Big Idea

The argument in one line.

A polished AI-built documentary comes from a deep reference board and several rounds of specific feedback to the model, not from one magic prompt.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A YouTube or short-form creator who wants documentary-style motion graphics over a voiceover without learning After Effects.
Someone who already tried prompting Claude or ChatGPT for an animated video and got a flat, PowerPoint-looking result.
A solo editor who wants to know how to split work between Claude for motion graphics and a second model for cutting and sound design.

SKIP IF…

You want a click-by-click timeline editing tutorial in a traditional editor, since this one is about prompting an agent that writes the animation in code.
You need the exact prompt in the video itself, because the author keeps it in the description rather than showing it in full.

TL;DR

The full version, fast.

The video shows a workflow for turning a voiceover clip into a documentary with continuous camera motion, built in code by Claude Opus 5.5 rather than generated by an AI video model. The author first spends 20 to 60 minutes collecting Pinterest references, then prompts the agent to build everything with code, hands it real historical photos as assets, and feeds back specific complaints about pace, readability and repeated phrases. Sound design is a final pass using a second model, GPT-6 Astra, with the author tweaking effects by hand afterward.

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 →

Chapters

Where the time goes.

00:00 – 00:56

01 · Cold open demo

The finished AI-history documentary plays, followed by the claim that this is one fluid movement.

00:56 – 01:46

02 · References come first

Why references matter more than the prompt and how he collects them on Pinterest across several search tabs.

01:46 – 02:45

03 · Setting up the agent

Selecting Claude Opus 5.5 in the Abacus AI Agent (sponsor) and the move to long continuous-motion clips.

02:45 – 03:41

04 · Prompt iteration

Build everything in code, give it real assets, then correct pacing, readability and repeated phrases.

03:41 – 04:39

05 · Sound design and voiceover

Splitting work between Claude and GPT-6 Astra, X High Mode, and generating a voiceover from a script.

04:40 – 05:43

06 · Second version

Another take on the same audio clip, with the sound effects offered in the comments.

Atomic Insights

Lines worth screenshotting.

References matter more than the prompt: the author spends 20 to 60 minutes collecting images before he writes a single instruction to the model.
Telling the model to build everything with code and no AI video generator is the first rule that produces controllable, continuous motion.
The first two versions looked like a PowerPoint until the model was given real photos of the historical figures to work with.
Asking the agent to list everything it needs, then handing those assets over, beats letting it improvise placeholders.
Specific complaints work better than vague ones: too fast to read, repeated phrases, not enough design time, rough camera movement.
Telling the model to ask questions when something is unclear reduces wrong guesses during iteration.
Once one section nails the style, keeping the same style makes the rest of the video fast to produce.
Using Claude for motion graphics and a second model for cutting footage and sound design splits the work along each model's strengths.
Most people quit Pinterest reference gathering after two or three minutes, which is exactly why their results look generic.
Newer prompting skill now produces 30 to 40 second clips of essentially continuous motion instead of stitched separate generations.

Takeaway

References and feedback beat one perfect prompt

WHAT TO LEARN

Good AI-built video comes from a long reference session and repeated, specific corrections rather than from a single clever prompt.

Spend 20 to 60 minutes gathering references before prompting, because a clear visual direction in your head makes every later instruction sharper.
Tell the model to build with code and avoid AI video generators when you want controllable, continuous camera motion instead of stitched clips.
Ask the agent to list everything it needs and then supply real assets such as actual photos, so it stops improvising weak placeholders.
Give feedback in concrete terms: too fast to read, a phrase repeating where it isn't needed, not enough design time, rough camera movement.
Invite the model to ask clarifying questions when anything is unclear, which cuts down on wrong assumptions across rounds.
Split roles between models: one for motion graphics, another for cutting footage and sound design, and be explicit about who does what.

Glossary

Terms worth knowing.

Reference board: A folder of saved images from Pinterest or similar that sets the visual direction before any generation or building starts.
Continuous motion clip: A single 30 to 40 second animated shot with one fluid camera move, instead of many short generated clips stitched together on a timeline.
Sound design: The final pass of adding sound effects and audio layers to match what happens on screen.

Resources

Things they pointed at.

02:16productAbacus AI Agent

02:16productClaude Opus 5.5

01:30toolPinterest

03:45productGPT-6 Astra

Quotables

Lines you could clip.

03:35

“The secret isn't one magic prompt. It just comes down to knowing how to work with it and not getting discouraged when it gives you a bad result.”

A clean, self-contained thesis about AI prompting.→ IG reel cold open↗ Tweet quote

01:20

“It all comes down to prompting and references, but I would argue that references play a bigger role.”

A short counterintuitive claim that stands alone.→ TikTok hook↗ Tweet quote

The Script

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.

Watch this full video and you will know how to turn your own audio clip into a high -end documentary that people actually want to watch. If you really want to hold attention and print views on social media, you're going to want to learn how to edit like this. And with it being 2026, by learn, I actually just mean steal my prompt.

1950. Alan Turing asks one question. Can machines think?

1956. Dartmouth gives it a name. Artificial intelligence.

1958 frank rosenblatt builds a machine that learns the new york times reports it will one day walk talk and be conscious of its existence then it fails ai goes dark for decades until jeffrey hinton and a few rebels refuse to quit don't program it train it billions of examples trillions of tiny adjustments until patterns become predictions and predictions start to look like thought Now, I don't know about you, but that looked pretty good to me, especially considering it was all one fluid movement.

This is basically unlimited power if you want to grow a YouTube channel. You just have to have enough brain cells to come up with a script that people actually want to listen to. Well, actually, AI can do that too.

Now, if you ever tried making something like this with ChatGPT or Claude, and you're thinking, well, mine didn't look like that, it's because it all comes down to prompting and references. but I would argue that references play a bigger role. Something behind the scenes that you don't see in my videos is I spend a lot of time looking for references.

I'm not creative in any way. If you met me in person, you would think that I'm brain dead. But on the bright side, I know how to steal, so let's go steal something.

So I'm going to load up Pinterest here, and I'm going to search 3D documentary in one tab, cinematic newspaper animations in another tab, and then documentary map animations in another tab. Now I can go between these tabs here and start clicking on different images and saving them to a folder. So each new image that we click here will bring up a new set of similar images so that we can start to build a similar style.

Most people would stop doing this after two or three minutes because they're getting overloaded with dopamine from how awesome these references are. But I personally keep going for 20 or 30 minutes or even an hour so that I'm super happy and I have a very good direction in my head for at least where I think I want my edit to go.

So if you have a Clotta subscription already, you can just use that. I'm going to be using Abacus AI Agent, which is the sponsor of today's video. And it gives you access to all of the models I've ever used for only $10 a month.

There's a link to it in the description. So I'm going to select down here and select the model I want, which is Claude Opus 5 .5. So as I'm getting better at prompting, I can start to get 30 to 40 second long clips that are essentially continuous motion.

Of course, before you would have to generate multiple different clips and the agent would place it on a timeline, which still works. But this is just a different style. And I didn't know it was this capable.

Now, this prompt here is in the description, so you can just copy and paste that. It's optimized for the demo I played in the intro, and it did take me a lot of tries to get it to flow that smoothly. So if you ever want to switch up the style completely, I'll kind of show you how I prompted it.

The first rule I gave it was build everything with code, no AI video generator. The first version here was not great. It kind of looked like a PowerPoint presentation.

The second one here wasn't great either, so I gave it real material to work with. I basically asked it to give me a list of everything it needs, and then I just handed those assets over to it. And this just included things like the actual photos of the historical people.

So this round was a lot better, but it still wasn't great. So I basically just told it everything that was wrong. I told it it was rushing too fast, so I couldn't read any of the words.

There'd be certain phrases like artificial intelligence that would just pop up in multiple areas when it wasn't needed. And so this is when I just told it to slow down, put more time into the design, smooth camera movement, and readability. I also told it to ask me questions if anything was unclear.

You can start to see the secret isn't one magic prompt. It just comes down to knowing how to work with it and not getting discouraged when it gives you a bad result. And then obviously, when you finally nail it like this, it's easy to keep up the fast pace as long as the style relatively stays the same.

So the last step is sound design. And for this, I used a timeline and ChatGPT -6 Astra. Now, Abacus AI Agent has something called X High Mode, which combines Opus 5 .5 and GPT -6 Astra.

It essentially has them work together. And this is perfect because these are the two models I use for my editing workflow.

You just have to be very clear here about which model you want doing what. And obviously in my last video I explained you want Claude for motion graphics and then ChatGPT6Astra for cutting up your footage and sound design. And in this case I already gave it direct access to my folders and sound design and then I just went in myself after and tweaked a few of the sound effects.

Now another thing I want to show you quick is on the home page you can see there's multiple tabs. If we select a voice and then audio generation we can feed it our script and it will generate a voiceover for your video. And of course, you can give this back to the agent, and it will layer all of the motion design on top of it.

If you want the sound effects I used in this video, comment down below, and I'll probably put a link in the description if that comment gets enough likes. Anyways, here's another version I did using the same audio clip. 1950.

Alan Turing asks one question. Can machines think? 1956.

Dartmouth gives it a name. Artificial intelligence. 1958.

Frank Rosenblatt builds a machine that learns. The New York Times reports it will one day walk, talk, and be conscious of its existence. Then it fails.

AI goes dark for decades. Until Geoffrey Hinton and a few rebels refuse to quit. Don't program it.

Train it. Billions of examples. Trillions of tiny adjustments.

Until patterns become predictions. And predictions start to look like thought.

The Hook

The bait, then the rug-pull.

Zane Hoyer opens with a finished documentary-style clip about the history of AI, then spends the rest of the video showing how it was built. The pitch is that the prompt is in the description, but the real lesson is the reference-gathering and feedback loop behind it.

Frameworks