My Next Guest with Daniel Ro
Train your AI editor on the edits you already made
Every CapCut project sitting on your drive is a record of how you cut. Claude can read those files and learn your style from them.
Watch the episode
Every edit you have ever finished is a record. CapCut saves it as a JSON file on your hard drive. The cuts, the transitions, the effects, the presets, the music, the b-roll.
Four years of those projects is what I used to train my AI video editor. Her name is Sofia, because she had to have a name.
The thing that was capping me
Ten plus years in this work. Started as a wedding photographer, moved to fitness and wellness because that is what I am into. Always wanted to be a YouTuber. The problem was never the ideas, it was the editing. There was no time for it.
0:34
Now Sofia edits my podcast, my blogs and my clients' content. Every time you add a new client, you add editing time, photo time, publishing time. That is the math that keeps a content business small.
What that looks like with a real client
One of my clients owns a gym and records his own workout videos. iPhone on a tripod, he shoots, he sends me a text message.
2:06
Then she publishes it to YouTube, cuts the reels and posts them to his Instagram. Another gym gets carousel slides, I just talk to Sofia about what I want to see. With a med spa owner we shot scripts for specific customers plus the treatments, and Sofia edits and publishes the reels.
Take this with you
Take one workflow you repeat every week and write out every step, from the file landing on your phone to the post going live. That list is the first thing you hand the agent.
Which model, and what it actually costs
People always ask which one is better. Honestly, all of them are great. It is preference. I started on ChatGPT, moved to Perplexity, then exported everything into Claude. If I switched again I would have to teach a new model my business, my voice, my clients, my services. So I stay.
Agents burn tokens. Think of tokens like a video game, you exchange them for weapons and maps. It is how the companies measure AI usage and charge you for it. Most people sit on the pro plan around twenty, twenty two dollars. The max plan is around a hundred. Once you run agents, you move up.
How the training actually works
Claude makes the rough cut. I finish it in CapCut and save the project. Then Claude reads my saved project and compares what it did to what I did.
9:54
Every difference becomes a rule.
On one reel the differences were obvious. Stock footage I added. B-roll where she does not look right, because Claude has a hard time finding the best moment in a clip. Hard cuts I wanted smoothed out, since a pause or a flubbed line needs a couple of extra seconds or another b-roll shot over it. And the end card, where I animate a location pin over a map to show where the client's business is. Captions and text overlays are not the hard part, Claude handles those.
Your old projects are the training data
Those files are already on your computer, in a CapCut folder. Point Claude at them and it starts finding patterns. Your transitions, your effects, your adjustments, the music you pick, the b-roll you cut to.
13:35
It also sorts them by type of client. Workout edits in one pile, med spas in another, gym owners, insurance, lawyers. So the next time you ask for that one client, it starts closer.
As of October 2026 it goes both ways. Claude can hand you a CapCut project back, same with Premiere and DaVinci Resolve. Some of the filmmakers and editors I know prefer that, because dragging a cut on a timeline is faster than going back and forth about whether the clip starts at seven seconds or twelve. I do it sometimes. My goal is still for Claude to handle every edit on its own.
Take this with you
Stop deleting old projects. Keep them in one folder, named by client, and treat that folder as the thing your agent learns from.
Six months of back and forth
You need to be patient. It took me about six months to train my video agent, going back and forth. That was months ago, when agents were just starting up. They get better every day, so by the time you are reading this it is further along than it was for me.
Remember giving ChatGPT a one line prompt and expecting magic. Same thing here. Hand Claude a folder of clips with no context and you will not get the reels you see on Instagram.
A basic prompt does not make a professional edit.
Where it still gets things wrong
Color grading is hit or miss. Those are minor adjustments, and the more technical you are, the more specific you can be. Fix the white balance, add an S-curve, lift the highlights, increase the contrast, add sharpness. If you know the language, use it.
Color does not matter much. The cut is everything.
The bottleneck moves
Editing was the ceiling on my business. Fixing it gave me leverage and a new problem.
18:39
Now I need more outreach, because I can take more clients than I could before.
Take this with you
Decide what you will do with the hours before you automate the thing eating them. Otherwise you built capacity you have nothing to put in.
Open CapCut and count your saved projects. Ten is enough to start.
If you do not have ten, make a few edits. Pull from the template section, swap in your own clips, export.
Open Claude Code, point it at your CapCut folder and ask it to read the JSON files and write rules from what it finds.
Save the reels you like on Instagram, download them or send the links, and have Claude study those alongside your own projects.
If you only take three things
Your old edits are training data. Keep the projects, sorted by client.
Correct the rough cut instead of explaining it. The correction is the lesson.
Plan for months, not days, and know what you will do with the time you get back.
Your next read
Tell me what you are working with
If you have a folder of old projects and no idea where to start, send me the situation. I will tell you the next step I would take.
Send me a text