Mission
Create a readable, reviewable caption layer for an approved video. Preserve the original cut and distinguish machine transcription, human corrections, translated text, burned-in captions, and sidecar files.
Use Video Assembly when captions are one part of a final timeline. Use this skill when caption accuracy, accessibility, or platform variants is the main deliverable.
Required inputs
- approved video parent and measured duration;
- source language, speaker names, terminology, numbers, and exact copy;
- caption type: transcript, subtitles, translated subtitles, or accessibility captions with sound cues;
- target format, ratio, safe area, line length, reading speed, style, and burn-in versus sidecar output;
- music/dialogue mix and whether non-speech audio cues are required; and
- rights, reviewer, disclosure, and definition of done.
Never silently rewrite uncertain speech, product claims, names, or legal copy. Mark uncertain spans and request a source transcript or human correction.
Pipeline
- Freeze the parent video and record duration, frame rate, language, speakers, and source rights. Do not caption an unapproved or changing cut.
- Generate or ingest a transcript using
media.caption_videoor an approved host-local path. Retain timestamps and confidence/uncertainty when exposed. - Normalize punctuation and line breaks without altering meaning. Apply approved terminology, speaker labels, sound cues, and translation rules.
- Check timing against speech, shot changes, music, and visual text. Keep captions within safe areas and readable at the target size; test high-motion backgrounds and mobile framing.
- Render a captioned draft or return a sidecar/timeline handoff. Poll the job and record the exact output. Do not assume caption generation burns text into a video or preserves every styling field.
- Review the complete output with video QA: spelling, names, numbers, claims, timing, contrast, overlap, omissions, reading speed, export compatibility, and disclosure.
Check MODELS.md and muapi-video-tools.md for the current caption operation, input shape, and output behavior.
Output contract
Return parent ID, transcript/caption version, source language, confidence or
uncertainty list, corrected terms, caption format/style, model/endpoint,
prediction ID, output URL or sidecar, QA findings, unresolved spans, and the
next reviewer action. Save permitted text and receipt artifacts under .video/
using reports-and-receipts.md.
Captioning is draft-only. It does not certify accessibility or platform compliance unless a current rule set and human review support that claim.