embedded-captions

Run hyperframes init on the clip, then scripts/prepare.sh to matte, transcribe, and compute safe zones in one command. Author a small JSON for the chosen identity, preview composite frames, and run render-and-composite.sh to produce final.mp4.

Pick from CATALOG.md by content register and scene luminance, never by engine name. Bright scenes above luma 150 use ink, verbatim explainers default to the anchor rail, and dark poetic scenes fit cream; the agent recommends one and the user confirms.

The skill refuses clips with multiple speakers, hard cuts, no human subject, under 3 seconds, no speech, existing burned-in captions, or garbage transcripts. Fast handheld footage also fails because the matte flickers.

No, the source video ships untouched; captions are the only addition and the matte only lets the subject occlude embedded text. The single exception is Theme mode's plate budget, which applies register-gated reaction effects after compositing.

Screen-blend cream templates fail above 180 luminance and those DNAs are locked against recoloring. The fix is to switch identity to ink, which uses near-black multiply blending designed for bright surfaces, or the opaque anchor rail.

It requires a built hyperframes checkout for the remove-background CLI, plus ffmpeg, sharp, and puppeteer resolved from that checkout. Transcription runs WhisperX through uvx, and matting weights auto-download once to the hyperframes cache.