video-speech-workflows · git:20260829.81fc760 · 2026-08-29 · sha256 b970e2805ab71816
video-speech-workflows git:20260829.81fc760A
Immutable. This exact content never changes and is served at /api/v1/blob/b970e2805ab71816.
--- name: video-speech-workflows description: Consolidated single-source voiced-video workflows for speech condensing, talking-head subtitles, and end-to-end caption/edit orchestration. Use when video-edit-agent routes a single voiced video to mode=speech-condense, mode=talking-head-subtitles, or mode=video-pipeline. --- # Video Speech Workflows This skill contains the specialized single-source speech/video workflows that sit beside `video-edit-assembly`. ## Route Modes Choose exactly one mode from the caller's args, then read and follow the matching workflow document: | Mode | Workflow | |---|---| | `speech-condense` | [workflows/speech-condense.md](workflows/speech-condense.md) | | `talking-head-subtitles` | [workflows/talking-head-subtitles.md](workflows/talking-head-subtitles.md) | | `video-pipeline` | [workflows/video-pipeline.md](workflows/video-pipeline.md) | If the args do not name a mode, infer the narrowest matching one: - Use `speech-condense` for one long unscripted recording that must be shortened, tightened, or cut to a target length. - Use `talking-head-subtitles` for straight captioning, subtitle repair, speaker-labelled captions, karaoke highlighting, bilingual lines, or subtitle placement/legibility work on one rendered video. - Use `video-pipeline` when a single voiced video needs an end-to-end stage plan that may combine condensing/editing with captions, speaker differentiation, karaoke, translation, or terminology notes. ## Boundaries - Multi-material source packs stay in `video-edit-assembly` or the common `video-edit-agent` workflow first. Subtitle the rendered output afterward so cue times match the final video. - The selected workflow's file contract and verification rules are authoritative for its own outputs. - `video_watch_segment(video_path=...)` can directly inspect tight windows without a prior full `video_ingest`; reserve full ingest for workflows that genuinely need dense whole-video contact sheets.