This week I posted a video on LinkedIn about an app I’ve been building. This isn’t a post about the app. This is a post about the subtitles I put in the video.
The subtitles had to do the talking
// the old style — Inter 600 in a chip, centred, every line the same
I’m not ready to narrate my own videos — yet. So the words on screen were the only voice the story had. And watching my 1-minute video pitch I realized quickly it looked like CRAP.
Why can’t mine look like that?
So it got me thinking that subtitles have more dimensions to them than a simple box with text. For example, Avatar movies mimic their world. Influencers manipulate their texts to give you constant dopamine hits. John Wick has the best subtitles in the business (a hill I will die on). All the while mine was the black box with white text you’d get when you clicked CC on Netflix.
So I asked for the system, not the style
> I know Hollywood and even personal bloggers have a system to create subtitles for their videos. Study these and build me that system. ⎘ jw.mov · John Wick: Chapter 3 ⎘ suit-and-tie · Justin Timberlake ⎘ (no file) · an Instagram reel, described from memory
paraphrased — the intent survived, the typos didn’t
I didn’t ask for cool subtitles. I told Claude there must be some kind of system behind the ones I envied, and handed it a movie, a music video, and a reel I could only describe in words as artifacts it could study.
It came back with a lab
It didn’t overlay captions on a video but gave me a full tool where I could manipulate and audition caption styles overlaid on my video, complete with scrubbing, four switchable styles, with colors, fonts, effects, and even a characters-per-second budget meter I never asked for.
It wasn't perfect
left: real text at 0.42×
Somtimes the font size was a bit too small. Or the caption was too big taking over the entire screen.
This was never a one-shot
Coming from video editing tools like Adobe Premiere and Capcut things much different with Claude. I went through multiple versions getting things like timing, size, placement, etc right because each fix had to be via a prompt vs using your mouse and editing the timeline yourself.
The model was never the ceiling.
What’s actually left limiting you
What you get out of AI is bounded by what you can imagine asking for, what you’ll keep iterating on after it not right, and the hours you can give it. Intead of going to version 40 on my video I could've just done everything in Premiere. I could have also asked Claude or Codex to build me a video editing tool too.
Here’s a short clip of how it looks now
Not done. But the lab makes the next pass cheap, and that’s the real deliverable.