I’m going to walk you through the AI UGC machine i'd run with Opus 5.5, from the lazy version where someone else handles the hard part to the full build in Claude Code, step by step...
because content changed shape a while ago, and now one good video isn't enough, you need hundreds of them, on brand, with the same creator, posted across TikTok, Reels and Shorts every single day
and before we touch a single tool, you need to hear the one rule that decides whether any of this is worth your time
a lot of people chase the fun part, prettier generations, a smoother face, a new model every week... and none of it counts if nobody sees the videos
a folder with 300 perfect clips and zero views is a hobby, not a growth system
so if you have no plan for where your videos go and who posts them, keep that in mind the whole way, because we're coming back to it
here's what we'll cover, in the order you'd actually build it:
• the easy route where the distribution side is handled for you
• what Opus 5.5 actually does in this setup
• the folder that runs the whole machine
• the five stations from angle to finished clip
• the quality check that keeps your accounts clean
• the weekly loop that keeps it all moving
• a four week plan so you're not staring at a blank folder
start with the easy route
i'll be straight with you, the Claude Code build later in this article is fun, but plenty of you don't want to build a distribution system, you want your videos posted and seen while you run the business
that's where[[Affiliate Network]](https://link.affiliatenetwork.com/dan fits, because it handles the part that breaks everyone once the videos start stacking up
you open a campaign, set your CPM, which is what you pay per 1,000 views, and write your rules in normal words, like you'd brief a new hire
creators apply, post your content, and every view runs through bot checks and your rule checks before you pay a cent, then payouts go out automatically
you can bring creators you already work with, or tap into a marketplace of around 200k to 250k creators ready to post
that last part is why this setup works, because once your Opus 5.5 machine fills a folder with finished clips, there's somewhere for them to go that isn't you logging into 40 accounts
what Opus 5.5 actually does here
Opus 5.5 came out on September 22, and the first thing to get right is that it doesn't make videos, it takes text and images in and gives text back
that sounds like a limitation, it's actually the point
video models are good at turning a prompt into a clip, they're bad at deciding what clip to make, keeping one creator consistent across 200 videos, and telling you which output is garbage
that's the part that eats your day, and that's where Opus 5.5 earns its spot
so think of it like this, Opus 5.5 is the director, your video tool is the camera, and distribution is the screen
a few things make this model a real fit for the job:
• input is $4 and output is $20 per million tokens, 20% below Opus 5, and Anthropic says typical jobs cost around 40% less because it uses fewer tokens per task
• cache reads cost $0.20 per million tokens, so the stuff you reuse in every run costs almost nothing
• it has a 1M token context window, so every winning script and every comment thread fits in one go
• through the Batch API it can return up to 300K tokens in one job, which is hundreds of scripts at once
• it reads images, so it can judge frames from your clips
• the effort setting goes from low to max, so you pay for deep thinking only where it matters
you can check all of that on [[Anthropic's Opus 5.5 page]](https://www.anthropic.com/claude-opus-5-5
The folder that runs the machine
the setup is one folder on your computer, and that folder is the whole brain of your UGC agent, so it's worth getting it right
inside it, i keep four files the agent reads every single run:
• a brief that explains the product, who buys, what they're scared of, and what the video should make them do
• a creator bible with the name, look, room, lighting, the way they talk, and three reference images of the same face
• a hook library with every hook that already worked, plus the numbers behind it
• a running log where every run adds a new entry and never deletes an old one
the creator bible is the file that saves your accounts, because the moment the face changes or the room changes, viewers feel it and scroll
the log is the one that saves you later, because three weeks from now you won't remember why a batch flopped, and the agent won't either unless it wrote it down
and here's the Opus 5.5 trick, you put the brief and the creator bible at the top of every request and cache them, so they ride along with every script for cents
The five stations
station one is angle mining, and every batch starts here, not with scripts
you give Opus 5.5 the raw material, top comments from competitor videos, product reviews, Reddit threads and your own best performers, and ask it to pull out the pains, the objections and the exact words people use
then you ask for 20 angles, each with one line on who it's for and why they'd stop scrolling
run this one at high effort, because one strong angle can carry fifty videos and this is where thinking harder pays for itself
station two is locking the creator, which you already did in the folder, so every request from here on reads the bible first
station three is scripts and prompts in bulk
for each angle, Opus 5.5 writes 10 scripts with different hooks, then turns each script into shot by shot prompts in the exact format your video tool wants
the structure i give it is simple, a hook in the first two seconds, the problem, the moment they found the product, the proof shown not told, and a soft close
run this through the Batch API at low effort, because you don't need deep thinking to write variation 37 of a hook you already approved
station four is generation, where the prompts go to your video tool's API
if your tool has no API, Opus 5.5 can drive it through computer use and click through the screens the way you would
the easiest way to wire all this is to let Opus 5.5 build the pipeline for you in Claude Code, you describe the loop in plain words and it writes the script that reads the angles, writes the prompts, calls the video tool and saves every clip with its script ID
it sounds scary, but it takes one sitting and you only do it once
station five is the quality check, and it gets its own section because it's the one everyone skips
The quality check
this is why so many AI UGC accounts fill up with weird hands and melting faces, nobody checked before posting
the agent pulls three frames from every clip and sends them to Opus 5.5 with the creator bible
then it scores each clip on four things:
• does the face match the reference
• does the room and lighting match
• are there broken hands, text or objects
• does the first frame actually work as a hook
anything under your bar gets regenerated, anything above it goes into the ready folder with a caption Opus 5.5 already wrote
the rule i give the agent is that every fail comes with the frame and the reason, and if it can't tell, it says so instead of guessing
a few warnings before you trust the scores:
• the model is judging stills, so it can miss a glitch that happens between frames
• lip sync and voice still need a human ear on a sample of every batch
• a clip that passes QC can still be boring, and only real views tell you that
Run the loop every week
this is the part that turns a one off batch into a system
the loop i run is short, and the agent repeats it every week:
• pull last week's views and results per video from your campaign dashboard
• tag which angles and hooks won and which ones died
• add the winners to the hook library and retire the losers
• mine five new angles from fresh comments
• batch the next round of scripts and prompts
• generate, QC and drop the clips in the ready folder
• wait for your yes before anything goes live
• write what happened into the log
one change at a time sounds slow, it is, on purpose, because if you change the creator, the hooks and the length in the same week, you'll never know which one did the work
and this is where the rule from the start comes back
a batch that looked amazing and got no views goes in the log as a miss, and a rough looking batch that pulled views and sales goes in as a win, no debate
that's also why distribution matters so much here, because one account posting 200 videos doesn't get 200 times the views, it gets flagged
with a campaign on \[Affiliate Network], the same ready folder goes out through real creators, and you only pay for views that passed the checks
here's what that looks like at scale on one dashboard, $65,603 spent, 48.8M views, at a $2.50 effective CPM
Your first four weeks
here's how i'd lay out the first month if you were starting from zero today, and i'd follow it in order instead of skipping to the fun part
in week one, you write the brief and the creator bible, pick your video tool, and set spend limits on the API before i is the first real run
in week two, Opus 5.5 mines your angles and Claude Code wires the five stations, then you run one small batch and read every QC report yourself
week three is launch week, you open your campaign, set the CPM and the rules, and send the first ready folder to creators
in week four, you turn on the weekly loop and start measuring, and you give each batch a couple of weeks before calling it a win or a miss
after that, the machine keeps going with you making one decision a week instead of writing scripts all day
this is the whole setup, and none of it needs you to be a video editor, it needs you to admit which videos get seen and which ones just look good
if you want to build it yourself, open Claude Code, make the folder, and write the creator bible tonight
and whatever you pick, make sure the videos have somewhere to go before you make a single extra clip
if you want the distribution side handled from day one, you can \[launch your UGC campaign on Affiliate Network here
