motion-ad-studio: finished video ad sets

Ad creative Video Desktop

What it does

Point it at a business, say how many videos you want, and get finished video ads back. Not prompts, not a storyboard: rendered 9:16 MP4s with voiceover, music, sound design, AI-generated scenes, motion graphics and burned-in captions, plus the Meta copy for each ad and a showcase page you can send to the client.

The thing that makes it different is how it builds a set. Most ad generators make three versions of the same ad: same structure, same look, same length, different words. This skill makes every ad its own world. Before a single script is written, it picks a different borrowed format for each ad, something your audience already watches that is not an ad, and runs the sales argument inside it.

A real example from the first test set: one ad is a wildlife documentary that roasts the agency model (the peacock is the pitch, a crowded nest is the junior juggling ten accounts), one is a live elimination broadcast where 200 contestants get knocked out stage by stage, and one is napkin math written live in ballpoint at a bar. Three different looks, three different emotions, lengths of 49, 31 and 20 seconds, one offer.

When to reach for it

You need paid video ads for Reels, Stories, TikTok or Shorts, and you want them finished, not as prompts.
You want a test set where each ad is a genuinely different bet, so the algorithm and the audience are not seeing the same ad three times.
The ad is narrated and 15 seconds or longer, up to VSL length for a high-ticket offer. Length follows the story.

Not for: classified-style text ads, stat reveals, ads under 15 seconds, variant batches or re-cuts of an approved ad (use motion-video); realistic AI footage of people as the whole ad (use seedance-video-ad-generator).

Where it runs

Desktop only

It renders video with HyperFrames, headless Chrome and ffmpeg on your machine, so it needs the desktop app.

How to run it

“/um-toolkit:motion-ad-studio 3 video ads for examplebrand.com, selling the free trial”

How it works

The whole run is one conversation with one approval stop. It takes about an hour for three ads.

1. Intake (one message)

It asks how many videos, what you are selling, where the ads send people, any assets you must use or avoid, and whether you want to approve the scripts first or let it run straight through. It tells you the cost up front: roughly 4 to 8 agent runs for three ads, plus a few image, video and voice credits on your own accounts.

2. Truth and assets

It verifies the product and offer on your live site (price, what is included, real ratings, real customer quotes), pulls your real logo, fonts and product photos from the brand library, and reads the voice of your customers: the words they use, what they wanted, what they were afraid of. Everything the ads will claim gets a source.

3. The slate: every ad its own world

It writes two concepts per ad you asked for and keeps the stronger of each pair. Every concept names:

The world it borrows: a nature documentary, a game-show final, a heist briefing, a weather report, a dating app, a bar napkin. Something recognizable in one frame that is not an ad.
The one emotion it delivers (laughter, suspense, relief, outrage, awe, calm), the second it peaks, and the physical reaction it is going for.
The story shape, the hook type, the opening frame, the look, the length, the voice and the music.

Then it runs the diversity gate. Any two ads in the set must differ on at least 7 of these 9 things, and never share a world, a visual style or a generated scene:


World / format	Story spine
Visual system	Dominant emotion
Opening frame	Hook type
Length	Voice and music
Generated scenes

It also bans the generic ad skeleton (hook, product and price, feature list, offer, rating card, end card) as a structure, and spreads lengths across the set: in a set of three, at least one ad runs about 20 seconds or less and at least one runs 45 or more, unless you say otherwise.

Each concept also has to pass five quick tests: the feed test (who it is for and what they do next), the logo-swap test (could a competitor run it with their logo? then it is not about you), the tautology test (the line must not just describe the picture), the mute test (it works with sound off) and the twice test (would someone watch it twice with the offer removed?).

4. Scripts and your approval

For each ad it writes the voiceover, a beat-by-beat storyboard, a psychology map (at least five named persuasion techniques placed at specific seconds, like an information gap at 0:00, precision on a big number, a double-bind close), a facts list with a source for every number, and the Meta copy: primary text, headline, description, button and two alternative hooks.

You see all of it once, before any credits are spent. Ask for changes and only that ad is rewritten.

5. Voice, music and scenes

It voices each ad with a different ElevenLabs voice, trims dead air and keeps the read natural to brisk (it never slows a voice down). It generates a music bed sized to the ad and the sound effects each beat needs. With Creatify it generates the scenes (stills plus 5-second motion clips), and every scene belongs to one ad. Every image and clip is looked at full size before it is used, with zoomed checks for stray text, fake logos and anything that looks off.

6. Build

One builder agent per ad, all running at once (10 to 35 minutes each). Each builds its ad in HyperFrames with your real logo and fonts, burned-in word-by-word captions, and every text element inside the Meta feed-safe area. Every line, X, circle or arrow drawn on screen must sit on the exact thing the voiceover is talking about, and the builder lists each one so it can be checked.

7. QA

Three layers, and all three are required:

Mechanical checks: size, frame rate, loudness (mastered to -14 LUFS), the hook visible at frame 0, the logo on screen by about 3 seconds, every text element in the safe area, no frozen stretches, enough visual change, and no em dashes.
A full-frame review: every scene change and every drawn mark looked at full size.
One cold critic, an agent that has not seen the scripts, who watches the finished ads as a skeptical scroller, a creative director and a craft inspector. It says where it would stop scrolling, where it would leave, and whether the set feels like three different ads or one ad in three colors.

One fix pass follows, then a confirmation.

8. Delivery

You get the mastered MP4s, a web version of each, a launch sheet with all the Meta copy, a run report with the cost, and a client-branded showcase page with each ad in a phone frame and its copy underneath.

How it writes

Money first. The job is ads that make money. The only limit is the truth: no invented numbers, quotes, reviews or results.
No hedging in the story. No disclaimers, terms footers, “results may vary” or volunteered downsides. Terms live on the landing page. If a claim would be false without a qualifier, it gets the shortest one possible.
Direct response every second. A hook with motion already moving at frame 0, every line opening the question the next line answers, something new on screen every couple of seconds, the offer early, the call to action on the emotional peak.

What it needs from you

The offer and the destination URL for each ad.
A brand library. If you do not have one, it creates one from your site the first time (logo, fonts, product photos).
ElevenLabs (ELEVENLABS_API_KEY) for voice, music and sound effects. Optional: it can use a local voice or a voiceover you record.
Creatify (CREATIFY_API_ID, CREATIFY_API_KEY) for AI scenes. Optional: without it, choose no AI imagery or generate the scenes in another tool from the prompts it writes.
ffmpeg and Chrome installed. It checks everything at the start and tells you exactly what is missing. See API keys.

What you get back

Finished MP4s, mastered and ready to upload, plus web versions.
Meta copy for every ad: primary text, headline, description, button, and two alternative hooks to test.
A showcase page with every ad and its copy, ready to send to a client.
The full source: scripts, storyboards and editable HyperFrames projects, so a copy change is an edit and a re-render.

A worked example

You: “/um-toolkit:motion-ad-studio 3 ads for our agency-alternative service, sending people to the application.”

The slate: a wildlife documentary (the agency lifecycle told with a peacock, a crowded nest and a bowerbird hoarding shiny trinkets, then a unicorn steps out of the mist), a live elimination broadcast (200 tiles knocked out through a five-stage vetting until one is left and you approve it) and napkin math (one eleventh of a brain at an agency against up to a third with a dedicated expert, drawn as two pies). Nine of nine rows different between every pair.

After approval: voices, music, eight scenes and three builds in parallel. The critic re-opened all three on fixable craft (a slow opening, an end card that cut away from the peak, a camera push cropping the key line), one fix pass closed them, and the set shipped in about an hour.

Tips and gotchas

Let the worlds be weird. The ads that stop thumbs are the ones that do not look like ads. A nature documentary about your customer’s problem beats a polished product card.
Approve at the script stage, not after the render. A changed line before the build is free; after it, it is a re-render.
Say if you want a long ad. High-ticket offers can carry 60 to 120 seconds when every line earns the next.
Real faces of real people need real photos. It will not invent a face for a named person or put an AI face next to a real review.

brand-library is where its assets come from. For text-card ads, stat reveals and variant batches, use motion-video. For realistic AI footage, use seedance-video-ad-generator.