Notes & essays

@paper + claude/codex is the single best way to make ig carousels, thumbnails, infographics for social

highly recommend

When I was making YouTube thumbnails, finding the perfect facial expression took forever. But YouTube is still a platform where you’re watching for a person, and it’s a click and watch platform, so you do want to put faces on the thumbnails generally.

So I built out an AI skill that helps me do this. So in this one video, it scanned 5,539 different frames. It threw out all the bad ones, ones where I like tucked my chin, mouth half open, mid-word, eyes closed, things like that. And it made sure that it was a real expression that I held, not like some face that I made for a split second.

It uses this model from Cloudflare called CLEF. It’s an image decision model to rank the different faces that I have. I get a sheet of the top faces to pick from.

It’s not perfect right now, but it is pretty good, and it saves me tons of time. At this moment, it doesn’t know exactly what I would want, so it can’t make the final decision, but it gives me a sheet of what I can choose from, and it’s getting pretty close.

Then I use this app called Paper to mock up the different thumbnails with all of those different facial expressions. So I can just run a command, and I can get different versions of the packaging, and I can just step away and do something else.

So if you want this particular skill, comment face, and I’ll send it over to you.

Spent some time this past weekend building out a full AI YouTube thumbnail pipeline from video → 234 thumbnail options (little excessive)

1. Find a face: Raw video → extract & filter frames → filter for shots based on eyes, framing, etc. → classify using @Cloudflare CLEF-flash → contact sheet → pick the best ones

2. Assemble thumbnail: My picks → 2x upscale via @topazlabs high fidelity v2 → references via benchmark.social → my own ideas for the framing → thumbnail angles and text options → different layouts with @paper = hundreds of versions

3. A/B/C test on YouTube Studio

It’s a bit janky right now but this is a great way to see tons of options at once.

I’ve been trying every single AI tool out there to see what’s best for my video editing workflow. And something that I’ve realized is that the best AI video editor out there just depends on the kind of video that you’re editing. Here are 5 that I’ve tried and what I actually like about them.

Number 1 is just straight code like Python and FFmpeg. If you don’t know, FFmpeg is a free software that you can use to edit videos, manipulate video, audio, whatever with text commands in your terminal instead of the timeline. I’m using it to cut, trim, speed up, crop, even add captions, color grade, clean up audio, EQ, the entire thing.

And a lot of times, plain FFmpeg is not only way faster, it can be way more token efficient and generally easier to give the model feedback.

However, it is extremely annoying to use when you want to make small tweaks that you could have just made on a video editing timeline. That’s because to see the results, you’re gonna have to wait for the entire video to re-render.

I think it’s great for quick edits, color grades, audio processing, short form lo-fi quick edits. FFmpeg sometimes gets the job done, and you don’t need anything else.

Second is HyperFrames. I think it’s great for animation. This is a free tool, open source by a company called HeyGen. Like a lot of these other tools, it’s turning code into video with the base of this being just straight HTML, and agents are really good at it because they know how to write HTML.

They’ve also released recently a new HyperFrames Studio, which is like a straight up video editor. You can hook it up to Claude and ChatGPT. Right now, I’d probably stick to using it for animation because that’s what it’s best at.

Number 3 is this new one called Palmier. This is like a proper video editor that they’re building out here. It’s a startup. It’s not as robust and built out as other video editors because they just got started, but it has great MCP support, so it can connect with Claude and ChatGPT and your other AI agents.

And one of the nice things is that it is pretty fast for a lot of basic edits. It can do most of the stuff that you would need for editing talking head videos, short form videos, adding captions. If you’re slapping on a LUT or something, it can do all that sort of stuff.

I had really no problems with Palmier. I just think there’s some basic table stakes that they’re building out just because other video editors are just so robust. And one of the nice things is that you can work on multiple edits at a same time.

Next up is Diffusion Studio. This is also free and open source. This can also run in your browser, but they also have a native app that you can download. The UI is really nice here, and one of the cool features is their art boards, which I think is great for repurposing content, resizing things. So if you’re trying to take like a 9 by 16 video, make it 16 by 9, it’s pretty good.

Diffusion Studio, I don’t know, maybe I just didn’t set it up properly, but I tried a lot and the agent was just struggling a little bit with it. Also, it’s a tad buggy. It’s missing some features here and there, but the UI is really great and it’s open source. You can actually go download it and just like tweak the software, if you want, for your own use case and like customize it, and they have examples on how to do that.

And then we also have DaVinci Resolve for proper video editing. Obviously, DaVinci Resolve is one of a kind for color, especially if you can get the skills right. It’s the editor that tons of filmmakers use. Claude can now control it. I still think that it’s not perfect because they haven’t exposed all of the full MCP support inside the DaVinci Resolve platform.

It’s still slower to edit some videos on this than compared to like Palmier or FFmpeg or one of these other tools. But if you’re native to DaVinci Resolve, you already use it, or you don’t mind using like an NLE where you’re working on like a bigger project, then I do think this is worth it to set up.

Every social media app is starting to look the same, and they’re all turning into YouTube.

In July, YouTube was fourteen point two percent of everything that Americans watched on TV. Netflix was at seven point eight percent. That means that people are watching almost twice as much YouTube on their TVs than Netflix, and every other social media app wants the same part of that pie.

Instagram went from ninety-second Reels to twenty minutes. They also made a TV app. Spotify has gone all in on video podcasts. They’re even letting you upload YouTube-like thumbnails onto the Spotify feed. If you open X right now, the timeline is littered with video. Even LinkedIn is pushing video, albeit to a smaller extent.

These platforms know that short-form video is great for getting attention, but when somebody sees your thirty-second Reel, maybe they like it, they probably keep scrolling, and then like two minutes later, they forgot the video that they just watched.

To keep you for twenty minutes, Instagram, at minimum, has to serve you twenty Reels. In reality, because everyone’s just skipping the videos that they’re watching, it’s more like seventy-some Reels that the Instagram algorithm is feeding you in that twenty-minute time span.

But YouTube only needs one video, and when you spend those twenty minutes with one person, you actually remember who they are.

It’s also not just that. The money aspect has a lot to do with it. If you’re spending twenty minutes doomscrolling on a short form platform, how many ads could they possibly fit in there? Maybe three, four. On top of that, all the other ads that the regular creators are doing as well. How many ads could you realistically see without getting upset?

YouTube, just a single twenty-minute video would have an ad before. Probably, you could run a midroll ad, maybe even an ad after that.

This is so much more lucrative, especially if Instagram can get both short form and long form working on top with their regular ads platform that they already have. That’s why every app feels like YouTube right now.

I’m literally prompting while I’m on the toilet, before I go to the gym, before I go to sleep. It’s like a never-ending thing right now because there’s always this feeling that you could always be more productive with your time.

Obviously, it’s a little toxic for me right now when I think about it, but here’s what I keep thinking about.

1 person could be extremely good at their job. Another person may be not as good, but could be doing 4 or 5 different things at once. They’re taking the time to set up the necessary stuff and the systems and the AI agents.

They think about, “Which agents do I have to run? How do they work together? What’s gonna take a while that maybe I can put on the back burner while I work on something else in the meantime?”

And I think those are really the skills that are going to matter now. And so this idea of the ADHD brain being the most valuable skill in like 5 years, or even right now, is something that I’ve been thinking about because in this AI world, it does benefit from being able to jump around multiple things at the same time.

Now, there are times where you have to sink in and go deep on something: writing, working on videos, strategy, big projects. But even then, there’s something else that you could be doing in the background while you’re working on something deep.

The hard part, though, is like knowing your limit of when is too much or when I have to go deep on something. Like for example, like sometimes I just have too many things running, too many agents running, and things can slip through the cracks. I just forget that I have to do them.

And us as humans, we all have a brain. Our brain has a limit on how many things we can actually keep in our heads. Um, I don’t know yet how people are going to raise that limit. I don’t think it’s actually possible, uh, without burning out.

What do you think? Do you think the ADHD brain will be valued in a world where everyone is just prompting AI to do knowledge work? Or will it be too much of a distraction where you have too many prompts and you actually never get anything done?