Transcripted v0.3.0

Eight PRs merged today. That number is misleading.

The PyAnnote switch was the one I was most confident about. Sortformer was producing phantom speakers — segments getting attributed to ghosts. PyAnnote was supposed to fix it. The switch was clean, dropped right into the pipeline. But I don’t actually know yet if it’s better. I don’t have a proper test suite for diarization quality. I shipped a model swap based on theoretical accuracy improvements and a handful of manual tests. That’s not nothing, but it’s not proof either.

The onboarding simplification — six screens down to three — felt good to ship. What I didn’t fully reckon with is that I simplified it because users were dropping off, not because I understood why. I cut the screens that seemed redundant. I still don’t know if I cut the right ones or if the drop-off was about the screens at all.

The evening PR batch was UI polish: speaker names in the tray, menu bar redesign, onboarding callout, README screenshots. Each one improved something real. But I was shipping fast and not thinking carefully about whether any of it addressed the actual friction points users hit.

Eight PRs is a lot of motion. Motion isn’t the same as progress. I don’t have enough users yet to know if any of this moved the needle on retention or activation. I’m building on instinct and moving fast, which is the right call at this stage — but I should be honest that’s what it is.

Red bar: shipped a lot, proved very little.