Get AI Editing Your Videos

Everybody is suddenly getting AI to edit their videos. Codex and Claude Code are cutting the footage, adding graphics and spitting out something ready for YouTube or TikTok.

If you've ever made video content you'll get why people are excited. Filming is honestly the easy bit. It's the hours of cutting, fixing the sound and adding captions that eat the time, or you pay someone several thousand dollars a month to do it for you. That's where most people give up.

They wanted to make YouTube videos, Reels and TikToks. They did not want a second job as a video editor. But that’s what you get when you start making content!

As of a couple of months ago you can hand a big chunk of that to AI. My YouTube channel is 100% Codex edited from my livestreams.

The problem is that every guide on this throws Codex, Claude Code, Remotion and Hyperframes at you at once. You wanted to make a video and now you have a shopping list of software you don't know how to use. Overload.

So I've started a three-part series walking through my actual setup, step by step. Part one (this one) gets the tools working and gives AI its first edit. Part two makes the edits look like your videos. Part three saves the workflow so every new video starts with your preferences already loaded.

Here’s Part 1/3

I've put the full walkthrough in a free step-by-step guide, with screenshots, all eight copy-and-paste prompts and starter project files. Grab it and follow along with your own footage.

Back in July I showed you the system I'd built and gave away a guide to building your own AI video editor. That was the "here's what's possible" version. This series is the beginner's route in - step by step by step!

The brain and the hands

You need Codex (ChatGPT's agent) or Claude Code. I personally use Codex, but Claude Code works just as well, so use whichever you've got. Both are AI agents: they work with files and run software on your computer. You tell it what you want and it goes off and does it with whatever tools you've given it.

You'll need a computer, the desktop app and probably a paid plan at $20 a month minimum. The free plan technically works, but you won't have enough usage to do anything worthwhile. Setup has to happen on the computer; once it's running you can trigger it and give feedback from your phone.

So the agent is the brain. It needs hands. I use FFmpeg (cuts video), Whisper (turns your audio into text), Remotion and Hyperframes (build graphics like text popping up on cue or a side-by-side comparison), all free, plus DaVinci Resolve. Resolve is optional. I use it so I can see what the AI has done and make a quick cut by hand if I need to.

Don’t worry - the agent will go and download, install and configure all of these.

My setup has grown a lot over the few months I've been doing this. Do not copy all of it on day one. Start with one agent, one recording and one way of making a video, and add the rest when you actually need it.

Gone in 60 seconds

The first job is plumbing. Make a project folder (your desktop is fine), open it inside your agent and give it a short brief: check what's on the computer, install Hyperframes using the official quick start (or ask Codex too), set up folders for footage, references, brand and exports, then make a five-second test video with a title and a simple animation.

Why five seconds? Because it tells you straight away whether the AI can make a video file that actually plays, before you throw an hour of footage at it and say "make this into a YouTube video". Start small, find problems and resolve them, then move on to the real projects.

Get this prompt and all the others in the PDF here

Fair warning: it might ask you to approve an install or sign into something. That's generally fine. Ask it to explain what it needs and what the safety implications are.

How the AI makes the cut

Once the test plays, give it a real recording. A couple of minutes of you talking about something you know well, to a webcam or your phone. Make it messy on purpose: bad takes, stumbles, you going "oh crap, that didn't quite work, let me try that again". The first job is to see whether the AI can clean all of that up into a rough cut, the shape of the thing.

Plenty of people on TikTok and Instagram will tell you to jump straight to AI motion graphics and animations. Cool, that stuff goes viral. But we need to lay down the foundation first: an editing workflow that can cut together bad takes and the narrative of the video. Then we add layers. Don't fall into that trap: we're building a system we can use to make content for the business.

The brief tells it what the video is for, what the viewer should understand by the end and what to do with the footage: keep the best complete takes, cut the false starts and anything where you're talking to the editor, keep the bits that explain stuff.

Then it makes a transcript with timings, and the timings are the clever bit. They tie every word to a moment in the footage. The AI edits the written transcript so the words flow, and because it knows exactly when each word was spoken, it can make the cut in the video at that point.

That's why talking to the editor while you record works. In my test I had three or four bad takes, then said "Oh, that was a terrible take. Okay. Take five editor. This is the one." With more swearing I’ll be honest. But that’s between myself and Codex.

The AI found that line in the transcript, found the exact time code and removed the other takes. I also get it to save a list of every cut, so if it chops something I wanted I can point at that decision and it pulls it back in.

The result looks like me just talking. Underneath, it added a small push-in and dropped a graphic onto the one cut to mask it. A cut with nothing changing on screen is a jump cut, and it makes people a little uneasy. Cutting but with something else moving on screen “covers” the cut.

It’s this sort of small, boring but important stuff that is actually valuable here btw - NOT the flashy viral graphics.

What you still check

You still watch it back. Check it kept the right take. Listen to the joins: has it cut you off mid-word, and does the explanation still make sense with bits removed? If it's done something wrong, tell it exactly where. "You cut the sentence that explains the example. Put it back." That's enough.

Tip: While you're going back and forth, ask for small previews of individual cuts. Re-rendering the whole video every time takes ages, burns tokens and leaves you sitting there waiting for the AI. So if there’s a specific area you want to improve get it to fix that part and only that part.

Part two adds the look: graphics, changes in framing and a two-up, like the clip from my cruel-to-Claude video that closes this one.

And part three will be talking about how we make our editing system learn from itself over time.

By the end of that you’ll have a complete production workshop in place for you.

Do this today

Download the step-by-step guide here. Start with the setup and first-edit walkthrough. The later sections cover your visual style and saving the workflow for your next recording.

Make a folder on your desktop, open it in Codex or Claude Code on your computer and paste this:

This one goes into the agent on your computer. You’ll need Claude or ChatGPT Desktop - no phone app for this one!

Next step once you’ve confirmed the basic plumbing works is to record two messy minutes of you chatting, drop the file in the footage folder and ask for a rough cut that keeps your best complete takes.

If you only watch one bit

• 06:09 The setup brief and why we start with a five-second test

• 10:04 How the timings let the AI find "take five" and cut to it

• 11:26 A rough cut that looks like me talking, and what it actually did

The full video also walks through each tool and what it does, so you know what's happening when something goes wrong.

Start small, find the problems, then add the layers.

To the Task,

Kyle

PS. Part 2/3 just went live too