Every time I watch a “made with AI” explainer video, I look for the same thing: the moment the details fall apart. Hands melt, on-screen text turns to soup, a building quietly changes shape between two shots. Image and video models are fantastic at vibes and pretty bad at precision.
So when I spotted this post from a LinkedIn creator who produced a full explainer video without a single AI image or video model, I read it twice. Instead of generating pixels, the author had the AI write code that animates a 3D scene. The tool doing the heavy lifting is Astra, running on Codex, and the whole thing worked on the first attempt.
The original poster laid out the process in three steps. I’ll walk through each one and add why it matters, because the rationale is where the real lesson sits.
The process at a glance
- Prepare the production brief
- Code the animation
- Final touch
Looks almost too simple. The catch is that nearly all the effort lives in step one, and the author says so directly.
Step 1: Prepare the production brief
Some backstory first. A week earlier, the creator used Astra to build a 3D model of a 12th-century palace. The source material was a historical painting and a construction manual from the same era. That model ended up with over 100,000 individual components.
Then came the realization that changed the whole approach. Every single one of those components could be animated simply by asking the AI to code the animation. No rendering a new clip for every shot. No hoping the model keeps the palace consistent between frames. The geometry already exists, so the AI only needs to write instructions for how it moves.
With that figured out, the author gathered the raw materials:
- Drafted the script, then refined it with AI
- Generated the voiceover on ElevenLabs
- Generated the music on ElevenLabs as well
Everything went into one production brief for Astra.
Why this step matters: the AI can only be precise about what you’ve told it. The script sets the timing, the voiceover sets the pacing, the music sets the mood, and the 3D model sets the visuals. Hand all of that over at once and the coding step has nothing left to guess about. In the author’s own words, “context and the production brief are the most important steps here.”
Step 2: Code the animation
Astra coded the full animation, and it worked on the first try.
That’s the entire step in the original post. One sentence. I think that says more than a whole paragraph could. When the brief is complete, the AI doesn’t need to iterate, and neither do you.
Why this step matters: coded animation is deterministic. If the camera should pan across the east wing at second 14, right when the voiceover mentions the east wing, the code does exactly that. A video model might give you something gorgeous that ignores the instruction completely. For a story about how a 12th-century palace was built, that difference is the whole ballgame.
Step 3: Final touch
No final touches this time. The creator was happy with the first version and openly admitted to being lazy that weekend. Fair enough.
Why this step matters anyway: keep it in your process. First-try success won’t happen every single time. But notice what “final touch” means in this workflow. It’s a code edit, not regenerating a whole clip and hoping for the best. If the timing is off by half a second, you change a number. That’s a completely different editing experience from AI video, where every retry is a fresh roll of the dice.
The bigger shift the author is pointing at
This industry pro has been using AI-generated video a lot, and still sees something new emerging: AI-coded animation and editing. The point isn’t that video models are bad. They’re incredible for mood pieces, b-roll, and anything where a bit of drift doesn’t hurt. But coded animation has the potential to help us tell stories that require precision, and that’s exactly where video models struggle.
Two different tools for two different jobs. Generate pixels when you want atmosphere. Generate code when you need every element to land exactly where the script says it should.
A few use cases where I think the coded approach wins outright:
- Architecture and history explainers, where the structure must stay accurate from shot to shot
- Product and mechanical walkthroughs, where parts move in a specific order
- Data stories, where numbers on screen have to match the narration
- Software demos, where the interface can’t hallucinate a button that doesn’t exist
FAQ: How many credits did it burn?
The original poster was refreshingly upfront here. This was a 2-hour exercise on Codex that used up 20% of a ChatGPT Pro (5x) weekly limit. Astra ran on high effort most of the time.
Why this matters: budget reality. Two hours and a fifth of a weekly limit for a finished explainer video with voiceover and music is a real cost, but it’s also a real result. Compare that to burning credits on video generations that need ten retries each and still come out inconsistent. High effort mode isn’t cheap, yet it’s what got the first-try success. Skimp there and you probably pay it back in revisions.
How to try this yourself
You don’t need a palace to start. Based on what the author shared, here’s how I’d approach it:
- Pick a subject where precision matters more than mood
- Build or source a 3D model with clearly separated components, so each piece can be addressed in code
- Write your script first, refine it with AI, and lock the timing
- Generate voiceover and music on ElevenLabs so the AI knows the exact duration of every beat
- Bundle script, audio, and model into one production brief and hand it over in a single pass
- Run the coding agent on high effort and let it do the full animation before you touch anything
- Treat revisions as code edits, not regenerations
The pattern I keep coming back to is front-loading. Spend your time on the brief, not on the retries. It’s the same lesson that shows up everywhere in AI work right now, but seeing it applied to a full animated video made it click for me in a new way.
Head over to the original LinkedIn post to watch the video itself and see the palace in motion. It’s worth a look just to notice how steady every detail stays from frame to frame!