Captions & subtitles

Two outputs, one transcription engine (mlx-whisper, same as transcribe):

• Sidecar file (.srt / .vtt) — for YouTube/Vimeo upload, re-editing, or other tools.
• Burned-in — pixels baked into the video (social clips where the player won't show a subtitle track). Rendered via PIL → ffmpeg overlay (mise ffmpeg lacks drawtext).

CRITICAL: use --clean for captions

This is the opposite of filler-removal. Whisper is verbatim by default here (keeps um/uh) because filler-removal needs them. Subtitles do not — viewers don't want "um, uh" on screen. Always pass --clean when transcribing for captions:

uv run video-agent transcribe video.mp4 --clean --srt -o subs.srt

Export a subtitle file (no burn-in)