quick context, so the rest makes sense: i'm building aesty.ai — an AI fashion agent for the wardrobe you already own: it plans your outfits, flags what's actually worth buying, asks your friends for their take, and tries looks on your avatar. currently we have ~560 active paying users
aesty.ai app store page
i'm a programmer, not a marketer — and i'm more and more convinced that marketing is turning into an engineering discipline. instead of leaning on a trained eye for what goes viral — the kind that takes years to build and that you either have or you don't — you encode the loop, generate and test dozens of hypotheses with AI, and let the data pick the winners so you don't out-clever the problem, you out-iterate it and the wild part is this already works today, on Opus 4.8 — imagine where it goes the moment we get our hands on GPT 5.6 and Fable (hopefully!!)
my last article covered the cheap layer of that engine — how i get 100k+ views a week on tiktok with carousels at basically $0, and it did really well. This one is the expensive layer of the same machine: video
some aesty top performers
video is also where the reach really compounds and it's the first point in the funnel where things stop being free. the good news is that a performing AI UGC video now runs me around $2-4 to make, where i used to pay creators $150-250 for the same thing. the harder problem isn't making the video cheaply, though - it's running 15+ accounts without quietly getting throttled, which is why this post covers both halves: how i build a video for a few dollars, and how i keep the accounts healthy while i do it. mostly healthy, anyway — i've taken exactly one shadowban, and i'll get into how that happened further down.
before any of that, the principle the whole thing rests on: video is expensive relative to carousels, so you only ever make one for something the cheaper layers have already proven. i think about it as a ladder of increasing cost that you climb one rung at a time, without skipping ahead:
- trendwatch (~free) — figuring out what's already winning before i make anything. Claude digs through competitors and their hidden and partner accounts for the formats that are actually landing right now, weighing saves and shares over likes, and pulling only from small-to-mid accounts where the format is doing the work rather than the follower count.
- carousels (cents per post) — testing which messages and which creator looks genuinely land with the audience
- video (real money and real time) — made only for the messages and faces that already proved themselves on carousels
- distribution via Postiz — once a piece is proven, Postiz ships it for me: it schedules and posts every variant across all the accounts with Claude Code
the reason most people decide organic doesn't work is that they start at the third rung: they spend their budget filming a "perfect" video for an idea nobody has validated, it flops, and they walk away. by the time i make a video it isn't a bet anymore but it's me pouring fuel on something that's already burning.
trendwatch — what to make (~free)
same starting point as always: i don't invent concepts, i find what's already working and copy the format. i packaged the whole process into a self-bootstrapping Claude Code skill — it pulls competitors' recent videos, filters for proven-but-transferable formats (placement + sub-500k accounts), reads the comments for the audience's actual language, and hands me ranked hook ideas with reference URLs i can go watch.
skill body
grab the trendwatch skill here
the output of this step for sure isn't "make this video" but a shortlist of hooks and formats worth testing cheaply first.
carousels — test cheap (this is your filter)
this is the rung everyone skips. before a single video gets made, i run the candidate messages and creator looks through carousels, because carousels cost cents and tell me two things i'd otherwise pay a lot to learn:
example of carousels
- which message actually resonates (save rate is the tell, not likes)
- which creator look the audience responds to (different face, different aesthetic — same hook)
a message that flops as a carousel will flop as a video too, just for 5x the cost. so carousels are my filter, only the hooks and faces that earn it on carousels graduate to video.
video — build only what graduated
once something's earned a video, the build is four parts. i mix and match the tools by what the shot actually needs, because the cost difference between them is large.
1. the creator's look — Higgsfield Soul. every video has a recurring AI persona, and i lock the look with Soul so the same face shows up consistently across everything that creator posts. The appearance should be catchy
some aesty AI creators
2. the first frame — copy what already worked. the opening frame is the whole ballgame for the thumb-stop, so i usually recreate the first frame of a video that already went viral with nano banana pro / GPT image 2. Always generate in good resolution (not 1k)
3. the video itself — match the model to the shot
- Kling 3 for simple UGC — basic actions, no phone in hand, no speech. it's cheaper and totally good enough for this
- Seedance 2 when i need to show product detail, lip-synced speech, or more complex scenes. it costs more, so i only reach for it when the shot genuinely needs it
4. the demo at the end — assembled not generated every time. the product demo (the screencast that actually shows aesty working) i build with Remotion on a fixed hook + demo structure. because it's assembled from clips, i can reuse different pieces across many videos instead of remaking the demo every time. For the first couple of times I recommend doing that manually with capcut or other video editor. Develop your working structure and follow it in every video you make. For me it is usually 3-5 sec video hook + 5-7 sec demo
putting all together - these videos were generated with Higgsfield
This is where the $2-4 comes from. the persona face is a one-time cost i reuse forever, the first frame is cents, and the demo is assembled for free from clips I already have — so the only real per-video spend is the generation itself: a couple of dollars on Kling 3 for a simple shot, a bit more when a scene genuinely needs Seedance 2. That's the entire reason a performing video costs at a few dollars instead of a few hundred, i'm only paying for the one part that has to be generated fresh.
video is the one stage i don't fully automate — i watch every one with my own eyes and the metric i'm watching is early retention: at least ~60% hold through the first 3-5 seconds. That's the line between a video the algo keeps pushing and one that dies in the test bucket so the ones that clear that bar are the ones i put paid budget behind later for meta/tik tok paid tests, the rest just keep running organically
distribution — Postiz
Postiz interface
same as with carousels, i mostly don't post this by hand. the whole video distribution layer runs through Postiz, and the part that matters most here is that Claude Code (or Codex/etc) drives it directly through the API. that means one proven video doesn't have to get posted just once — you can spin up variants with different hooks, different openings, and a different ordering of the demo clips, then schedule each one through
Postiz under its own angle across all the accounts. a single video that works turns into a whole week of posts that way, each framed a little differently and all booked automatically, without me ever touching a calendar.
the part that keeps it alive — not getting shadowbanned
none of this matters if the accounts get throttled, and once you're running more than a handful that becomes a real risk. the reassuring part is that staying on the algorithm's good side isn't about clever tricks and it's mostly about behaving like a real person who's genuinely worth watching
the things i actually stick to:
- never post duplicates. the one time i got shadowbanned was entirely my own fault — i accidentally posted 3 identical posts in a row. it recovered, but the lesson stuck: vary every output and never let a feed look copy-pasted.
- warm new accounts up slowly. a brand-new account that blasts 10 posts on day one reads as automated. ease in, post at a human cadence, and let it settle into a niche over the first week or two.
- keep each account coherent. one account should behave like one real person with a consistent niche, voice, and rhythm — not a generic firehose pointed at everything.
- optimize for saves and watch-time. the algorithm rewards content people actually find useful, and chasing fake engagement is exactly the kind of thing that gets accounts flagged.
it's just giving the algorithm what it's already optimizing for and staying inside the rules while you do it. boring, but it's what keeps the engine running.
the stack
- Claude — runs the trendwatch skill, decides what graduates from carousel to video, and drives the re-cutting + scheduling.
- Higgsfield Soul — locks the recurring creator look.
- nano banana / GPT image 2 — the first frame.
- Kling 3 — simple UGC shots (no phone, no speech).
- Seedance 2 — detail, speech, and complex scenes (when it's worth the cost).
- Remotion — assembles the reusable hook + demo screencast.
- Postiz — schedules every variant across every account, driven by Claude via API.
the through-line is the same as the carousel post - climb the ladder, spend only where the data already points, and let the machine handle everything that doesn't actually need your double check
