Launch video sound
Sound carries half of a launch film. Measured on a studio-made reference (a ~100 BPM track): the big story turns start on musical section changes, on bar lines; the music stops down (drops out for about one beat) just before a turn; shots inside a section cut freely. The rule this skill enforces: turns on bars, everything else free.
Steps
-
Choose the order. Music-first with a library track: lock the track, grid it, then time the storyboard to its bars. Picture-first when generating: fix each act's length in the storyboard, then generate to those lengths (ElevenLabs
musicskill, composition-plan chunks of 3 s or more) or score the finished silent cut withvideo_to_music(add_sfxandseparate_tracksreturn music and effects as separate stems). Done when the project's brief names the order and the source. -
Source the music from the table below, under a licence that covers commercial use in ads and on social. Done when the file is in
public/audio/andpublic/audio/CREDITS.mdnames its source and licence. -
Grid it. Decode and run the bundled script:
npx remotion ffmpeg -i public/audio/music.mp3 -ac 1 -ar 22050 -c:a pcm_s16le /tmp/music.wav python3 ~/.claude/skills/launch-video-sound/scripts/beat_grid.py /tmp/music.wav --fps 30 --out src/beats.jsonIt needs
numpy(a throwaway venv works). Done when the printed BPM matches tapping along, the first downbeat sounds like "one" (confidence under ~1.0 means check by ear), andsrc/beats.jsonholds bars, sections and drop-outs. -
Time the turns. Every act boundary, hero reveal and the logo hit starts on a bar (a listed section change is the strongest choice). One helper is the single source of musical time:
import grid from './beats.json'; export const bar = (n: number, beat = 0) => Math.round(grid.bars[n].frame + beat * grid.beat_frames);Done when every act boundary in the timeline is a
bar(n)call, with no frame literals. -
Cue the effects. Write each on-screen event into
src/cues.tsas{frame, file, gainDb}usingsfx-palette.md, and render all effects from that one list. Done when every visible action the palette names has a cue and every cue's file exists inpublic/audio/sfx/. -
Mix. Automate the music with
<Audio volume={(f) => …}>: the stop-down before each turn (to 0 over 2 frames, one beat long, back on the downbeat with an impact), and a −10 dB duck under any voice. The film ends on a sting or hit on the logo frame, letting its tail ring out. Done when a Studio playthrough has every cue audible over the music and no clipping. -
Master and measure.
npx remotion ffmpeg -i out/video.mp4 -af ebur128=peak=true -f null - 2>&1 | tail -12Target −14 LUFS integrated (±1) and true peak ≤ −1 dBTP. Out of range: run two-pass
loudnorm=I=-14:TP=-1:LRA=11on the audio and remux with-c:v copy. Done when the measured numbers are in range.
Music sources
| Source | Best for | Licence |
|---|---|---|
| Library: Artlist, Epidemic Sound, Musicbed | The hero track; what studios use | Subscription; check the plan covers ads |
ElevenLabs music (composition plan) | A track cut exactly to the storyboard's act lengths | Paid plan for commercial use; free tier is non-commercial with attribution |
ElevenLabs video_to_music | Scoring a finished silent cut, with effect stems | Same as above |
Synthesised in code (engine/ from the open-source animation-studio skill, MIT) | Free, no account, every sound exactly on its frame; how the Beemcute film was scored (see the kit's sound/README.md) | Yours |
When scoring in code: put every on-screen event's frame in one EV list copied from the scenes, write music per act (tension before the turn, lift after), give each visible action one synthesised sound from sfx-palette.md, mix on separate buses with effects about twice the bed, and master to −14 LUFS in the same script.
Feeds autoplay muted, so the on-screen text tells the whole story; sound is the reward for unmuting.