this is a practical guide: by the end u will know how to write a ugc script that doesn't sound like slop, generate the video, cut it, and test it on a real human. plus what it all costs. around 20 minutes once ur set up, a few dollars
so u know where this is coming from: not sponsored by higgsfield or anyone else. im not a ugc operator either, i make ai ads. ive been following the people actually printing with this closely since december, mostly my friends @maxxmalist and @Mho_23, along with some others, and running tests of my own. every number has a source or a screenshot, and the screenshots are me doing this exact run for the first time
why now
ive always tested ai videos on people around me to see if they believe its real
for years we could basically always tell. the skin gave it away, that plasticy dead look that turns anything cool into ai slop
then seedance 2.5 dropped and the skin finally looks human. this post from @maxxmalist is just one of many examples that show how crazy this model really is:
https://x.com/maxxmalist/status/2085330197524390113
https://x.com/eptwts/status/2085141769755242580
start with the script, not the model
the script is where you lose or keep the viewer. always will be
every ai ugc guide spends its words on the generation part, partly because generation is what the sponsor sells. ep's (@eptwts) seedance workflow breakdown is genuinely good
but scroll ur own feed: the clips that hold u are the ones where the first line hits and the next line makes u wait. thats writing, and the model cant do it for u
step 1: build ur judge panel (one time, ~10 min)
u dont learn to write scripts by writing more of them alone. u learn by getting torn apart fast, hundreds of times, by readers meaner than ur audience
heres the system:
- make a separate claude agent that only does scripts
- give it a panel of judge skills, each with one job: one for pacing, one for vocab, one for ideas, one for structure
- train each judge on a big swipe file from whoever fits ur style and taste: transcripts, scripts, hooks, ideas of the writers and creators u rate. gary provost for rhythm is a classic pick, but anyone whose writing stops ur scroll works. my actual suggestion: find an imperfect source, like a real tiktoker, and upload their entire profile to the scripting agent. polished writing is not what a ugc script should sound like
- feed the agent a bank of scripts that already performed, urs or ones u study
- every script u write goes through the full panel before u generate a single frame
a skill is just a file of instructions the agent loads: who the judge is, what it cares about, what makes it kill a script. write each one like a brief to a brutal crew member with exactly one job
the best thing you can do here imo is to make at least one of the judges 100% trained on a real UGC creator, you can scrape all their videos + scripts
(if you're confused just paste this article into claude)
i run seven judges on everything i publish, including this article (size urs however u want). it is the single reason my stuff doesnt read like slop
write ur script, let the panel judge it, rewrite, run it again. when it survives, u have something worth spending money on (usually helps to do 10+ at a time for ai ugc)
step 2: generate the video
seedance 2.5 is the model that changed my mind. higgsfield might be the easiest place to run it, dreamina and freepik work too
if a ugc video already converted for a product like urs, use it: feed it to gemini, ask for a 1:1 timestamped breakdown of everything on screen. the breakdown becomes ur prompt, ur judge-approved script goes in as the dialogue (ep wrote the deep version of this workflow)
starting from scratch instead? two prompts: a first frame, then the video
make the first frame in gpt-images-2 (nano banana works too). the trick is stacking realism keywords so u dont get the polished ai look:
ultra realistic iphone front camera selfie, woman in her mid 20s sitting in the drivers seat of her parked car, natural afternoon light through the windshield, messy bun, oversized hoodie, real skin texture with visible pores, light under-eye shadows, candid mid-sentence expression, one hand raised talking to camera, shallow depth of field, soft warm tones, authentic tiktok vlog aesthetic, no text, no captions, 9:16
then feed that image into seedance 2.5 as the reference frame and prompt the video around ur script. structure it like a shot breakdown, second by second:
subject: the woman from the reference frame, same hair, same hoodie, same car
camera: handheld front camera selfie perspective, chest-up framing, natural micro shakes, 9:16
audio: clear phone-mic voice with light room tone, no background music
00:00-00:03 she leans slightly toward camera, mid-sentence energy: [ur hook line]
00:03-00:12 natural blinks, small head movements, one pause mid-thought: [the body of ur script]
00:12-00:15 slight smile, closing line straight to lens: [ur cta]
keep the skin texture, no beauty-filter look, lips synced, 15 seconds
set 9:16, start at 720p, upscale when u like the take. match the video length to ur timestamps or the delivery desyncs
seedance will render a weak script just as beautifully as a strong one. thats the trap the whole industry is about to walk into
step 3: edit it like real footage
the generation is raw footage, so treat it like footage: cut the dead space at the start, tighten every pause, add captions, drop a sound under it, the same way u would edit a real creator's selfie video
when ur flow works, build the machine
keep running ur own tests until the clips consistently pass as real. tweak the judges, the prompts, the edit. that flow is ur actual asset
then wispr flow the entire thing into fable 5 and build ur own machine: an agent that runs ur exact flow at scale. dont do this until uve perfected the flow manually, automating a broken flow just makes slop faster
what it costs
on dreamina a 720p 8 second clip runs about 296 credits, around $3.50 on the $19 entry plan. across platforms ur looking at roughly $2.50 to $6 per clip at 720p (source
(prices and model access move fast, check before u believe me. my real spend for this run is in the screenshots above)
so ur first believable clip: about 20 minutes on the clock and less than lunch. now u understand why the people running this at scale are printing
ur first 20 minutes
one time, before the clock: build the judge agent (~10 min)
- 0-8: write ur first script, let the panel tear it apart, rewrite til it sounds natural
- 8-14: gpt-images-2 first frame, paste ur script into the video prompt, generate on seedance 2.5
- 14-18: capcut pass. dead space, captions, sound
- 18-20: watch it back once like a viewer. if u wouldnt stop scrolling, the problem is the script, back to the judges
bookmark this and run it once tonight
for more ai video sauce, join my free tg: t.me/beechhq
