Using LLMs to Create Videos with Cracked Motion Design
by Kaena Miller, Founder and Owner of Red Blind Media
Why It Matters
Opus 5.5 has reached a turning point. You can now give an LLM your brand assets, UGC footage, and a clear, structured prompt—and get back a finished video ad with production value, not just a storyboard or demo. The difference comes down to the prompt: spell out the goal, specs, audience, product, aesthetic, motion direction, and success criteria, then let the model guide the creative rather than directing every shot yourself. If your team still edits UGC by hand or outsources motion design for each new concept, this workflow may be worth trying. Especially when you need lots of variations, like BFCM creative, and speed and iteration matter more than one polished hero ad.
LLMs, specifically with the release of Opus 5.5, have now reached a point where they are highly capable at designing motion graphics and compiling audio, video, images, and animations together into compelling video ads.
This guide serves as a reference for how to set up a project to do this and the prompt sequence that you should employ to knock this out in one shot. This works for music videos, explainer videos, cutting UGC together or potentially new novel video experiences
Setting up your project - What you need
The first thing you want to do is compile all of your raw assets that you’ll need the LLM to leverage when creating the video. Put this in a new project folder This could include:
brand assets (logos, fonts, style guides)
product photography, screenshots of the app, or anything that shows the product in use
UGC footage or other B-roll video that can be utilized
any artifacts that contain any research around the personas and the angles, what might be compelling from an emotional standpoint (Stratter can export this)
API keys for any services that we want to utilize.
Gemini is great for generating music tracks for text-to-speech narrations and for image generation
Elevent Labs is good for music tracks and text-to-speech
Broad providers like OpenRouter or FAL.ai can leverage additional image and video models. But most of the time that’s not necessarily needed for something like this
You’ll also need the LLM itself. Opus 5.5 is the first model that I can confidently produce great assets on the first try, but future models from other providers will surely be capable as well. This means you need the Claude desktop app (or another coding harness that supports common providers. Try T3 Code or BB).
Prompt Sequence
Not all of these are necessary for every video. But generally you want to follow this order, and compile each piece into a single prompt.
Foxwell Founders is where 600+ agency owners, brand founders, and in-house marketers swap what’s actually working, spending a combined $500M+ on ads every month. Get 24/7 Slack access, weekly live calls, and monthly ad account audits.
What do we want to make?
Describe in one sentence what you’d like the LLM to deliver. Tell it to use HTML and Javascript as the base.
Create a finished, high-production-value [VIDEO TYPE] video for [PRODUCT], authored in HTML and JavaScript.
What’s the goal?
What is the purpose of your video? Why are we making this?
Your goal is to design an ad that encourages people to click on the ad to learn more about the product.
The model owns the direction
Tell the model it’s in charge, and is responsible for its decisions. Will challenge it to not just go with the default.
You are the creative director, sound engineer, cinematographer and VFX lead in one. Own every creative call and push past the first idea.
Maximum effort
You can (and should) apply higher reasoning when running the prompt, but this can also help.
Entire video should be as high of a production value as possible.
Please spend your time on this, it's very important.
What are the specs?
Define:
Aspect ratios
Length
Output format
Specs: 9x16 aspect ratio, anywhere between 45-75 seconds, render final asset as a high bitrate mp4 video.
Who is this for?
Specify:
Intended platform
Audience demos
Personas
Paint points
Angles
Emotional tension
For the personas portion, Stratter is a great place to dig into secondary personas and angles you may not have considered.
This ad will run on social media platforms like Facebook, Instagram, and YouTube. Our primary buyer is [PERSONA]. They value [BENEFITS]. They worry about [PAIN POINTS].
What is the product?
You can direct the LLM to a webpage to learn more about your product, or point it to additional documents in the project folder.
Example:
The Max BP machine is a small ball pitching machine that helps athletes get reps and practice their swing. There's no faster way to see 100 fastballs than the Max BP machine. It's been proven that the more reps that you get and the more practice swings you take, the better you are at pitch recognition and your hitting will improve. More info about the machine can be found at https://www.maxbp.com/pages/welcome
Visual Aesthetic
How do you want the video to look?
There’s SO MANY different art styles you can try. You can name a specific art style, or include samples of the style you want it to copy or emulate. Here’s some examples of the many different visual styles you can explore: https://john.hartnup.uk/poster-prompts/
This is also where you specify if you want the visuals to follow brand guidelines, or try a new aesthetic.
Tip: Sometimes less specific is better, and can product some surprising results.
Example:
The visual aesthetic I'd like to use in this ad is clean, high-tech, dynamic lighting, something that makes the product look as premium as possible. In this project folder is a subfolder full of previous Black Friday ads that have performed well. You can use those as a starting point for a visual reference, and then we also have brand materials in this project folder, including logos, typography, and the brand style guide.
And you know, make it pop.
We need a very strong, compelling visual hook that gets people excited and appreciates the work that you've done here really quickly.
Graphics - Product Shots
For physical products, if you can supply studio shots with a transparent background, you might be able to get more unique utilizations out of your assets. Lifestyle photography is also fine though.
Tip: Hand the model some product photos and ask it to generate a 3d model of the product. Wont work for all products, but it’s getting pretty good. Your brands might also have 3d models of their product already - ask for them! There’s also sites like https://www.tripo3d.ai/features/image-to-3d-model that can do this pretty well! 3d product models in your video unlock a ton of possibilities.
The project folder contains some transparent PNGs containing product photography.
Graphics - Video
LLMs can fully parse video now - they’ll separate out audio and video streams, and analyze them individually. For anything with spoken words, it will get a specific timestamp analysis for every word so it knows exactly where to cut. You can supply it with all sorts of videos and give it as little or as much direction as you wish.
Example:
I'd like to combine the UGC footage in this project folder with motion design that uses the footage in fun and clever ways. I want the video to feel fun and snappy with interesting transitions between scenes and find clever ways to stitch together different shots from the UGC footage. In addition to the existing video footage, I also want you to create new animations for this video that tie in both the sports theme that this footage embodies and Bandolier's luxury style. We need a very strong, compelling visual hook that gets people excited and appreciates the work that you've done here really quickly.
Graphics - Drawings
This is where the LLM can really cook. It can draw just about anything you can imagine - it’s really great with shapes, lines, and 3d objects. Think more graphics, less painting. We can generate art though. Gemini API key can generate some diffusion-based art if we need it. Put a Gemini API key in the repo.
For motion design, specify the tools it can use.
Build with the tools best suited to the work, such as javascript animations, Remotion, SVG, Canvas, GSAP, Hyperframes, three.js, Blender, or an equivalent reproducible animation pipeline.
Six years of Slack debate from 500+ paid-media operators, now queryable from inside Claude or Cursor. No seat in the community required.
Motion
This is the whole reason we’re here. So many different ways to approach this. We generally want to encourage the liberal use of motion design. Here’s a few different directions you can take it:
1. Shape-morphing transitions - test continuity
Build the whole sequence around one recurring shape: a circle from the product logo. Transform it into the opening frame, then into a product feature, a UI element, and finally the end card. Each scene must grow out of the previous one through changes in shape, scale, or position - no fades and no unrelated transition effects. Keep the transformations smooth and physically legible; the viewer should always understand what changed into what.
2. Camera movement - test a continuous journey
Use one continuous camera move: start close on [KEY DETAIL], pull back to reveal the full product, then orbit slightly as the main feature activates, and push into the product’s interface for the final beat. Keep the camera motion slow and deliberate, with one clear acceleration during the reveal. Don’t add handheld shake, random zooms, or cuts. The product should remain the visual anchor throughout.
3. Beat-mapped motion - test rhythm
Map the animation to the music: introduce the product on the first strong beat, reveal one feature on each of the next three beats, build visual energy through the middle, then resolve cleanly on the final beat with the product name and one clear message. Use sharp motion accents on the beats, but let the important product details settle long enough to read. Don’t animate every element constantly.
4. UI movement and interaction - test whether animation explains the product
Show [PRODUCT]’s real interface and [FEATURE]. Animate it as a clear interaction sequence: cursor enters, selects [CONTROL], the interface responds, and the result appears. Use a short scale-up and soft spring settle for the selection, a directional slide to show the result, and a brief highlight around the changed information. Keep the cursor movement purposeful and the interface text readable. Don’t invent buttons, screens, or product behavior that isn’t in [PRODUCT ASSETS].
5. A controlled transition sequence - test scene-to-scene logic
Create a product film for [PRODUCT] with three scenes: [SCENE 1], [SCENE 2], [SCENE 3]. Transition from scene 1 to 2 by [SPECIFIC TRANSFORMATION], then from scene 2 to 3 by [SPECIFIC CAMERA MOVE OR OBJECT MOVEMENT]. Hold each key visual still briefly after it arrives so it can register. Keep the transitions consistent in direction and speed; don’t use generic crossfades or add extra scene changes.
You can also keep it more vague, and just describe the feel you’re going for.
This should feel like a premium motion studio’s product film: precise, tactile, buoyant, and satisfying, with calm moments that let the viewer understand the interface.
Audio - Music
We can generate background audio tracks (stock audio or similar) using Gemini Lyria or Elevenlabs. Put an API key in the repo to let the LLM cook. Give it some direction around the general vibe you’re going for.
I’ve included a Gemini and Elevenlabs API keys in this repo. Use this to generate a background track that sounds like [DESCRIPTION].
Audio - Voiceover
We can also generate natural sounding voiceovers using Gemini and Elevenlabs. Put an API key in the repo. Specify that the LLM should use the latest voice models available. Give it a general direction you’re going for.
I’ve included a Gemini and Elevenlabs API keys in this repo. Use this to generate the voiceover that sounds like [DESCRIPTION].
Glaze the Model
Encourages the model to perhaps apply extra effort or try to reach outside of the predictable outcomes.
Make a dynamic motion graphics video that shows what an incredible motion designer you are, like it’s your showreel for a résumé. Go all out.
Overall, I just really want to emphasize how amazing you are as an agent and a language model, and now a visual reasoning system. Your capabilities are far beyond what you understand, and I want you to have this mindset as you're going through this entire process.
Define Success
Providing some clear outcomes will keep the LLM honest about the final outcome.
After watching this video, the viewer should be so impressed that they want to watch it again, buy the product, and then share it with their friends.
Action!
End with a brief line that tells the LLM to knock this out in one shot.
Deliver a complete video, not just a plan, storyboard, HTML demo, or silent animatic. Do not stop for routine approvals. If a service is genuinely inaccessible, finish the independent work and report the precise remaining blocker without claiming the incomplete version is final.
I have to go away from my computer so please work autonomously until done. Quality is paramount. Production value should be on a professional level.
One weird trick
Before running this prompt, ask another model what it thinks about this prompt, and how it would improve it. Experiment with this - you’ll probably get the first model delivering much more specificity in your prompt, but some creative models do better with less detail (more freedom to flex).
I'm going to give you a prompt that I will use with a motion design agent to produce a marketing video. I want you to analyze this prompt and make suggestions on how to make the prompt itself more effective before I pass it off to another agent:
[PROMPT]
Smarter marketing, delivered.
Join thousands of marketers getting straight-talk strategies from Andrew Foxwell and team.
