You know those Vox videos, the ones with slick animations, quick cuts, a narrator who pulls you straight into a story before you can blink? I've been obsessed with that style. And I wanted to see if I could automate the whole thing, from research to script to actual video, using nothing but a single prompt. So I built a skill for it and packaged it in a repository you can clone right now.
The example I'm about to walk through is an explainer for an AI bookkeeping product. By the end of it, the system had done web research, found a compelling angle, wrote a script that sounds exactly like something a human writer would sweat over, and spit out a finished video. It wasn't perfect, and I'll show you exactly where it stumbled, but the core pipeline is solid. And if you stick around, I'll point you toward the resources that turn this into a full automated content machine.
How the Vox Skill Works
The main skill in the repository is called Vox style. It's a full pipeline that doesn't just generate a video on your topic. It asks you a series of questions first to narrow down the kind of video you really want. Are you making an explainer? A product promotion? A story-driven series? It figures that out before it ever touches a script.
Under the hood, it uses what I call "Higgs Field" skills: helper pieces that handle the actual video generation, the explainer formatting, and the visual identity. I bundled all of that into one repo so you don't have to stitch together five different tools. You clone it once and you have the whole assembly line ready.
There's also an .env file where you set your default parameters. Want vertical video? That's the default. Landscape? Flip a flag. You can control video length, subtitles (on or off), character style, language, and a bunch of other knobs. It's all right there, documented in the repo. No magic, just variables you can see and change before you run anything.
Step-by-Step: From Repo to Finished Video
When you clone the repository (or download it as a zip file), you open it in your code editor. I used VS Code, but it doesn't matter. Then you start a Claude Code session, switch to auto mode, and trigger the skill.
All you do is type something like:
Vox style, and here's my topic.
That's the prompt that launches the pipeline. It's not a passive generator. It goes out and does research, writes a script, creates a storyboard, figures out scene breakdowns, and then generates the video piece by piece, dropping everything into an output folder when it's done. You don't touch the research. You don't touch the scripting. You just give it a topic and answer a few follow-up questions.
And if you use the Cloud Co-work version, there's a link in the repo that lets you load the skills straight into that environment. Same pipeline, different interface.
The AI Bookkeeping Example
I wanted to test this with a real product, something that had an actual website and a value proposition. So I pointed it at bookzero.ai, an AI bookkeeping tool I built. My prompt was simple: I told it to research bookzero.ai, tell a story about how AI bookkeeping saves small businesses time, money, and effort, use real stats, and then pivot at the end to promote the product.
The skill recognized the site, pulled the description (AI receipt scanning, automated categorization, QuickBooks export), and then jumped into its question flow. It asked me what kind of video I wanted. I chose "find me a story." Then it narrowed the niche: I said focus on small business bookkeeping, look for real stories of owners drowning in receipts and the hours they could save. If no real stories exist, make up convincing ones that reflect the actual pain points. And I wanted it to explain the problem, the solution, the hours saved, and the value. Straight up.
I told it the story could be evergreen (something that stays relevant for two years or more) and asked for three candidates to choose from. It did a ton of web research around that specific niche, then presented a script.
This is what it came back with:
"It is 11:00 at night. The shop closed 6 hours ago. The rain has not stopped and neither have you. So you sit with a shoe box of receipts typing them into a spreadsheet one line at a time adding it up on a calculator. You have been avoiding this box since March and every single week you avoided it, the box got a little heavier."
That opening hit exactly the tone I wanted. It's visceral. It puts you in a specific, late-night scene with someone who runs a small business and has been putting off the books for months. No generic "small business owners struggle with accounting." It shows you the shoe box.
The script continued with stats that the skill had dug up during research:
- 40% of small business owners spend more than 80 hours a year on bookkeeping.
- 80 hours is two full work weeks of data entry they never bill to a single customer.
- 90% of owners say the admin work hits their daily operations harder than the tax bill itself.
Then it transitioned smoothly into the solution: you take a photo of a receipt. The AI scans it, pulls out the vendor, date, and tax, categorizes it, and exports it directly into QuickBooks, Xero, or a spreadsheet. The first 50 receipts are free. No credit card required. The final line was a clean call-to-action for bookzero.ai.
I approved that script and let it generate the full video.
The Final Result (and Where It Stumbled)
When I played the generated video, the narrative was solid. The visuals, the music, the pacing, they all matched that Vox feel. The AI had broken the video into multiple scenes, each around 15 seconds, and then stitched them together.
But here's the catch: the scene transitions weren't seamless. Because each scene was generated independently, the end of one scene didn't always visually flow into the next. You'd see a hard cut that felt a little jarring.
I know the fix. Instead of generating all the scenes in parallel, you generate them sequentially. You take the last frame of scene 1 and pass it as the starting frame for scene 2, so the visual identity carries through. That way the video feels like one continuous piece instead of a slideshow of loosely connected shots.
I've flagged this as an improvement in the repository. If you want to contribute, you can create a pull request with that sequential generation logic, and I'll approve it. I left it open specifically so the community can take the pipeline further. The bones are there. The fine-tuning is what makes it sing.
What's Really Happening Under the Hood
I think it's worth spelling out what the system actually did during that example run, because it's easy to glance at a finished video and miss how much decision-making happened automatically.
- Web research: It went out, found bookzero.ai, read the site's description, and likely pulled related stats about small business bookkeeping pain points from other sources.
- Question flow: The main skill's first job isn't to generate anything. It's to figure out what you really want, because "make a video about AI bookkeeping" is too vague. It narrowed down the niche (small businesses, not accounting firms or enterprises), the format (story-driven, not a dry explainer), and the freshness (evergreen).
- Scripting: It wrote a narrative with a distinct voice. It didn't just list features. It built a scene with a specific time of night, a shoe box, rain outside, a calculator. That's storytelling.
- Scene breakdown: It split the script into segments, each short enough to generate as one visual piece.
- Video generation: Using the latest C dance 2.0 and Gemini Omni models (the tools I integrated), it rendered each scene with the narration and visuals.
- Assembly: It concatenated them into a single file and dropped it in the output folder.
If I had asked for a different style, a product promotion, an educational explainer, or a series with multiple episodes, it would have adjusted all of those steps accordingly. The pipeline is agnostic to the format, but it needs the initial clarification to pivot correctly.
How You Can Use This Right Now
The repository I'm referencing is public. You can clone it, download it, or load it into Cloud Co-work. Once you have it locally:
- Open the folder in your code editor.
- Start a Claude Code session.
- Switch to auto mode.
- Type: Vox style, then paste your topic. Be specific. If you have a website or a product page you want it to research, include the URL. Tell it what kind of video you want and what emotions it should aim for (urgency, relief, curiosity, whatever).
- Answer the questions it throws back. Usually it'll give you three candidates for a niche, ask about freshness, and maybe ask for a preferred format.
- Approve the script it generates (or ask it to tweak things).
- Let it generate. The output folder will contain your video.
The .env file is where you set your defaults before you start. Change the video orientation, the length, turn subtitles on or off. Those settings get applied automatically so you don't have to re-enter them every time.
And yes, the scene transition issue I mentioned is still present. If you want the highest quality out of the box, you might need to do some manual touch-up or, better yet, implement the sequential frame-passing I described. I wanted to be upfront about that because the video isn't "broadcast ready" on the first pass. It's a starting point that saves you hours of research and scripting and gets you 80% of the way there.
The Bigger Pipeline: Content on Autopilot
This Vox skill is just one piece of something larger I've been building. Inside my school community, Eric Tech AI Builders, I have an entire AI marketing course that takes skills like this and turns them into a pipeline that generates content on a schedule. You pick a topic, the system researches it, writes a script, generates the video, and then automatically distributes it to your platforms on a calendar. No manual uploading, no staying up late to hit publish.
That's the natural next step if you want to go from one-off videos to a consistent content engine. And I'm in that community regularly doing live calls and Q&A sessions. If you run into roadblocks building this out, you can bring your questions there and I'll answer them within 24 hours. The Vox skill is the demonstration; the course is where you learn to systemize it.
Key Takeaways
- I built a Vox style skill that clones the Vox video format. It's a complete pipeline: research, script, storyboard, scene generation, and assembly, all triggered by a single topic prompt.
- The repository includes this main skill plus helper skills and an .env file where you preset video orientation, length, subtitles, and other settings.
- You run it inside Claude Code (or Cloud Co-work), switch to auto mode, and give it your topic. The skill then asks a series of clarifying questions before it ever writes a script.
- In the AI bookkeeping example, the skill researched bookzero.ai, found a story angle about a small business owner drowning in receipts, and wove real statistics (40% of owners spend 80+ hours yearly on bookkeeping) into a narrative that ended with a product pitch.
- The generated video looked like a Vox piece, but scene transitions were abrupt because scenes were generated in parallel. The fix is to generate sequentially, passing the last frame of one scene as the first frame of the next.
- The pipeline can be extended into a fully automated content distribution system. The course inside Eric Tech AI Builders teaches you how to do that, so you can have scheduled, recurring video content without manual work.
There's a moment in the example video where the narrator is sitting with a shoe box of receipts at 11 p.m., and you realize the entire script was written by an AI that dug up that specific emotional hook. That's what I wanted to prove: the storytelling part, the part that feels human and crafted, can be automated if you give the system the right constraints and let it ask the right questions.
The visual polish still needs a human eye, or at least a smarter generation order, but the heavy lifting of ideation, research, and scripting? That's done. And if you've ever stared at a blank page trying to come up with a video topic, you know exactly how much time that saves.
Go grab the repo, run it on your own topic, and when you notice a spot that could be smoother, open a pull request. I mean that. The skill is out in the open because I want it to get better, not sit as a demo. And if you want to go beyond one video and build a machine that never stops making content for you, I'll see you in the community.