Long-form write-ups from inside the work. The problems we hit, the decisions we made, and the ones we got wrong before we got them right.
We want an agent to understand what happened in a screen recording, frame by frame, even when there is no narration to lean on. This is the research behind that feature: why we anchor on pixels instead of audio, why we picked a text-first pipeline over feeding screenshots to a vision model, and the two things I had backwards before I looked them up.
A 46-second recording took five minutes and forty-two seconds to export on a machine with a hardware encoder. This is how we found the real cause, threw out a confident wrong guess about threading on the way, and got it down to a minute and a half.
Recast started as a Windows app. Here is what it took to run it natively on macOS and Linux: the two decisions that kept it from becoming a mess, the freeze that only happened on a Mac, and exactly where each platform stands today.
We replaced the editor's video element with a decode pipeline we control: WebCodecs for video, a second decoder to hide cut latency, resolution-aware caching, range-based loading for large recordings, and a Web Audio engine that keeps sound aligned with cuts.