Svipall

Last time I posted Svipall here I admitted it couldn't read a page that's only a video, and that bugged me more than I expected 😅. It started with an animation I saw in a repo, one of those flows that play in the background to show how the tool works under the hood. I gave Claude what I thought was a decent prompt and one screenshot of the flow, and it didn't get it quite right. Once I gave it three screenshots of three different states of the flow, it copied it pretty well. That's when I thought: if I ever want to copy something harder, like the animation in a product demo video, I'm not taking screenshots by hand all afternoon hehe. So in 1.2.0 Svipall reads video. You give it the link, it picks up to 24 frames at the moments where the picture changes, and it lines them up with the video's own subtitles. If there are no subtitles, it turns the audio into them, locally on your machine. I tried it on that same flow first. It worked. Then I tried a showcase animation, and I was really happy with the results without taking a single screenshot. It also saved me on a YouTube video that was too good not to keep for research: no copying subtitles by hand, just the link. A lot of open source tools still don't do this, and with models like Opus 5.5 being this good at design, handing them the video instead of a pile of screenshots felt like something that should exist. I want Svipall to be the ultimate web exploration tool, and I thought it was already pretty good, but audio and video were two big holes. They matter a lot when you're researching, or when something only exists in a video, and typing it out by hand is a pain. There are still two things I don't love. You can't pick which seconds of the video to use as the reference. The AI usually figures out which part you mean from the frames, but a range would be much more useful, and I left it out for practical reasons for now. And some audio, like Opus inside WebM or MPEG-TS segments, can't be decoded yet. When you're searching around the web you don't check what format each thing is, so you get an error and you're a bit blind unless you go read the logs 🙃. I want to add the range soon and cover more formats, so the only thing you have to worry about is finding the right video. Repo's here if you're curious: https://github.com/ilien-dev/svipall Here's one of the results I got, a GIF animation for one of my own projects. https://preview.redd.it/wgagmkdzthsh1.png?width=1800&format=png&auto=webp&s=8329c65af85192be8df6248ebaa682df0e0afa83 And this is the prompt I used to do it with this tool: Watch the animation in https://github.com/vectorize-io/hindsight#the-three-operations and recreate it for my current project: same kind of flow and motion, but with my project's own components and steps. You can see the original animation I wanted to copy at the link in the prompt.