AI video editing: make better YouTube videos.

You have something useful to say. The difficult part is making the video as clear as the idea in your head.

In this episode, I use plain-English instructions to create graphics and custom B-roll around my own recording. Codex coordinates the work, Remotion builds the animation, and fal.ai provides access to video-generation models when I need moving footage.

The useful shift is that I can describe the result, inspect it and ask for a specific change. I still choose the message, judge the output and decide where the visuals belong. The demonstration focuses on supporting visuals; a complete AI edit is a separate project.

Get a small test working.

Open a project folder in Codex. You need a working Node.js installation, Remotion and its Agent Skills. If you already have a project, ask the agent to inspect it before installing anything. Let it consult the current official instructions for the versions available to you.

The project-scoped skills installation used for this episode was npx remotion skills add. Restart the agent after installation so it can discover the skills. See the official Remotion skills guide for current installation guidance.

Your first goal is a five-second test with a title, one animated shape and a clear entrance and exit. Have the agent open it in Remotion Studio and render a review file. A working test is a better starting point than a long list of things the agent says it has installed.

Turn a reference into rules.

Choose a short reference with movement you can study. I used the Presscut demo in Remotion’s showcase to examine how cards enter, settle and guide attention.

Ask the agent to inspect frames before, during and after each important change. What moves? Where does it begin? How long does it take? Which element receives attention next? A link alone may not provide access to the frames; supply a short clip or frame sequence if needed.

The output should be a short, time-coded recipe. Use its movement principles with your own words, colours, layout and footage. Keep the reference’s content and identity out of your finished work.

Build one short passage.

Select five to ten seconds with one clear idea. Give the agent the clip, its transcript and your visual recipe. Include any brand rules it should follow.

Ask for large readable text, graphics that stay clear of your face, and emphasis that lands on the right spoken cue. Preserve the original audio and source recording. Render a separate review version so you can compare what changed.

One passage. One visual idea. One short render.

A short fragment gives you enough detail to judge timing and movement without having to inspect an entire video after every change.

Correct the movement.

In the original three-card treatment, the entrances were compressed into too short an interval. At normal speed, their different directions were difficult to follow. I isolated that entrance and specified the direction, duration and spacing of each card’s journey.

The correction separated the starts by 16 frames and used approximately 25–26 frames for each journey. Those numbers belong to that composition’s frame rate; use timing that fits your own passage.

The more obvious rough-versus-improved comparison in the episode uses a deliberately rough teaching baseline. The original first render is preserved separately. That distinction matters when you show a before-and-after result.

Change one family of elements at a time. Watch the revision at normal speed before deciding it is better. Frame stepping helps diagnose a problem, but the viewer experiences the movement in real time.

Add custom B-roll.

Remotion is useful for words, charts and diagrams. A shot of a place, an action or a physical object can communicate something different. Decide what the footage needs to explain before describing its appearance.

Give the generation model a subject, action, setting, camera movement, lighting, duration and aspect ratio. With a configured fal.ai connection, the agent can prepare and submit that request. Model availability and cost vary, so check the chosen model and budget before generating.

Review the returned clip, then fit it to your spoken passage. Keep the voice continuous unless you deliberately want an audio change. Treat generated footage as an illustration, and check whether it helps the viewer understand the point.

Save what works.

Once you approve the treatment, save the components that produced it: typography, colours, layout, movement rules, prompts and render settings. Keep episode-specific words and footage separate from reusable parts.

Ask the agent to document how to use the system with a new passage. Then test it with different wording. A template that only works with the first sentence is not yet reusable.

You do not need every component in every video. Choose the ones that make the next idea easier to understand. That is where the time saving starts to accumulate.