operate.
we run it for you, every week. a director picks the winner and the audiences, agents make the versions, and a director signs each one before your media buyer sees it.
you own: the work · every file, nothing to renew
a DTC brand spending $1M+ a month on Meta made every ad one creator at a time. we rebuilt one of their winners with AI, then made its opening again with new people for new audiences, without booking a single creator.

for years, every new ad meant the same routine: find a creator, brief them, send the product, pay for the video, then wait to see how it did.
they had the scripts and knew what worked. their top ten UGC ads each spent over $1M on Meta. but every new face cost a booking, a parcel and a shoot, about $150 a video, and every creator reads the script their own way.
about $150
about $150
about $150
about $150
their way · one creator at a time · filmed, shown with permission
we took one of their winners, kept its script and its structure, and rebuilt the first 56 seconds with an AI presenter. no creator, no shoot.
play both, with sound. the first is theirs, filmed with a real person and shown with permission. the second is ours, and every frame of it is AI-generated.
their winner, filmedover $1M spent on Meta
our version, AI-generatedsame script, no creator, no shoot
the winner gives the engine three things: its script, its angles and its footage. the engine is Claude agents working through two skills, a pool of close-ups cut from their own b-roll, and locked looks for each person, room and voice.
it runs the same eight stages every time, and it never spends money on its own: a person approves every paid run. what comes out is the same ad, opened by a new person for each audience.

over $1M spent on Meta · filmed, shown with permission

JessAI-generatedyoung parents

CarolAI-generatedbuyers over 60

MattAI-generatedthe original audience

DevAI-generatedstudents
fig. 01 · the engine: one winner in, new people out
then the engine made the opening again with new people. the same script and shots, with a new person, room and voice for each audience: Jess for young parents, Carol for buyers over 60, Matt for the original audience, Dev for students.
Jess’s version and Dev’s shorter cut play with sound. every person here is AI-generated, and they present and demonstrate, nothing more: they never pose as customers or experts, never show results, and only say claims from the approved list.
the engine also makes the inserts and b-roll an ad needs, built from elements already in the video.
JessAI-generatedyoung parents
CarolAI-generatedbuyers over 60
MattAI-generatedthe original audience
DevAI-generatedstudents
new faces · approved shots from each person’s generation, cut together · every person here is AI-generated
the cap, held to the lightAI-generated
a drop on the fabricAI-generated
fabric close-upAI-generated
down the drainAI-generated
inside the empty plasticAI-generated
new visuals · inserts and b-roll built from the same elements
it used to work like this: you picked the audience (age, interests, lookalikes) and Meta showed your ad to those people. one strong ad could run to every audience you chose.
since Andromeda, Meta does most of the targeting. it reads each ad, decides who it fits, and groups ads that look alike as one. so the way to reach more kinds of buyers is to give Meta more ads that are actually different.
near copies land in the same spot and compete for the same people.
JessAI-generatedyoung parents
CarolAI-generatedbuyers over 60
MattAI-generatedthe original audience
DevAI-generatedstudentsdifferent ads land in different spots and reach different people.
fig. 02 · why one winner now needs many versions
bid × estimated action rate + ad qualityMeta’s auction ranks every ad this way. creative moves the two underlined parts, so an ad that fits the person in front of it can win the auction for less.
building it taught us a procedure, so we wrote it down: two skill files, playbooks the agents follow step by step. ugc-ad-pipeline holds 8 stages, 4 reference files and 3 tools. shot-builder holds 10 reference files and the locks for the person, the place and the look.
the agents run the same eight stages in the same order on every run, a person decides at six of them, and every mistake becomes a new rule. one from the first build: never start a voice track on silence.



fig. 03 · where the skills came from



decidespicks the person, approves the sheet
decidesapproves each sheetfig. 04 · eight stages, a person decides at six
when an ad wins, you should not have to start over to get more of it. each week the engine makes it again for the audiences you choose.
a director picks the winner and the audiences, agents make the versions after a yes, and a director signs every version before your media buyer sees it. what performs comes back as next week’s brief.


fig. 05 · keep every audience supplied with what already works
one creator at a time.
find a creator, brief them, send the product, pay for the video, wait to see how it did. about $150 a video, one face per ad, and every creator reads the script their own way.
one script, a new face per audience.
agents run two skills on their script and their own footage. a person approves every paid run and signs every version before a media buyer sees it. about $1.50 of model cost per new version.
same script, same claims, a new person for each audience.
model costs come from our run logs, and the creator figure is their own average. direction and editing time is priced in the proposal, so read these as model costs, not the whole bill.
generated people present and demonstrate. they never pose as customers or experts, never show results, and only say claims from the approved list. that rule is in the skill files, not in somebody’s memory.
the costs on this page are model costs, from our run logs. a person still directs, edits and signs every version, and that time is priced in the proposal, not hidden in the 99%.
the next ad starts from these files. a new person is one new sheet. the engine points at any account where one ad already wins and the people it has not reached yet are the ones you want.
from their winner to yours
send one winning ad and any raw footage you have. within 48 hours you get a written proposal with two ways to work together. within 5 working days you get that ad made again for three of your audiences, free.
we run it for you, every week. a director picks the winner and the audiences, agents make the versions, and a director signs each one before your media buyer sees it.
you own: the work · every file, nothing to renew
we build it inside your team and train the people who run it: the two skills, the locks, the gates. then the engine is yours to run.
you own: the system · skills, locks, prompts, training