Video Captions
Read caption workflows. Choose exactly one caption input: explicit words, a transcript JSON file, or local transcription. Word entries are {text,start,end} in seconds. Never estimate timestamps and present them as measured alignment.
Use the optional faster-whisper environment for transcribe: true. Its timestamps are ASR estimates. Review technical terms and alignment against the sound, especially after retakes or repeated phrases. Empty or out-of-range captions stop the job.
The renderer groups short phrases and highlights active words; the runner preserves transcript.json and SRT. Check punctuation, line length, contrast, interface safe areas, and placement against the subject. Font coverage must be tested for the chosen language.
Translation is an agent-authored reference workflow. Keep the original timing unless meaning requires revised segmentation and alignment. Bilingual styling and ASS sidecars are not built-in outputs in v0.1. Read the reference for available approaches and engine limitations.
For reels, use the protected content band described in reel direction. Keep phrase grouping readable across cuts; words can emphasize actual speech, while arbitrary beat timing must not replace measured word timestamps.