This year, I started a new hobby. I started to make AI videos. At first, I was playing with tools, but over time I moved into two separate schemes. The first is making music videos for myself. Second, I played with making adult content. I’m really a hobbyist here. My music channel on YouTube has 30 views and no subscribers. But I find the work engaging and interesting – and who knows – maybe I’m unfound talent now.
One thing I have observed in the last year of AI video: I started with Wan, then moved to LTX, while playing with other tools. The “generation” cycle for Video AI and Image AI seems to be about 6 months long. Since it takes me a month to make a video, that’s a lot of change.
Picture engines ship new versions every month. Video engines change what is possible every few weeks. MiniMax H3 landed in August 2026. It generates picture and sound in a single pass, just like LTX does. But this is years ahead of LTX. Instead of choppy action and ignoring my prompt, MiniMax does what I say. That is exciting. It is also the fourth engine I have had to absorb this year.
A few months ago, I started using Claude Code to Vibe Code my own image queue system. Just a small tool. A way to queue up multiple pictures and rerun them with just a small prompt change. I told Claude what to do, and Claude did the work. It only took an hour. Now, a few months later, my Video Queue Program is a very sophisticated engine that automates a lot of tasks for me. I can queue things locally or on RunPod, automatically generate a dozen related pictures using any picture engine, and quickly compare related tools. The time it takes me to make base pictures, animate them, and organize them has fallen from months to days.
So this is my gift to you – some ideas to help you get started with your own Vibe-Coded solution to AI Image and Video Processing. A bit of sanity for our rapidly changing technology stack.
What is Vibe Coding
The term comes from Andrej Karpathy. He posted it on February 2, 2025. His version was deliberately extreme. You “fully give in to the vibes” and “forget that the code even exists.”
The phrase escaped him almost immediately. Collins named it Word of the Year for 2025. Their definition is broader and far more useful. It is “the use of artificial intelligence prompted by natural language to assist with the writing of computer code.”
I am not shipping software (my “Day Job”). Nobody but me will use my tools. There is no spec, no help desk, and no product returns. This is a personal project between Claude Code and me to create exactly what I need today. Audience of one. And when a new model drops tomorrow – Claude and I will add it.
AI Video Is Excruciatingly Detailed Hands-On Work
My first encounter with ComfyUI was so horrifying I deleted the app in one day. I said -no human can ever comprehend this. Little did I know, it is the cleanest and best tool imaginable for private and secure AI Video production. Web tools make it easy, but web tools cost 10-20 times more per second, and they are neither private nor secure.
For AI videos, I always start with an AI-generated still photo. I used to use web interfaces for these, but I switched to my Vibe Code app to save money. Grok AI has great Image generation – 40 photos at a time on a $30 per month account. Or I can generate them one by one for about $ 0.05 each. I like the control. I only generate about 20 images per month. xAI – here is my $1 for August.
I have a full ComfyUI setup on my local machine. This is great for testing and refining my prompts. I test Videos in really small sizes – 256p or less. Fewer pixels in a small-size video, so generation time is shorter.
For production-size videos, 796p or higher, my local PC is really slow -20-30 minutes for a 15-second clip. So I use a rented RunPod GPU for production work. I rent a server for $2 per hour and run 4-8 hours at a time.
I use some API consolidators, and they come in two classes worth knowing about. One is metered. Atlas Cloud puts hundreds of models behind a single key with no subscription, at roughly four cents an image. The other is a flat fee with a much bigger promise. Venice promises privacy and less moderation but charges five times more. Atlas, Venice and RunPod have dozens of competitors and alternative platforms.
Every one of those routes has its own interface. Every interface wants you present. Pick a prompt. Submit it. Wait. Look at the result. Fix one word. Submit again. My failure rate for video clips is about 90% – that is, it takes me 10 tries to get a usable 15 seconds of video.
My Vibe Code Tool – The Early Versions
I started with pictures. The first tool took one prompt and sent it to many engines. OpenAI, Gemini and xAI directly, then the consolidators as another route. One prompt in, a spread of results out, no clicking.
Then I extended it to ComfyUI. Primarily for i2i workflows to edit pictures. I built my workflows and settings once, properly, in ComfyUI itself.
I added Videos to my Queue system, and this is a game changer. I can queue and test prompts locally – that’s the 9/10 failure rate. I run them through Local ComfyUI. Then I queue the production size until I am ready to launch a RunPod. Then, when the RunPod is connected, I have it process my queue – total nonstop generation – for hours at a time.
This system requires identical workflows and models on my local PC and RunPod. So I Vibe-Coded a few helper batch files to move them.
The same submission form reaches local ComfyUI, pod ComfyUI and hosted engines. The turbo and step controls sit right there, which matters more than it looks – see the last section.
I have an LTX workflow that adds sound to a Wan video. Works nice. More recently, I have a Picture Person and Scene randomizer. It can generate a dozen pictures at a time. It selects pre-set descriptions, backgrounds and clothing. I can have it spit out 20 pictures at a time, and then choose just 3 or 4 to build a scene around.
One prompt box, every picture engine I use, grouped by provider. The randomized prompt script selector is the dropdown under the prompt.
I’m handling SFW and NSFW hobby work from the same PC. Clearing it between sessions destroyed the tests from the night before. So I made named queues. If I had a developer, it would have taken days, and a reason that I don’t want to share with my colleagues. My buddy Claude and I had it done in 20 minutes.
How to Vibe Code Personal Tools on Windows
This is less involved than people expect. Four steps.
Pick an AI coding tool. I use Claude Code. The equivalents from xAI, Microsoft, Google and OpenAI all work for this. They are close enough that the choice is not worth agonizing over.
Designate a work area. One folder, separate from anything you care about. This is where the AI can operate. I don’t allow Claude access to the rest of my PC.
Install Python and PHP locally on Windows 11. Both are free. I know PHP a little; I don’t know Python, and frankly, I have never looked at my VidQueue code at all. You don’t need to learn either, but the AI needs them to run what it writes.
As I work, I decide on new things. I tell Claude. Claude does them. Every day, for five months, I have evolved my Vibe Code app into a true workhorse.
What It Is Actually Like
I have a project management background. I have also managed offshore teams. So the workflow feels familiar. I describe an outcome, somebody else implements it, I review the result.
But the difference is – these are personal tools. I have full rein to change my mind. That matters enormously here, because the engines change under me constantly. I have rewritten the same feature three times in a month because a model shipped.
A human team would have quit me months ago. I know that, because I have led people. Nobody wants to build the same thing a third time for an audience of one.
Claude does not resent my overbearing nature. It does not resent my tiresome attention to detail. That is the real value proposition, and it is not the one you usually hear. AI is not cheaper labor. It is tolerant labor.
Now the honest half.
- Do I swear at it? Yes.
- Has it erased critical code? Yes.
- Did I swear more that time? Yes.
My Claude code space is a designated place in my directory structure. Claude doesn’t access or touch 95% of my files, my history, or my critical records.
I don’t rewrite every vendor directly. So there are some time slips. Microsoft shipped MAI-Image-2.6 three weeks ago. The API consolidator I use is still on 2.5. They’ll update soon, and I would rather wait than drop everything and create a direct billing account with Microsoft AI.
How Efficient is Vibe Coding?
My first “finished” video took four months. The second one took two months. I have about five half-finished projects now. But I’m down to about 20 hours per video, and for me that’s about 10 days’ worth of hobby time.
The money moved too. In my early months of picture and video generation, I dropped about $200 in subscriptions and platform fees. Now I’m down to my Claude account ($20), about $10 for the image APIs, and about $20-30 for direct RunPod hours. My Local ComfyUI is free. A four-hour RunPod session runs about eight dollars for the size I use.
The biggest savings are the ones nobody advertises. Specialist video generation websites charge as much as $0.50 per second of video. Given my 10-1 re-prompting rate, it would cost nearly $500 to get a 6-minute finished result. And then – the question is – who owns it – them or me?
What it is Like Now
I am ninety percent done with a music video. It is a song I wrote, based on a Ming dynasty folk tale about a keeper of records who wants to be a fish and discovers he is a god. It uses Tang dynasty painting styles, with scenes from Lhasa and western China.
Then I compared MiniMax H3 against LTX for production. Their capabilities are so much better that I resolved to reshoot the whole thing. I also changed my mind about which picture engine to use for the stills. So all the pictures go too.
Ninety percent done became ten percent done, in an afternoon, by my own choice.
Under the old workflow, that would have been six weeks in the bin. It wasn’t, and that’s the only lesson in this article that really matters. The prompts survived. The renders are disposable. The prompt library is the durable asset, and the tools exist to run it against whatever engine is best this month.
So a total teardown is a re-run rather than a rebuild. That is also why generating forty unattended candidates is not indulgence. When output is disposable by design, volume is the correct architecture.
One closing note if you are heading down this road with MiniMax H3. Check the license before you publish anything. The territory terms catch out US and EU users in a way that is easy to miss, and I wrote up how I got cleared.