Cutback: Edit Video by Talking

Cutback is a short-video editor you drive by talking. You drop in a clip, AssemblyAI writes the transcript, and a voice agent takes instructions like "cut the intro," "caption this," or "make it vertical." Each instruction becomes a real edit on a timeline, not a chat reply you still have to apply by hand. The player, the on-screen captions, and the export all read the same keep-spans. What you preview is what FFmpeg renders. You can export MP4 or WebM at 1080p or 720p, with burned-in captions and a separate subtitle file. Several clips sit end to end. You can undo, redo, reframe, and mix speech against music. On a phone the mic stays on the bottom edge; on a desktop it sits under the picture. Voice is the editor, not a side panel. The browser holds a live AssemblyAI voice session. When you speak, playback pauses so the video is not treated as an instruction. The agent calls tools for cuts, captions, framing, and audio. Ambiguous requests ask a follow-up instead of guessing. A proposal is shown before it lands, and the same call cannot cut twice. Limits are honest. A project holds up to two minutes and 200 MB per file. Cutback does not invent missing speech or auto-track a face. Transcription and the live session use AssemblyAI only. The stack is TypeScript, React, a Fastify server, and FFmpeg.