The argument in one line.
A repeatable loop of research, script, image, and last-frame-chained video generation lets one AI coding agent direct three separate AI models into a single coherent short film for about a dollar.
Read if. Skip if.
READ IF YOU ARE…
- A creator or marketer who wants a repeatable AI pipeline for producing short explainer or documentary-style videos without hand-animating anything.
- Someone already comfortable directing an AI coding agent (Claude Code, Cursor, etc.) who wants to chain multiple image and video generation models into one workflow.
- A solo creator deciding between building this pipeline themselves or paying for a template plus prompts to skip the setup.
SKIP IF…
- You want a video editing tutorial in a traditional NLE — this is entirely prompt-and-agent driven, no manual keyframing or editing software shown.
- You're looking for unbiased tool comparisons — the walkthrough doubles as a pitch for the creator's own paid mastermind course.
The full version, fast.
A creator walks through the exact AI pipeline he used to build a 30-second Vox-style animated short about HMS Victory. He first fed roughly ten reference videos into NotebookLM to distill Vox's editing conventions into a style brief, then wrote a short chunked script inside Claude Code. Each scene follows the same loop: generate a still image with GPT Image 2 (leaving blank space for text), animate it into a clip with Google Omni, extract the clip's last frame with FFmpeg, and feed that frame back in as the starting point for the next scene. Once proven manually, the loop runs unattended end to end. The finished three-scene video cost about a dollar in model fees.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:24
01 · The One-Shot Result
Cold open teases the finished HMS Victory clip and previews that everything was built in Claude Code, then states the day's premise: showing the whole pipeline for GPT Image 2 + Google Omni.
00:24 – 01:39
02 · Researching the Vox Style
Uses Apify to pull about ten YouTube videos about Vox's editing style, loads them into a NotebookLM notebook, and asks it for a written breakdown of the style plus a style prompt and animation prompt.
01:39 – 02:21
03 · The Free Pipeline Files
Shows the local project folder holding the reusable prompts, docs, and Claude skills that make up the pipeline, offered as a free download.
02:21 – 03:39
04 · Writing the Script
Inside Claude, requests a 30-second HMS Victory script split into three ~10-second sections, then asks for a punchier rewrite and a stronger opening line.
03:39 – 04:36
05 · Image & Video Models
Explains the model choices — OpenAI's GPT Image 2 for stills and Google's Omni for animating them — routed through the Kie AI aggregator instead of separate provider API keys.
04:36 – 05:31
06 · Creating the First Image
Claude generates the first still frame (grayscale halftone HMS Victory engraving on torn paper) with instructions to leave blank negative space and no on-image text for later overlays.
05:31 – 08:00
07 · Prompting the First Video
Claude writes a detailed motion/animation prompt (camera moves, label pop-ins, paper-texture rules) before generating video from the approved still, with a mastermind-course pitch woven into the wait time.
08:00 – 08:11
08 · First Scene Result
The first rendered clip plays back showing the ship, title card, and narration in sync.
08:11 – 09:47
09 · Chaining the Scenes
Explains and demonstrates using FFmpeg to grab the last frame of the finished clip, feeding it back to Claude so it can plan and generate the next scene from where the last one left off.
09:47 – 10:40
10 · Finishing the Video
Tells Claude to complete the remaining scenes and stitch all three clips together unattended, then returns to review the assembled cut.
10:40 – 11:16
11 · The Final Result
Plays the finished 30-second HMS Victory video in full, covering the ship's history, gun count, rigging, crew size, and Trafalgar.
11:16 – 12:35
12 · Process Recap
Recaps the full loop — style research, script, image, video, last-frame chaining — and notes the whole 30-second clip cost about a dollar in model fees.
12:35 – 12:57
13 · Outro
Points back to the free pipeline download and the Applied AI Mastermind, then signs off.
Lines worth screenshotting.
- A thirty-second Vox-style animated video was produced end to end for about a dollar in AI model costs.
- Feeding the last frame of one generated video back into the image model as the next scene's starting frame keeps a multi-clip video visually continuous.
- NotebookLM can synthesize roughly ten reference YouTube videos into one written style brief, replacing manual screenshot analysis.
- It is cheaper and faster to iterate on a still image than on a generated video clip, so revisions should happen before the video-generation step, not after.
- Routing image and video generation through a single third-party API aggregator avoids juggling separate API keys and credit balances for each model provider.
- Claude Code is capable of running an entire multi-scene generation pipeline unattended, but it is noticeably slower than tools like Codex or Cursor's Grok models for the same task.
- Explicitly instructing the image model to leave blank negative space and omit on-image text keeps room for text overlays to be added in a later step.
- A full automation run — generate image, generate video, extract last frame, plan next scene, generate next video, repeat, and stitch — can complete in a few minutes per scene once the workflow is proven.
- Breaking a script into short chunks of ten seconds or less before generating any visuals keeps each AI-generated clip narrowly scoped and easier to correct.
How to chain still-image and video AI models into one continuous short.
AI VIDEO PIPELINE
A repeatable loop — script, generate an image, animate it, extract its last frame, and use that frame to start the next scene — turns three separate AI models into one continuous short film for around a dollar.
02Researching the Vox Style
- Before writing any prompts, gather about ten reference videos in the target style and have NotebookLM synthesize them into one written style breakdown rather than eyeballing screenshots.
- Ask the research step to produce two separate outputs: a style prompt (how things should look) and an animation prompt (how things should move) — image and motion are different problems.
04Writing the Script
- Cap each script section at ten seconds or less so a single AI-generated clip only has to carry one beat of narration.
- Review and tighten the script's wording before generating any visuals — ask for a punchier opening line, since a script is cheap to iterate on and expensive to fix once video exists.
05Image & Video Models
- Split the job across two specialized models rather than one: a dedicated image model for stills, a dedicated video model to animate them.
- Route calls to multiple providers' models through one aggregator API instead of separate accounts, so you don't manage several sets of API keys and credit balances.
06Creating the First Image
- Explicitly tell the image model to leave blank negative space and omit on-image text, so titles and captions can be composited in cleanly afterward.
- Approve the still image before spending on video generation — correcting a bad image is cheaper and faster than correcting a bad video clip.
07Prompting the First Video
- Write the motion prompt as its own explicit step covering camera movement, timing, and what must NOT change (background, colors, framing), not just "animate this image."
- Generate and review one scene at a time before letting the agent continue, so mistakes get caught before they propagate into later scenes.
09Chaining the Scenes
- Use FFmpeg to extract the last frame of a finished clip and feed it back to the model as the starting image for the next scene, so consecutive clips flow without a visual jump.
- Have the AI analyze that extracted last frame before planning the next scene's prompt — planning blind, without seeing what actually rendered, produces prompts that don't match reality.
10Finishing the Video
- Once the per-scene loop is proven manually, it can be handed to an agent to run unattended: generate, extract frame, plan next scene, generate, repeat, then stitch every clip together with FFmpeg.
12Process Recap
- The full loop is research the style, write a chunked script, generate an image, animate it, extract its last frame, repeat, then stitch — a small number of repeatable steps regardless of how many scenes you need.
- A 30-second, three-scene video assembled this way cost about a dollar in AI model fees, which is the real headline result to evaluate the workflow against.
Terms worth knowing.
- Vox style
- An editorial documentary motion-graphics style built from paper-cutout illustrations, hand-drawn labels, and kinetic typography, associated with Vox's YouTube explainer videos.
- GPT Image 2
- OpenAI's image-generation model, used in this workflow to produce the starting still frame for each scene.
- Google Omni
- The Google video-generation model used to animate a static starting image into a short motion clip.
- Kie AI
- A third-party API aggregator that exposes multiple providers' image and video models through one API key instead of separate accounts per provider.
- Last-frame chaining
- Extracting the final frame of one generated video clip and feeding it back into the image model as the starting frame for the next scene, so consecutive scenes flow without a visual jump.
- NotebookLM
- Google's AI research tool, used here to synthesize several reference videos about a target editing style into one written style brief.
Things they pointed at.
00:24toolNotebookLM
00:24toolApify (YouTube scraper actor)
03:39toolGPT Image 2 (OpenAI)
03:39toolGoogle Omni
03:39toolKie AI
Lines you could clip.
00:00
“I just one shotted this with AI.”
05:18
“It is far cheaper and quicker to request changes to images like this than it is to request changes to videos.”
11:20
“It's pretty staggering what we can now create for, you know, about a dollar.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphorstory
00:00I just one shotted this with AI.
00:02This ship has been at war for two hundred and sixty years, and she's still in commission. Meet HMS Victory.
00:10104 guns, 27 miles of rigging, 820
00:15men crammed So today, I'm gonna show you how I created this Vox style video using QPT Image two and Google Omni, and I'm gonna show you how you can create it too. So let's get into it. Okay.
00:25So the first thing we need to do in order to create a Vox style video is to define what a Vox style video is. We need to give Claude a load of information about what this actually means. Now there are different ways that you could do this.
00:38The way that I did it was that I used Apify to go and find a bunch of YouTube videos that talk about Vox's editing style and video production. I said perhaps 10 of them, then put them in a notebook, LM notebook, and then from that give me a comprehensive breakdown of what the editing style is.
00:54I found this is one of the best ways to do research. You could alternatively just go and take screenshots and ask it to analyze it, but I used this method instead. So it created me a notebook that looked a little bit like You can see it it pulled some sources, and it created a notebook.
01:09And then I asked it to create me a comprehensive style prompt and an animation prompt, because we're gonna be doing two things here. We're gonna be creating images, and we're gonna be prompting animations. We're gonna be doing both of those things.
01:19And it went ahead, and it made that style prompt and that animation prompt. Now, these prompts are part of the video production pipeline that I've created as well as a bunch of skills, which basically fix problems that I came up against whilst building it. If you want access to this entire pipeline, so like the skills, the prompts, everything like that, then there's a link down below where you can download that absolutely for free.
01:39But here in my files, can see I've got a folder called Vox video and if I open it up, I've got a load of different information in there. And this is the main one here, the Vox sequential scene workflow.
01:48And if I open this up, you can see it basically gives the instructions of this entire video which I'm showing you. Now, what this allows you to do is create the video in one shot. You simply say, create a video by x and it will create it.
01:59However, what I'm gonna show you in this video is step by step. I'm gonna show you the different steps so that if you wanna create this yourself, then you can.
02:05Or if you wanna be able to, you know, reverse engineer what I created, so maybe you can, like, take parts of what I've done and apply it to your own project, then you can do that. And also, it provides a better result cause you can check it at every step along the way rather than just saying, hey, create this and letting it do everything, but that is an option.
02:21So let's hop into Claude and let's actually get started creating it. I'm gonna use Fable five. I'm within this folder.
02:27That's the same folder as linked down below, and I want to create a video about the HMS victory. I'm a proud Brit, so let's create it about that fine bit of British history. And I'm gonna say, I wanna create a Vox style video all about the HMS victory.
02:42Please, can you create me a thirty second script broken up into three chunks of ten seconds or less. Just give me the script for now. Okay.
02:52So as I said, it can do everything, but I wanna check-in this pass through. So it's gonna tell me what the script is gonna be, and then can edit it because the script is what is gonna guide everything.
03:02Okay. So here we are. Here are the three sections.
03:04They seem okay. They seem a little bit long, and I need a little bit more punch.
03:09So I'm just gonna say, can we cut down the size of each one of these? You can rewrite it completely if you need to.
03:17And also the first section, can we start it with like something a little bit more punchy, please? Always say pleased AI. You never know when the AI overlords are gonna come, and they're gonna remember who said please and who didn't.
03:28So let's see what we get after that. Okay.
03:32This ship has been at war for two hundred and sixty years and still in commission. Meet HMS Victory. Yeah.
03:37Okay. That is much much better. Good.
03:39Okay. So now the next step is we were gonna create the first image. Now let me talk about the tools that we're using here and the models.
03:46We created the first image so that we can feed that image into the video model, which will produce the video. The image model that we're using is OpenAI's image GPT image two, which at the moment is the best image model out there.
04:00It's not that expensive either, and then we're using Google's Omni as the video model. Now to create these images and videos, I am connected to Kai AI.
04:08I'm not connecting directly to OpenAI's API and to Google's API. Instead, I'm just doing it through one API. It just makes it easier as you don't have to deal with a load of API keys and a load of, you know, making sure everything's got credits.
04:20Kai is good, equally good, like, and other things like that. In order to set this up, all that I did was to ask Claude to look over Kai's documentation and set up the API routes for Google Omni and for GPT two. Okay.
04:32So now we're back in Claude, and I was pleading to say, okay. That looks good. Please can you create me just the image of the first scene?
04:42Okay. And we'll send that off. Now if you are running this workflow, let me give you one little tip.
04:46Fable five is good and the Claude models are good. They are incredibly slow.
04:50So if you actually want to produce something quickly and like time is important, then don't use Claude. Use something like Codex or use something like Cursor like the Grok models.
04:59However, we are gonna keep using Claude. It is pretty good. Okay.
05:02So I to reset my EMV variable, but I've got that image. Let's have a look at it. So here it is.
05:07That is pretty good. Now, part of what the prompt says is in the first image of the first video, don't put any overlays of any text or anything because we need to add that in later. Good.
05:18So now what I'm gonna say, now that I'm happy with that, if you are not happy with that, then you can obviously request any changes. It is far cheaper and quicker to request changes to images like this than it is to request changes to videos. So if you've got changes to make, make it here.
05:32Cool. So I'm gonna say, that looks good.
05:34Now can you tell me the prompt that you're gonna use to create the first video from this image? So again, you can just run this through all the way without being involved, but because this is kinda like a little bit of a walkthrough and showing you how it works, I'm doing it step by step.
05:46Okay. So here we go. It's giving us the prompt for first this video and we can look it over just to make sure that we like what it's gonna show us.
05:53So let's see. As the narration says, at war for two hundred sixty years, a large stamp appears for two hundred and sixty years.
06:00Yep. Still in commission. Rack focused gently on the ship.
06:04The ship sharpens and lifts slightly in contrast with the sea. Okay. Yeah.
06:08That looks fine. I'm gonna say, great. Go ahead and create this first video clip for me and then show it to me.
06:15Don't create more videos after this. Okay. So again, we're just doing this first video, and then we're go on from that and create the the second videos.
06:22Essentially, I haven't said before, this thirty second long clip is gonna be three separate videos which we're gonna stitch together. Now what that generates, if you are interested in creating content with AI for your business, for your marketing, whatever it may be, then inside the applied AI mastermind, I've got some assets which can be very useful.
06:38There's a full course all about AI content fundamentals covering everything from image creation to, um, hyper frames, and actually I'll give you my own software which I have built bespokely to edit my own videos, but essentially from creating videos, creating images. There's also a full course about hyperframes and how to use AI to edit videos for you, and this is just a small selection of what is included within the applied AI mastermind.
07:00You can see that what is inside is a selection of full length courses on things from Claude and AI, vibe coding, Hermes, AI marketing to how to generate leads, NA 10, which is still a very important tool. But most importantly, there is the AI applied to business module, which is how to see the opportunities within your business or a client's business where you can actually use AI and apply it to a business so that the business can move forward, whether that's in terms of sales and increasing top line or cutting costs and essentially improving the bottom line.
07:31We are seeing some great results inside the mastermind for some of our members, and there is a fourteen day money back guarantee. So you can join, and if you don't like what you see, if it's not for you, absolutely fine. Just request your money back, and there is zero risk.
07:45And also, we're about to increase the prices. So if you wanna join and potentially save a load of money, lock in your price, then you can join with the link down below in the description, and I'd love to see you inside. Anyway, let's get back to it.
07:57Okay. So this first scene is generated. Let's have a little look at it.
08:00This ship has been at war for two hundred and sixty years, and she's still in commission. Meet HMS Victory. Okay.
08:09There we go. So, um, let's close that. And now what we're gonna do within this process is that it's gonna read, in order to create the next video.
08:16So we've we've got three videos stitched together. We will do. In order to create the first frame for the next video, we're gonna take this last frame here.
08:24We're gonna have AI look at it because what AI is gonna do is it's gonna look at this, and then it's gonna plan on the next things that should happen from this frame here. Because I've tried it before where you just ask it to create all the prompts without seeing what's happening, and because it can't see what's happening, it creates prompts which don't really make a lot of sense.
08:42So what in fact we're gonna do is feed it this last image here. It's going to use FFmpeg to strip it. It's gonna analyze it, um, and then it's then gonna take this and be able to plan on the next steps.
08:53So I'm gonna say, great.
08:56Use FFmpeg to do your process and look at the last frame of that video that we created, and then using that, please plan out the next video for us. Don't actually create it, just tell me the prompts. Now, whilst we're waiting for this, I will mention that in the description below, I've got a list of free AI resources.
09:11If you're looking to help AI improve your business or clients businesses, then those might just help you, and they're free. So go ahead and check them out.
09:18Great. Okay. So here is the second prompt.
09:21So it's got some text in it. It's going to keep this text exactly the same, but it's gonna render on some more text as well. No hard cuts.
09:30Narration is gonna say a 104 guns while rough pencil circles are drawn around two or three gun ports on the ship's hull. Okay. Cool.
09:38Yeah. I'm not gonna read through all that. I'm just going to assume the Fable has done a good job, and actually I'm gonna say that looks good.
09:45I want you to go ahead and just finish the entire process for me. So create this second video, and then take the last frame of that, and create the third video from that, and then go ahead and stitch all three of those videos together.
09:58I want you to just give me the finished video. So here we're gonna use that automation process which is built into this select skill or project that I've built, where we can just say, okay. Just go ahead and do it.
10:09And I've given it some detail here about what it should do, but that's only just to make sure it stays on track. I have described before what it needs to do, so it should it should know if I just say complete it, that it just needs to complete it and do in order to get it working.
10:23So it's gonna take probably like five minutes for it to create that image to create that video, then strip the image, analyze the image, create the prompt, create the other video, stitch them all together.
10:33So let's come back when that's done, and let's see how it fares. Okay. So it's come back.
10:37All of those three scenes stitched together into one video. Let's put them all together with FFmpeg. Let's have a little look and see what it looks like.
10:44This ship has been at war for two hundred and sixty years, and she's still in commission. Meet HMS Victory, 104 guns.
10:5427 miles of rigging, 820 men . Trafalgar, eighteen o five.
11:02Nelson's flagship shattered two fleets, and he died on her deck victorious. There you go. Pretty good.
11:08There are definitely some things we could change and we could improve, but, you know, it's pretty staggering what we can now create for, you know, about a dollar. Just a little over a dollar is what that costed.
11:19So that's the entire process. The entire process is creating a script using the Vox style guide, which we created via pulling it from a load of YouTube videos. First, creating the first image and then turning that first image into a video, and then we take the last frame of that video and use it as the first frame for the next video, and we do that for however many bits of video you wanna create.
11:39I've just done three here, you can do as many as you like. Then once all of those video files are created, in fact, let me show you all of those video files. So here we can see if we go in here, we've got scene one, scene two, and scene three, and we've got that raw files as well.
11:54And simply they just stitched together using FFmpeg, which just is how you manipulate video files or image files as well, and stitched it together into one video. Now we've gone through this step by step, but you can simply say create me a video about the Napoleonic war or whatever it is you wanna create a video walk away for ten minutes and come back and it will create you a video thirty seconds long or three minutes long or thirty minutes long, however long you want.
12:19So there we go. I hope you have found this useful. If you do want kinda like the file that I used and the skill that I used, the project that I used to create this with all the prompts and all the star guides and everything like that, then I'll leave that down below in the description.
12:30All you need to do is add your own API key for Kai or Fowl or whichever one you wanna use. It's about applying AI to your business, creating content with AI, or generally improving your business with AI with prism strategies, then do check out the applied AI mastermind which will be linked down below in the description. Found this video interesting?
12:46Then please do give it a like. Subscribe if you wanna see more videos like this, and I would highly recommend going and watching this video next. I think you'll find it very interesting.
12:56See you in the next video.
The Hook
The bait, then the rug-pull.
The video opens mid-boast — "I just one shotted this with AI" — before immediately rewinding to show the actual multi-step pipeline behind that single Vox-style clip about HMS Victory.
Named ideas worth stealing.
01:39model
Vox Sequential Scene Workflow
- Research style with NotebookLM
- Write chunked script
- Generate scene image
- Generate scene video from image
- Extract last frame via FFmpeg
- Analyze frame + plan next scene
- Repeat per scene
- Stitch all clips with FFmpeg
The reusable loop the creator packaged as a downloadable pipeline of prompts and Claude Code skills for building any Vox-style AI video.
Steal forany AI-generated explainer, documentary-style, or motion-graphics short that needs multiple continuous scenes
How they asked for the click.
VERBAL ASK
06:27product
“If you're interested in creating content with AI for your business, inside the Applied AI Mastermind I've got some assets... a full course all about AI content fundamentals... fourteen day money back guarantee... and we're about to increase the prices.”
Woven into the walkthrough during a natural wait (video rendering in the background), framed as bonus value tied directly to the tools just demonstrated; closes with urgency (price increase) and risk reversal (money-back guarantee).
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
Visual structure at a glance.
