Heyn Motion Kit fixes narrated video with a $39 Claude Code skill

Ad-hoc chat requests for video produce three recurring problems: paid voice synthesis fires before the script is confirmed, UI code gets rewritten each time so visuals drift, and a finished render is treated as done even when the picture froze or the audio came out quiet. The skill file addresses each by fixing the order of steps and writing the completion bar into the file itself — script approval sits before any API call, the agent only authors a spec JSON with timing anchors like "@why+0.35", and a QA step returns OK/FIX lines the agent must clear before re-rendering.

The technical middle of the piece shows how ElevenLabs character-level timestamps become word timings: splitting on spaces for English, and using Intl.Segmenter plus a rule that attaches particles and punctuation for Japanese. The author notes that newer ElevenLabs models (eleven_v3 / eleven_v4) produced captions 0.7–1 second early in their testing, and suggests comparing silence detected in the audio with punctuation in the timestamps as a cheap numeric check.

The QA layer relies on ffmpeg's freezedetect for still stretches over 0.5 seconds and ebur128 for loudness, with -14 LUFS as a social-media target; in the author's demo the first pass flagged a 1.07-second freeze and -25 LUFS, fixed with a slow 8% zoom and two-pass loudnorm. The piece is candid that numeric QA is a floor, not a taste judgment, and that anything the agent cannot verify — like how the voice sounds — should be reported as unverified. The commercial kit extends the same design to beat-synced cuts and 9:16 / 16:9 / 1:1 outputs.