Start on Pinterest. Find the aesthetic you actually want, Y2K, old money, harajuku, clean girl, whatever it is, and pick one image to carry across as your reference.
One, not a pile. Arcads takes a reference image, and the job you give it has to be specific or it gets averaged away. Tell it what the reference is for, in capitals, rather than saying "match this": REFERENCE IS FOR the colour grade and the quality of the light. Nothing else.
Images are the cheap part. Arcads gives you 100 free images a day on Nano Banana Pro and 100 on Seedream, and their own credits page says those "do not consume paid credits within the daily free limit". So generate a lot and be fussy. Make it vertical from the start, because cropping a wide image later destroys the off-centre framing you are about to ask for.
A candid photo of [describe your person: age, hair, what they are wearing], [what they are doing], shot on a phone front camera in [describe the room and the light source]. Vertical 9:16, 1080 by 1920.
LENS
Roughly a 24mm equivalent, deep depth of field. The room behind her stays nearly as sharp as her face. No background blur, no bokeh.
SKIN
Real skin at close range: visible pores across the cheeks and nose, a little oil on the nose and forehead, faint redness under the eyes and around the nostrils, uneven natural tone, fine flyaway hairs at the hairline. Do not smooth, retouch or even out the skin.
LIGHT
Available light only, from [the window / the ceiling light], raking across one side of her face so it skims the skin and casts small shadows in the texture. A mild colour cast from the room. Not studio lighting, not a ring light, no fill.
EXPRESSION AND FRAME
Mid-sentence and slightly asymmetric rather than posed or smiling at the camera. Framing casual and a little off-centre, as if nobody composed it.
REFERENCE IS FOR the colour grade and the quality of the light. Nothing else.This is where most people lose it. A perfect still turns into obviously-generated video the moment the motion is too smooth, so the flaws here are camera flaws rather than skin flaws.
Animate this still as one continuous handheld phone clip, vertical 9:16, 1080 by 1920, [8] seconds, 30fps, no cuts.
SOURCE LOCK
Same person, same face, same hair, same clothing and jewellery as the still, in every frame. Keep the skin texture from the still exactly as it is: the pores, the oil, the faint redness, the flyaway hairs all stay. Nothing gets smoothed, brightened, slimmed or relit as it moves.
CAMERA
Front camera held in one hand at arm's length, roughly a 24mm equivalent, deep depth of field so the room behind her stays nearly as sharp as her face. Constant small drift and micro-shake: the frame never settles perfectly still and it never glides. The framing creeps off-centre and she corrects it once. Focus snaps once in a single step rather than racking smoothly. Digital stabilisation makes the edges of the frame breathe slightly.
EXPOSURE
Phone auto-exposure, visibly working. When she turns toward the window the bright side clips to pure white with a thin halo along the edge, then the whole frame pulls down and resettles over about half a second. Shadows stay lifted and slightly noisy, luminance noise in the dark parts rather than film grain.
SHUTTER
1/60 at 30fps, so a quick hand or a head turn smears a little instead of coming out crisp.
MOTION
She moves like a person mid-thought, not a performance: one small weight shift, a hand comes up near her collarbone and goes back down. Her eyes leave the lens once and come back. One blink that is not on a beat. Her hair swings slightly late and settles.
AUDIO
Room tone only: the hum of the space, faint traffic or a fan through the wall.smooth skin, plastic skin, airbrushed, beauty filter, shallow depth of field, creamy bokeh, studio lighting, ring light, colour grade, slow motion, gimbal glide, captions, subtitles, watermark, extra fingers, warped hands, morphing face, flicker. If it has no negative field, delete the list rather than pasting it anywhere.A UGC ad is someone talking, so this is the step that decides whether yours works. Arcads is built for exactly this: you pick the actor, write the script, and it generates a lip-synced video, at 800 credits per minute per actor. The mouth is the part everyone worries about and the part the models are now decent at. What gives a talking clip away is the eyes going still, the pauses, and a voice that sounds like a narrator instead of a person who picked up their phone.
Animate this still as one continuous vertical 9:16 clip, 1080 by 1920, [8] seconds, 30fps, no cuts.
She is talking to her own front camera, phone held in one hand at arm's length.
She says: "[your line, about twenty words for eight seconds, short sentences, numbers spelled out as words]"
VOICE
Her own speaking voice, close on a phone mic, a little bright and a little boxy from the room. Conversational and uneven: she leans harder on one word than the rest, runs two words together, and takes one audible breath before the last sentence. Not a narrator, not an announcer.
MOUTH
Her lips close fully for a frame on every p, b and m. Her jaw opens only as far as each vowel needs. When she stops speaking her lips come to rest closed and stay still, no chewing.
FACE
Her eyes do the work. Her gaze flicks off the lens and back once while she thinks. Two blinks across the clip, neither on a beat. A small nod into the word she stresses.
CAMERA
Front camera in one hand, roughly 24mm, deep depth of field so the room stays nearly as sharp as her face. Constant drift and micro-shake, never a glide. Focus snaps once rather than racking.
ROOM SOUND
Her voice sits in a real room: faint reverb off a hard wall, the hum of the space. Her voice gets slightly louder as she leans in.
LOCKS
Same face, hair, clothing and jewellery as the source image in every frame. Keep the skin texture exactly as it is. Clean frame, nothing written on it.Add any of these to any prompt. Reach for two or three, not all of them, because piling on every flaw at once reads as a filter rather than as reality.
SKIN AND FACE
visible pores across the cheeks and nose
oil on the nose and forehead catching the light
faint redness under the eyes and around the nostrils
a few fine flyaway hairs at the hairline
slightly chapped lips
uneven undertone, one cheek warmer than the other
an asymmetric expression caught mid-sentence
LENS AND SENSOR
deep depth of field, the room nearly as sharp as her face
luminance noise in the shadows, not film grain
the window clipped to pure white with a thin halo
digital stabilisation making the frame edges breathe
a slight fingerprint smudge softening one corner
vertical lines in the room skewing for a frame on a quick move
CAMERA
focus snapping once in a single step, never racking smoothly
exposure resettling over half a second after she turns to the light
handheld micro-shake, never a glide
framing creeping off-centre, corrected once
VOICE AND MOUTH
one audible breath before the last sentence
she leans harder on one word than the rest
two words run together
a half-beat pause mid-thought, not mid-sentence
lips fully closed for a frame on every p, b and m
gaze flicking off the lens and back once while she thinks
two blinks, neither on a beat
ROOM AND LIGHT
one practical light source with a colour cast
light raking across one side of the face
a slightly cluttered background that nobody tidied
mixed lighting, warm lamp against cool daylightAn ad set is three to ten clips of the same actor. Every generation is a blank slate, so continuity is something you enforce by restating, never by writing "as before".
UGC means ads, and ads mean variants. Once one actor and one room are working, the expensive part is done.
Hold fixed: the actor, the room, the light, the lens language, the lock block. Change only: the spoken line and the first three seconds. That is where the hook lives and it is the only part worth testing.
Six to ten variants is enough to learn something. Changing the actor and the hook at once teaches you nothing, because you cannot tell which one moved the number.
She says: and regenerate.Build AI systems with me through recorded lessons you can follow at your own pace. Join the waitlist for first access and the founding price.