video-use

Transcribe footage with word-level verbatim ASR, pack the transcript into phrase-level lines with timestamps, then select cut ranges that snap to word boundaries. This Skill automates that flow: transcribe, propose a strategy, confirm, then render cuts, grades, overlays, and subtitles with ffmpeg.

Extract per-segment clips whose start and end times align to word boundaries from the transcript, pad each edge by 30-200ms, and add 30ms audio fades at every boundary. Concatenate the segments with lossless -c copy rather than a single-pass filtergraph to avoid double encoding.

Subtitles must be applied last in the ffmpeg filter chain, after every overlay. If subtitles are burned into the base video before compositing, overlay frames render on top of the caption pixels and hide them.

This Skill supports HyperFrames for HTML/CSS/GSAP compositions, Remotion for React-based compositions, Manim for formal diagrams and equations, and PIL with PNG sequences for simple cards like counters and typewriter text. Each animation renders in its own slot directory via a parallel sub-agent.

Yes, transcription uses the hosted ElevenLabs Scribe API, so an ELEVENLABS_API_KEY must be set in the environment or in a .env file at the skill repo root. ffmpeg, ffprobe, and Python dependencies must also be installed; Node.js is needed only for HyperFrames or Remotion animations.

Hard cuts between segments create waveform discontinuities that produce audible clicks. Apply a 30ms audio fade-in and fade-out at every segment boundary, and verify the rendered output's waveform at each cut to confirm no spike slipped through.