How I Edit Videos with AI Using Codex and ChatCut

Video editing is one of those jobs that can take longer than recording the video itself. There are pauses to remove, explanations to tighten, and repeated takes to sort through before you even start thinking about a title and description.

In my latest Smart Alex tutorial, I try a workflow using Codex and ChatCut Desktop to handle that first editing pass. The result was encouraging: during the demo, I report that roughly 30 minutes of footage had been reduced to around 15. But there’s a catch. Getting AI to assemble an edit feels much better than trying to fix a complicated timeline by hand in the interface I tested.

Watch the workflow

This walkthrough covers connecting the agent, importing footage, refining the edit, and creating the YouTube publishing details. The observations below come from the version and session shown in the video; interfaces and available integrations can change.

1. Start a project in ChatCut Desktop

I begin by creating a new project in ChatCut Desktop. In the video, I also show the cloud portal, but the desktop workflow is the one I use to connect my own AI agent and work with the footage.

Before importing everything, give the project a clear purpose. A tutorial explaining how to generate an anime clip needs a different edit from a short montage. The agent needs to know what the viewer is supposed to learn, not just that you want a shorter video.

Watch the project setup at 1:21.

2. Connect Codex to the editing project

Next, I connect the agent to the ChatCut project through the plugin shown in the demo. The connection indicator turns green and confirms that an AI agent is connected.

That connection is what makes this more useful than asking a chatbot for editing advice. The agent can work with the project and make timeline changes, so I can see the edit taking shape while it works.

I use Codex in this walkthrough. Use the connection instructions available in your installed version rather than assuming that every agent, subscription, or platform supports exactly the same setup.

See the connection step at 1:50.

3. Import the footage and let the transcripts do some work

The footage I import is for a ComfyUI tutorial about generating anime clips. After uploading it to the library, I show the transcript-ready indicators beside the clips.

Those transcripts are a big part of the workflow. In the video, I explain that the agent works with structured project information, text, and some visual material such as screenshots. It isn’t useful to imagine a human editor sitting there watching every second in exactly the way you would.

For a spoken tutorial, the transcript gives the agent a way to understand the explanation, find relevant sections, and assemble them into an order. In this session, it trims gaps and starts building the edit before I move on to more specific feedback.

This is also why I’d check important names, technical terms, and instructions. A bad transcription can give the editor a misleading picture of what was said. And a transcript alone won’t tell you whether an on-screen demonstration has been cut too quickly for someone to follow.

Watch the footage import at 2:10, then the transcript explanation at 2:49.

4. Refine the edit around what the viewer needs

Once the first assembly is there, I give it more direction. My aim is to remove unnecessary explanation and keep the tutorial focused on teaching the viewer how to generate the clips.

During the demo, I describe the footage going from about 30 minutes to 15 minutes. That’s a result from this particular session, not a benchmark or a promise that every recording can be halved without losing something useful.

The more useful question is whether the remaining video still teaches the process properly. A shorter explanation is only better if the viewer can follow it.

Here’s a suggested prompt you can adapt. It’s a starting point, rather than a verbatim copy of the prompt in the video:

Build a clear first edit of this tutorial. Remove long pauses, false starts, and repeated explanations. Keep the steps in a logical order and preserve the context needed to follow the on-screen demonstration. Flag uncertain cuts for review.

Then make the follow-up specific: name the explanation that drags, the repeated section, or the transition that needs attention. Reviewing one concrete change is easier than asking it to “make everything better.”

Watch the follow-up editing stage at 3:37.

5. Use the project context for the publishing details

One of my favourite parts comes after the editing. Because the agent already has the project context, I ask it to generate a YouTube title and description. It also produces chapters.

That saves me from explaining the entire video again in another chat. The same context that helped assemble the tutorial can help describe what the viewer is about to watch.

I’d still check the title against the actual result and test every chapter timestamp against the final export. If you make another round of cuts after generating chapters, those timestamps may need updating too.

See the publishing details at 4:16.

Where the workflow gets awkward

I’m positive about the AI editing, but I’m much less positive about the manual timeline experience in the version I tested. In the video, I explain that moving clips, putting removed sections back, and correcting small mistakes can feel finicky compared with the editors I’m used to.

That matters when a project needs careful B-roll planning, storyboarding, or lots of precise manual adjustments. The agent might give you a promising assembly, but you still need a comfortable way to fix anything it gets wrong.

For talking-head videos and explainers, this experiment gives me a reason to keep using the workflow. For more involved edits, I’d judge the tool by how easily I can revise the result, as well as how impressive the first pass looks.

Watch my results and limitations at 4:41.

Try it on one straightforward video

If you’re experimenting with AI editing, I’d start with a spoken tutorial or a simple talking-head recording. Give the agent a clear objective, let it make an assembly, and watch the result carefully before asking for another pass.

For me, the appeal is having less repetitive editing between recording an idea and sharing it. This session shows a useful step in that direction, with enough rough edges that reviewing the finished video still matters.

If you want to follow along, watch the full Codex and ChatCut tutorial. You can also read my ComfyUI anime clip walkthrough for the kind of footage I’m working with in this demo.