From Zero to Promo Video in 30 Minutes with AI

September 18, 2025

Retain creative control when AI writes the code

I created this 74-second promo video draft in about 30 minutes. AI wrote every line of Remotion code, but I was the creative force.

I made a video for Air, a new Python web framework too recent for training data. A perfect test case.

AI cannot feel a musical phrase change, catch fast text, or know what matters to me. This post breaks down the session: providing context, taste, and critical feedback to stay in control.

This was inspired by Greg Ceccarelli creating his company's product
video with
AI

Getting Started: The Blank Canvas

If a tool has a standard generator, use it. I grabbed the hello-world starter from the Remotion docs.

npx create-video@latest --hello-world

Key tip: Even when doing agentic coding, I read docs, learn, and figure out how to give better context and directions.

Step 1: Providing Good Context

Good output requires good input. I provided tools for the Air code and docs.

Prompt Snippet:

You have tools for both Air and Air Docs, so explore those, understand
how they work, and keep it factually correct for a technical audience.

AI could pull core concepts directly from the source of truth.

AI Chat History:

Tool use: mcp__air__fetch_air_documentation

This made it easy to direct AI to specific parts of the library or docs.

Step 2: The Initial Prompt

I voice-transcribed a prompt about Air's key features, my vision, the
feel I wanted, and the target audience.

Key tip: Get as much of your vision, opinions, intuition, feelings,
and goals into context as possible.

Step 3: Adding Taste

Time for the iteration loop: review, critique, command.

Feedback: Readability and Credit

My first reactions:

Everything was too small and cramped
I wanted to credit Air's creators, Daniel and Audrey Roy Greenfeld, more. I found their site and downloaded images.
I wanted the Air logo, so I grabbed the SVG.

I voice-transcribed more instructions, adding context each step.

My Prompt Snippet:

"Everything is so hard to read. Bump it up and make things easy to
see. I've also added the real logo... and images of the two authors
(Danny and Audrey) so you can credit both."

Feedback: Highlighting Key Features

The initial draft missed two key features: strong docs and strong
testing. I prompted the AI to add them to the "Core Features" section.

It misunderstood; air.test doesn't exist.

My Prompt Snippet:

"Remove the built for testing piece. There isn't an air.test module.
This is about maintainers testing the code thoroughly for quality."

Feedback: Better Code Examples

Showing code isn't enough. You have to explain why it's interesting.
I asked for animated highlighting and annotation callouts and a slide on combining Air + FastAPI.

My Prompt Snippet:

"We can do more with the code by highlighting and annotating. We need
to call out why something is interesting. Here's some ideas..."

The code demonstrations became clearer and more complete.

Step 4: Syncing with Audio

A video needs music. I found a royalty-free track on Pixabay and gave it
to the AI. I listened to a few and looked for:

Songs that were close to the same length of my video
Had an impactful phrase change about 30 seconds in
Had a light, airy beginning

AI can't "feel" the music on its own. It needs a human to provide the keyframe.

My Prompt Snippet:

"I want 'Simple Air Application' to come in right at 35 seconds to
match the major phrase change in the song. Figure out how to adjust
the beginning to do that and feel natural."

I could be vague about timing because I picked a song that didn't
require many changes. If I'd needed bigger shifts, I would have edited
the song.

I matched the video's duration to the song's length (1:14) by
holding the final "Call to Action" slide so the music could fade out
naturally.

The Result: A Powerful First Draft

This took about 30 minutes. Is it perfect? No. The pacing could improve, some slides need more
space, and the copy could be better. But as a first draft, it saved me
hours.

AI is powerful, but it's most effective with clear
vision and domain expertise. The final quality depends on human feedback
and context. You need taste.