, audio models, and workflows, which is really cool.
And Astra can route work across Cdance, Kling, Sol, Recraft, and the rest of them without separate provider integrations of each one. Higgs field is basically the harness around those models. It handles the queuing, the GPU scaling, the shared submit, check, download flow, and exposing public pricing and cost estimates before each generation run, which is important because we want to keep our gross margin high, 70%, 80%, 90%.
The key stays on the server and every request, result, failure, and actual charts stays attached to the client job. The fourth prompt is going to build the model router because we're going to need the system to basically go to the right creative model so we get the best output. So the fourth prompt turns the Higgs field model catalog into this production system.
So Higgs field provides the models and the workflows and the pricing and the execution layer. And I wanted Astra to decide which model fits each step and actually explains why. So I have it actually explain why in studio operator.
I can, I did add a feature where I can. override asterisk choice and immediately see how that changes the job economics. Problem five is adding quality insurance and repair.
So it's going to be all in charge of correcting things. So every output gets checked against the approved brief for the right format, duration, copy, product details, required scenes. brand rules, motion, audio, and anything else the client's going to approve.
The system decides whether the asset is ready or it needs a controlled edit or it needs to be regenerated or it needs a human decision. You might be wondering why some of the outputs on Studio Operator look so good and it's because part of it is the QA is really good. Look at this.
absolutely stunning. This is for a perfume brand, but just looks really cool. And it's got a lot of good aura.
And a part of that is because of the QA prompt. So don't forget that. The sixth prompt adds the controls that lets the autonomous firm keep moving safely.
So each job has spending limits, repair limits, maximum attempts. Approved models, rules for client messages, and a clear approval gates. So Astra can draft updates, interpret routine feedback, which is really cool.
Remember the client's preferences. There's a whole area for the memory on studio operator. Look at major scope changes and final delivery come back to the human that we can actually approve.
The audit log records every decision. Generation cost. change, repair, and everything.
You can just see all the activity that happens. And that was important for me because I want to make sure that the stuff that gets delivered is high quality. And if there's some mistake, I can trace back to it.
The orders of the prompts do matter a lot. I gave Astra the business one layer at a time. You know, what we sell, how a job moves, how production connects, how models are chosen, what correct means, and which decisions require a person.
And then I put it together. It's not like I just gave it all at once. And I think you're going to get a better result if you do that.
Once those six layers exist, changing the service template creates an actual different firm. So a launch video template can become like a whole podcast promotion pack. Maybe it becomes a real estate listing pack, a monthly restaurant content service, these sorts of things.
The interface doesn't really change, but the buyer, the deliverable, the production recipe, the rule book, and the approval points could change. And that's what you want because you're going to have different types of clients come in. Okay, so I've shown you studio operator, how to take a job from Upwork or Fiverr, take it through production, repair and delivery.
But there's so many creative models. And I don't know about you, if you're anything like me, you're just like, what model is... is each good at?
And I know that you have Astra orchestrating the thing and routing the models, but it's important that if you're going to do this, you understand what each model is good at and what it's bad at. So before choosing a model, I ask four things. Am I making a still image, a video, or something with speech?
Am I creating something new or repairing an asset that is already closed? Am I exploring new ideas cheaply or am I making the final version? And what absolutely cannot change?
The product, the person, the words, or the movement? Once you answer those questions, the model choice becomes a lot easier. So we're going to go through just a primer of all these different models.
You've heard their names. By the end of this little section, you're going to understand what you should use and when you should use it. If the job is a person and the image needs some amount of taste, I would use Soul.
It's like really good for fashion and like editorial images, identity, and it's got like these culturally current portraits and founder photo shoots. But the weak spot is like things like packaging, logos, and small legal copy. So it'd handle those in like a separate step.
If the job is like a product scene, Seedream or Flux. They're really good at taking one clean product reference and placing it into different lifestyle images or campaign worlds while keeping the subject recognizable.
But I will watch out. I would give you a few watchouts. Labels, the exact colors, sometimes small geometry.
You need to check those against the original product.
If the creative asset needs real words and layout, ideogram or recraft are going to be your best bet. So that's the lane for posters, covers, local promotions, campaign graphics, and other design assets where typography is part of the product. Obviously, I'm not going to go reach out to ideogram if I'm trying to do cinematic motion and things like that.
If the image is already good, but like one detail's wrong, quen. So it's really good at repairing a label or replacing one object, localizing a headline, or correct like one small visual mistake, but you want to preserve everything else. So I use it as like a repair shop, but not a place where I want to develop the entire creative direction.
For cheap still image exploration, Zimage and Marketing Studio Image can produce a lot of directions quickly. For cheap video explorations, Pixverse, the faster WAN models, and lower -cost Kling configurations can test hooks, openings, and camera ideas.
Their job is to help you decide what is worth making properly, so don't confuse the most interesting test with the final deliverable. If the job is a longer cinematic scenes, I would use C -Dense. It becomes useful when you need multiple references, a sequence that develops over time, native audio, or the ability to edit part of the result.
It is a stronger production model once the direction is clear. Rather than the model, I would use for 30 random experiments. If the job is the hero shot, I'd use Clang.
That's where it's been more for camera movement, physical realism, and that final touch that you know that the customer is going to notice. There are more expensive ways to discover that the original idea was weak, which is why I would arrive at the shot already planned. If the person needs to speak and the audio should be generated with the scene, WAN or Minimax.
They're useful for creative ads, spokesperson videos, localized performances. But I would watch out for lip syncing, logos. Those things need a QA pass from them.
If you already had the exact movement you want, I would use Kling Motion Control. So you give it the movement reference and apply that performance to a person or character, which makes it really useful for things like fitness, dance, demonstrations, and repeatable creative formats. Think about it.
if you figure out the format you give it you give it to cling motion control and you know it's it's pretty pretty darn good the source image and motion reference have to be clean though because bad inputs uh do be car do become part of the output then i would use topaz lip sync voice captions, and export tools to finish the work.
So these tools improve resolution, they localize and improve performance, and create the actual files the customers want to publish at the end of the day. So they can polish a good asset, but they're not going to rescue a bad concept. I will say, I wish the Higgs field API had nano banana.
Wish it had it, it doesn't have it. So the way to think about all these models, Sol is for creating people, designing people.
Cdream and Flux are for products. Ideogram and Recraft are for words and layouts. Quen is your repair shop.
Cdance is for those longer scenes. Juan and Minimax are for speaking. Motion Control is for copying or performance.
And Topaz and Finishing Tools get the work done to deliver. That's basically the primer of if you're using Higgs Field Launch API, those are the best ones to be using. Okay, so I'm going to give you three startup ideas that are all about this whole AI native service firm that you can build with Astra and you can build with these creative models.
Run them end to end. I hope people build them. And if at the very least just gets your creative juices flowing, let's go.
So the first one is an always on ad studio. So every e -commerce brand has the exact same problem. the ads get tired faster than its creative team can replace them.
So I would sell a weekly creative pack for one SKU. So the brand gives you the product page, the original photography, the brand rules, and the approved claims. And then every Friday, it receives new hooks and product scenes and still ads and short videos that are just ready for the media buyer to test.
You can call it something like Friday ads or something like that. And then the production system is pretty simple. You can have Astra reading the product pages, the reviews, social, current ads, to find the language that buyers are already using.
And you'll use that for the basis of your ads. Cheap models. explore 20 angles before you spend real money on them.
So customers are going to be happy about that. You can use like Flux or Seadream to keep the product consistent across the strongest scenes. Then you can use Quinn as the repair shop.
Things are going to need help on packaging and copy. And then faster video models test movement, right? And premium models produce the few ideas worth putting in front of customers.
So you have QA comparing the label, the colors, the claims, the formats, and the required shots with the source assets before anything leaves the system. If I was starting this tomorrow, I would choose one category and build a list of like 100 brands. I'd probably open the meta ad library and look for companies that are clearly spending money but have not a lot of creative variety.
It looks all the same. It's like the same product hook, the same background, one or two or three formats across most of their active ads. Then I would be looking at their reviews, their five -star reviews, even their one -star reviews to get some insight around how can I test a bunch of different hook angles.
For the best prospects, I would make three private watermark concepts and record a 90 -second loom. And the pitch becomes, hey, you have this really good product, but most of your ads are making the same argument. So I've got these three unused hooks from customer reviews, and I mocked up what they could look like.
You're actually showing them the work, and you're showing value. And then I would sell the first engagement as a paid creative sprint for just one SKU. Maybe you charge like $750 to $1 ,500.
Depends on how many deliverables they want. And then once something performs, you can move the brand into like a monthly package where you get that batch every week and you're getting that recurring revenue. I would say like a really interesting growth tactic for this business.
is media buying agencies. So they already have clients who need more creative. And if you can build a production layer that plugs into the account...
That's super valuable. Then you're learning in real time. You're building this library of hooks and failure patterns and winning creative structure that Astra then could reuse without making brands look the same.
That data is really, really powerful. That's a good one. It doesn't require a lot of money to start.
I like the monthly recurring revenue. I like that you're validating because it's validated. These people are spending money already.
The big if is if you could create actually good ads. And I believe what I showed you today with Studio Operator, you can. The second startup idea is one of those boring business ideas that some people might really like.
So franchises. A franchise might create one good national campaign and then they need 50 local versions with different addresses, prices, offers, languages sometimes, phone numbers, all that sort of stuff. So the way it works is today, someone at like HQ or in an agency they work with, they rebuild the same thing over and over.
That is a beautiful job for an AI native service because the creative direction is already approved. The remaining work is repetitive and it's checkable. So how do you build it?
Well, the client uploads the approved master campaign, the brand rules, and a spreadsheet of all the locations. Then you have Astra create the production plan and keep every location's data attached to the right asset. You have Quen and Ideogram and Recraft handle copy and layout changes.
Voice and lip sync models create the approved language versions. And then QA checks every address, price, offer, phone number, and disclaimer against the spreadsheet before you actually deliver the stuff. I would target brands in the 20 to 200 location because they're big enough.
to feel the pain every month, but the marketing team is still reachable. You're not going after McDonald's here.
The international...
Franchise association directory actually low -key gives you a brand list. And LinkedIn gives you the head of franchise marketing, the local store marketing, or brand operations. You can also, by the way, look at the brand's location page and social accounts to see whether each franchisee is improvising its own creative or have AI do that for you too, just going and checking that with browser use.
Okay, so how do you think about... Outreach, how do we get customers to this idea? Take one public national campaign and just privately mock up how it could become three local versions.
You have the real store information, you watermark everything, and then you send a before and after video. And the pitch is like, hey, you've got this amazing national campaign. Here's what the approved creative could look like for Miami, Austin, and Montreal.
without asking a bunch of teams to rebuild it. And you come in there undercutting agencies by so much. You have so much more margin to play with.
And it's like, did you know you can get it for $500 a month? I would start with a paid five location pilot. The point is to brew that after every address, offer, phone number, language, and disclaimer survives the workflow correctly.
They want to make sure that... this actually works. And once it works, price the service per location, per campaign, or as a monthly localization desk.
So the expansion is built into the account, right? Five locations become 50, one campaign becomes the monthly calendar, and then one language becomes four. And then you sell it through the ecosystem.
You know, there's franchise marketing agencies or multi -unit operators, someone who owns 20 Chick -fil -A's. print vendors and private equity firms who actually have a bunch of brands with the exact same problem and are looking to reduce costs. The cool thing is one agency relationship or one of these relationships can bring you 10 franchise systems.
And then you're just built into their system, everyone's happy, they're paying less, you're producing good work, and the customers aren't getting campaigns that are not in their language or aren't catered to them. The third startup idea is how can you turn an industrial catalog into a sales team? Let me explain what I mean by that.
There are thousands. Yes, thousands of manufacturers who are selling expensive products with dense PDFs, bad photos, and product pages that look like they were last touched in 2009 because they probably were last touched in 2009. The service turns one SKU into a short product demo, an application video, a trade show loop, a distributor training clip, and a set of sales graphics.
How do you build it? Well, you get ASHA to read the spec sheets, you get those CAD files, the manuals, the approved claims, and that turns into a visual plan. The creative models show the product in the environments where it is actually used.
While the rulebook checks every dimension and feature and compatibility claim and safety statement against the actual source material. So the customer gets a useful sales package and they don't have to arrange this new shoot for every SKU, which is actually a pain point. If I were going to do this, you want to start with one vertical.
Maybe it's commercial kitchen equipment or dental labs. and just learn its language. ThomasNet has hundreds of thousands of manufacturers and suppliers organized by category and trade show exhibitor lists are another ready -made prospect database.
Look for a company selling a five -figure product with a detailed spec and product page and it has three dim photos and no useful video. The worse the marketing looks, besides the price of the machine, the stronger the lead. And I can tell you no one on X is building something like this.
These people want something like this. And then you start with one SKU and you expand to multiple SKUs over time as they do new product launches or customized versions of their products. So you're building the sales.
assets for them in real time. It's actually another boring idea that I think could print. So I gave you four business ideas.
I gave you studio operator. I gave you three more startup ideas. But there are probably hundreds of versions of this hiding across different industries.
The harder question is how do you decide which one is actually worth building? So I would start by making a list of repeated jobs from Upwork, Fiverr, industry directories and agencies that serve a niche. Then I would ignore how exciting the industry sounds and I would actually just look at the shape of the work.
So I would use four filters. The first filter I would use is, does someone actually pay for the work already? An existing budget is way easier to capture than a completely new category to explain.
The second, well, the unit has to be obvious. One SKU pack, one localized campaign, one product video, or one listing package is easier for the customer to buy and it's easier for the agent to understand. The third is just the work repeat.
Repetition gives the workflow a chance to improve and it gives you a reason to sell a subscription or an ongoing product or contract so you're getting paid monthly. Then you can quit your job or just you've built something with a lot more value in terms of exit value. Meaning when you have recurring revenue, that business is worth a lot more than something with one -time revenue.
And then the fourth is, is there a concrete definition of correct? So obviously, creativity and creative taste can be subjective.
But what can't be subjective is the product name, the SKU number, the claims, the formats, and the required scene that can all be checked. So the strongest opportunities, in my opinion, have a creative surface but a boring operational core. So the customer...
notices the campaign or the film, the company makes money because it controls intake, cost, versioning, rights, QA, and delivery. Every job leaves behind a useful lesson. Which missing input predicts revisions?
Which model preserves the important detail? Which route stays inside the margin? And which beautiful output fails the client's actual requirements?
But over time, those lessons become the software. And that's what's so cool about this age of building AI -native service firms is you're building a loop. And in the past, you were building a loop around human beings.
But human beings would leave, right? If you started an agency, and over time your agency would get smarter and smarter, but those people would leave. And because of that, they went to other places, and those lessons went to other places.
With an AI -native service firm, you're building this loop and it's getting better and better and better. The more client work you do, the more outputs you generate, the more lessons you learn, the more workflows you do. It really is a new way of doing business.
And that's what's so exciting to me about this age we're in where you can have Astra plus these high -end creative models. Creative AI is finally reaching the point where you can build a company around it. The models are becoming specialists and the APIs make them programmable.
And Astra and models, whatever Claude and Google and others are going to come out with, can run the messy work between the client request and the final delivery. If you want to start a business or if you're interested in this, I would start with one narrow... package, and then five customers.
You want to stay close to the work, write down every missing input, mistake, revision, and place the margin disappeared, and then turn those lessons into the intake, into the routing, into the QA, and into approval rules. So you get paid, basically, while you're discovering what the software needs to become. I think the next wave of one -person businesses will look like these small, capable companies that happen to have most of their staff inside workflows like this, this AI native workflows.
So I hope this got the creative juices flowing. This is an incredible time to be building. Like I said, I haven't seen too many people talk about this in this way.
So my goal with all these things is just to help you connect the dots, hopefully a light bulb or two go off. And if I do that, then I've done my job. I'll see you next time.
Have a creative day.
The Hook
The bait, then the rug-pull.
Greg Isenberg opens with a claim: the combination of Astra and creative models is barely talked about, and it's about to produce a wave of one-person businesses charging $500 to $5,000 or more for jobs found on Upwork and Fiverr. He says he's giving away the exact prompts and a working demo to prove it.
Frameworks
Named ideas worth stealing.
01:46list
Three Shifts, All at Once
Creative models got good enough
Astra became a capable orchestrator
One production API (Higgsfield) unified the models
The argument for why solo AI-native service firms are possible now rather than earlier.
Steal forany pitch for why an AI-driven business model is viable this year, not hypothetically
20:08list
The Six Build Prompts
Build the firm
Brief intake
The production layer
The model router
QA and repair
Controls and approval gates
The sequence Greg used to build Studio Operator, one layer of the business per prompt, in a deliberate order.
Steal forany Astra or agent build for a productized service business
26:42list
Four Questions Before Choosing a Model
Still image, video, or speech?
New asset or a repair?
Cheap exploration or final version?
What absolutely cannot change?
A quick filter for routing a job to the right creative model before generating anything.
Steal foranyone manually choosing between Flux, Kling, Seedance, or similar models
43:15list
Four Filters for Picking an Idea
Does someone already pay for the work?
Is the deliverable unit obvious?
Does the work repeat?
Is there a concrete definition of correct?
The filter Greg uses to separate a strong AI-native service idea from an exciting-sounding but weak one.
Steal forevaluating any productized-service business idea, AI or not
CTA Breakdown
How they asked for the click.
VERBAL ASK
20:20link
“I'm giving away all these prompts in the link in the description.”
Ties the lead magnet directly to the exact build just demoed, so the ask lands right after the viewer has seen the payoff rather than as a generic description-link mention.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Greg Isenberg and Remy put the invite-only Instinct agent through real errands, a haircut, a restaurant table, and a Bali visa, and find out it keeps a copy of your email even after you disconnect it.
A two-person breakdown of OpenAI's top-tier model that skips the game demos and goes straight to code audits, nine money-making agent prompts, and a Raspberry Pi speaker built and shipped in about 30 minutes.
A solo walkthrough of five open-source GitHub repos getting traction right now, an AI writing editor, an agent-run CRM, a video-editing agent, a skill security scanner, and a phone-controlling harness, each with the exact install command and the first small workflow to try.
Greg Isenberg names the role he thinks AI agents are about to make the most valuable job in tech, and hands over the folder structure, tool stack, and 30-day plan to become one.
Greg Isenberg and developer Vinny break down WebMCP, the experimental browser feature that lets any AI agent search, compare, and buy on a website without scraping the page, then price two startup ideas built on it.
00:37
01:29
02:40
03:16
03:58
04:14
04:49
05:10
06:15
06:55
07:35
08:15
08:54
09:34
10:14
10:54
11:43
13:42
14:41
15:27
16:40
17:39
18:39
19:38
21:07
21:46
22:25
23:05
23:44
24:35
24:51
25:42
26:22
26:42
27:39
27:56
29:13
29:42
30:27
30:54
31:41
32:09
33:17
33:40
34:13
34:56
35:39
36:21
37:04
38:30
39:13
39:55
40:38
41:50
42:24
42:58
43:15
44:04
44:38
45:35
45:49
46:19
46:35
47:40
47:54
27:30
22:52
24:43
35:18
28:58
44:21