Contents
0%Hey!
I found a creatine ad aimed at men that I loved, and rebuilt it with AI for a creatine brand aimed at women. Almost none of the original words survived, and that is exactly the point.
The finished ad
Their product, for menTheir product, for men
My product, for women
StarpopMy product, for womenStarpopThat is the reference ad playing, with both products beside it. The competitor sells a men's creatine gummy. Mine is a creatine built for women. Same category, opposite buyer, and that gap is the entire reason this workflow exists.
Voiceover plus B-roll is the most flexible AI ad format there is, because the visuals can be any style at all. Learn it once and crochet, claymation or paper animation are the same job.
Why you cannot just swap the product
The reference ad opens with a man being accused of buying cheap protein powder by his wife. It is funny, and it works, because it is built on something men buying creatine actually worry about.
My buyer is a woman. She is not worried about cheap protein powder. She is worried about bloating, puffiness and looking bulky, and an ad that handles the wrong objection quietly does nothing.
Find and replace the brand name in a competitor's script and you get an ad written by someone who has never met your customer. The structure transfers. The words do not.
The pipeline
The AI ad creation pipeline: analyse, research the new avatar, script from the research, voiceover then B-roll
Same start as any clone, then one extra step that does all the heavy lifting.
Step 1: Break down why it works, not what it says
The workflow: AI video breakdown, then research across Reddit, Facebook, TikTok and Instagram, then writing, then generation
Hand over the reference and you get the script, the shots and a shot-by-shot visual description. For this workflow you are reading it for pacing and structure, not for lines to lift.
Mine came back as a classic story arc. A frustration, a discovery, a transformation, a soft close. That shape is the only thing I kept.
Step 2: Let AI research the new buyer (the whole game)
Instead of writing a script, I asked it to go find out what my actual customer says. Not what a marketer imagines she says.
Real Reddit threads from a fitness subreddit, pulled in as customer research
These are real threads from real subreddits, and you can click into every one. It is not a model inventing what women say about creatine, it is women saying it.
AI research output: top objections and top desires for the new buyer, pulled from real threads
Out the other side comes a ranked list of objections and desires. Bloating and puffiness at the top. Glute growth, plateau frustration and supplement overwhelm underneath.
This is how good copywriters have always worked. They read how the customer actually talks until they know her better than she knows herself. The research step does that in two minutes instead of two days.
Step 3: Watch the research turn into hooks
Three AI-written hook options with the visual, the script and why each one works
Three hooks came back, each with its visual, its script and a note on why it works. Every one traces straight back to a line in the research.
"I've been doing hip thrusts for 6 months and my booty looks the exact same. My man's over here popping creatine, making gains."
That is the plateau frustration, written as a first line. It is a better hook than the one in the reference ad, because it is aimed at someone specific.
I picked it over the AI's own recommendation. You should override it too when you have a read on your customer that the data does not capture.
Step 4: Lock the characters, then the storyboard
Hero character drafts for the two recurring characters in the ad
Story ads need the same faces in every shot, so the recurring characters get designed and approved before a single scene gets generated.
Two things that saved me. If the style comes back weak, feed it a reference image and say what level of detail you want, which fixes it faster than rewriting prompts. And keep the visuals honest to the script, so if your character is meant to be frustrated about her results, do not let the image give her the results already.
Step 5: Voiceover first, then cut the silence
Generate the full script as one continuous voiceover before you touch a single visual. The audio is what defines the pacing of the entire ad.
Cutting the dead air out of an AI-generated voiceover in the editor
Then cut the dead air out of it. My 2 minute voiceover came down by almost 30 seconds with no speed-up at all, purely by removing the gaps the voice model left between sentences.
The trimmed voiceover is your timeline. Once it is tight you know exactly how long each shot is, and the B-roll gets generated and cut to fit it instead of the other way around.
Then it is B-roll. Each storyboard shot becomes a clip, you cut on the beat, and the ad assembles itself against the audio.
The reason it holds together
Research, script, characters, storyboard, voiceover and every clip all happen in one chat, sharing the same context. The hooks know the research, the storyboard knows the hooks, and the B-roll knows the storyboard.
That continuity is what makes an ad built this way feel written rather than assembled. Split it across a research tool, a doc, an image generator and a video generator, and each step forgets what the last one decided.
What I would fix
The ad came out well and the script is genuinely decent for the effort. Two honest criticisms.
I introduced the product way too early, which kills the curiosity gap a story ad depends on. And the script ran long, so the pacing is looser than it should be.
Both are script problems, not AI problems. If you are doing this for real, read the script before you generate anything and apply normal copywriting discipline to it.
Next Steps
Take an ad you admire that sells to someone completely unlike your buyer. Break it down for structure, then run the research on your own customer and let the hooks come from there.
If the ad already sells to your buyer, skip all of this and do the straight clone instead. The secrets of cloning competitor ads with AI covers how to tell the difference.
