Found something good?

first, a bit of context so the rest makes sense. i'm building aesty.ai — the "cookbook for outfits". it's an AI app that digitizes your wardrobe and shows you in outfits built from clothes you already own: we recognize the clothes in your gallery, then style looks and render them on you, so "i have nothing to wear" turns into "you have 47 outfits, you just couldn't see them." two engineers, no marketing background and no team behind us.

that last part is exactly why this pipeline exists. for a tiny consumer app, paid acquisition is brutal — tiktok and meta ads carry real CPMs, and the CAC quietly eats your money before you've even learned what message works. but the people we need are already on tiktok, screenshotting outfits they'll never actually recreate, so the channel was never the question. the problem was that doing it properly — posting every day across enough accounts to matter is a full-time job i didn't have and couldn't afford to hire out. so i built a machine that earns it for basically the cost of ~150 image generations.

and here's the part i'd want a younger me to hear: i'm not a marketer, i'm a programmer. i don't have a natural instinct for hooks or a feel for what goes viral, and for a long time i assumed that ruled me out of growth altogether. what changed is that with Claude Code you can just vibe-code the whole engine now — something that iteratively hunts for formats, tests them, keeps what lands and kills what doesn't, and wins not by being a marketing genius but by sheer number of iterations.

i'm in amplitude with claude a lot, and digging into the analytics i found that now >50% of aesty.ai's initial-purchase revenue was coming in with almost no ad spend behind it. it was all from tiktok carousels: accounts pulling 100k+ views a week in total and bringing in roughly 15-25 new paid users a week at a CAC of basically nothing. i'm barely involved at this point, and i don't mean "i hired a team to run it," i mean the machine genuinely runs itself.

the unit economics are what made me commit. for us it works out to roughly one purchase per 4-6k views, and my honest rule is simple: if a channel converts above zero and costs me nothing beyond some image generations (that i can reuse with different text overlays), i'm going to scale it. this one does both, so here we are.

Claude turned "what do i even post" into a thirty-second answer. the only thing left that genuinely doesn't scale is the boring middle — making content every single day, across a lot of accounts, and getting it out at the right time without a human sitting there doing it. so that's the actual game now. not making content, but shipping it at volume without babysitting it.

and the timing of this post is perfect. Claude Fable 5 dropped this week — anthropic's new model that regular people can actually get their hands on — and my entire engine got smarter overnight without me touching a single line of it. i swapped the model, re-ran trendwatch, and everything downstream — the hook analysis, the copy, the scheduling decisions — now runs on a better brain. that's the part i want you to sit with: when your growth is a pipeline driven by a model instead of a team of humans, every model release is a free upgrade to the whole machine. an agency gets more expensive every year, this thing gets smarter every release day, for the same near-zero cost

the recipe has four parts: trendwatch, generate, distribute, and compound. the whole thing looks like this:

here's how each part works

1. trendwatch — don't guess, reverse-engineer

most people trying to grow do it backwards. they sit down to invent the perfect hook, make it once, post it, pray, and then burn out three weeks later when it doesn't land. my number 1 rule is don't invent concepts, find what's already working and copy the format. every carousel that already popped has done the hard part of proving the hook works, so there's no reason to gamble on your own when proven ones are sitting right there.

the catch is that you have to read the data correctly, and most of what looks like signal isn't. a few things i learned the hard way:

i packaged this whole process into a self-bootstrapping Claude Code skill. you drop it into your product's repo, point it at your landing page, and it handles the rest: it asks around five strategic questions, pulls competitors' recent videos with a double filter (placement keyword plus a sub-500k follower cap), classifies them, generates ten hook ideas each with a reference URL you can actually go watch, and — the part that matters most — it accumulates learnings across runs so it gets sharper the more you use it

the first thing to do with any skill or prompt is ask it to explain how it's built — i'll sometimes ask for a flowchart. that's how you actually understand it well enough to control it and adapt it to yourself instead of just running it blind.

iterating is where the real compounding happens. a brand-new account gets a small test bucket of a couple hundred views per post, so breaking 500-700 views on a cold account means you've passed the algo's first gate, and that's the signal to stop ideating and start iterating.

on attribution, don't overthink it. either ask "where did you hear about us?" on onboarding, or hand out a promo code and that's plenty to see what's actually driving paying users

2. generation — consistency first, then volume

making content is cheap now, but "cheap" has a specific shape that most people get wrong, and getting it right is the whole reason i can run dozens of carousels a month for almost nothing. genuinely expensive step is generating a photorealistic face (however faceless content is also an option), so i generate a face exactly once per persona and reuse it forever. each persona gets one locked reference selfie plus around 15 "cover" photos edited from it, which comes to roughly $5.50 to stand a persona up, and after that every carousel is built by compositing caption text over those same covers for free.

one-time per persona (the only money you spend):

every carousel = a banked cover + app screens + PIL caption text

- 4-7 slides at $0, the same covers reused across 10+ carousels

two rules keep that both cheap and consistent. the first is never edit an edit: every cover is a single-hop image-edit of the original locked reference, never an edit of an edit, because editing a generated image compounds drift and within a few hops the face stops being recognizable

the second is no baked text in the photos, ever — the covers are generated with an explicit no-text clause, and every word (captions, headlines, the @ handle watermark) gets composited on afterwards in a separate PIL layer. that second rule is also what fixed the "text printed twice on top of itself" problem i used to hit, because once the text lives in an overlay instead of the image it's always editable, correctly spelled and on-brand, and you never re-pay a generation just to fix a typo. consistency itself i hold by writing the persona's lore into a folder and referencing it on every build, so the same face and the same voice come out across every post — the same format reads completely differently for a 22yo british budget-stylist than a 34yo busy mum, and that difference is the entire point. Generated with new Fable 5 model:

which model i reach for depends on the slide: Higgsfield Soul (via MCP) for developing the character, GPT image 2 for the faceless carousels, and nano banana pro for the human-face covers. whatever you generate, mind the carousel safe zones so the text and CTA don't get cropped by the UI.

the last step before anything goes out is a hard gate, and it's non-negotiable because the pipeline auto-schedules. i render all the slides with nothing posting, then actually claude looks at them to confirm the same face and defining features show up on every slide, the covers read UGC-natural instead of glossy, and the captions and @ handle are correct — and only the carousels claude approves get scheduled. underneath all of it the mindset is still that volume beats everything: more posts simply mean more data and more hypotheses tested, and a low conversion rate per post is completely fine when each post costs almost nothing to make

3. distribution — stop posting by hand, use Postiz

running many accounts by hand means many logins, dozens of uploads and constant timezone math, every single day, and nobody keeps that up for long — it's a full-time job, and it's exactly why most people's "content engine" quietly dies in week three.

i run the entire distribution layer through Postiz. it's open-source, and the part that matters to me is that its API works really well with Claude Code, so i don't open the dashboard at all, and it even adds music to the carousels on the way out. you map each account to its Postiz channel once and after that the batch scheduler does the rest. the bit i lean on hardest is that it's gap-aware: it builds the pool of upcoming time slots (i run 08:30 / 11:30 / 21:00 local), asks Postiz which slots are already booked on each channel, drops those, and drops each carousel into the earliest free slot, so it never double-books anything

there's one more rule baked into the scheduling: no slide-1 "hook photo" repeats within a single posting day, and if a cover has to repeat it gets relegated to the CTA slide instead, so a feed never looks like copy-paste. the one time i got shadowbanned was entirely my own fault — i accidentally posted 5 identical posts in a row — and even then everything recovered after a couple of days. the lesson is just to vary your output and never fire literal duplicates, because the system runs on trust, not on spamming the same thing.

4. compound — exploit winners, spin up new accounts

once posts are flowing, the loop gets simple. you watch which formats are landing, then exploit them by running variant after variant while holding the message constant, and when a format has clearly proven itself i spin up new accounts modeled on the winners and run the same playbook there. the best hooks don't even have to stay carousels — i take the top performers and turn them into video (with Seedance 2 for example), which gives a proven hook a new surface and more reach for basically no extra creative risk.

300-500 views per post, three posts a day, across ~15 accounts, over seven days, comes out to 100k+ views a week. at our rate of roughly one purchase per 4-6k views, that's 15-25 new paid users every week, at a CAC of effectively ~zero. if something costs nothing and converts above zero, you scale it. build the engine once, let it run, and go back to building the product.

the stack (everything i used)

content was never the moat, the engine is. build it once, feed it data, and go back to your product — mine is shipping a new retention feature for stylists while this thing ships carousels. let's go!