A Prompt for Synchronizing Sound with AI Motion-Design Videos

If motion-design videos created by Opus 5.5 still sound like cheap electronic squeaks, try this prompt.

Prompt for the assistant:

Add sound to this video. You cannot hear, so measure everything and do not rely on subjective impressions. If ffmpeg and librosa are not installed, install them first.

1. Use music with a suitable license: either the track I provide or a composition distributed under a free Mixkit license. State the source and license. Do not generate music or sound effects programmatically.
2. Determine the musical beats with librosa. Sometimes the library locks onto weak beats by mistake, so plot the waveform in the range below 120 Hz, overlay a beat grid, and zoom into several bars. If the kick drum hits do not align with the grid, shift the entire grid.
3. Sync major motions, such as scene cuts and title appearances, to the musical beats. If you can change the animation, change it. If you cannot, shift the start time of the music.
4. For each frame change and each significant motion, use a separate ready-made sound-effect file: a whoosh for movement, a pop for element appearance, a click for presses, and a keyboard typing sound for text input. First use the existing project sound library, for example the one shipped with HyperFrames. If there is no such library, ask me.
5. If the video has a voice-over, it must be louder than the other sounds, and the music volume must be reduced during speech. Sound effects should be louder than the music. Aim for an overall loudness of about -16 LUFS and a peak level below -4 dBFS to avoid clipping after AAC compression and playback through a phone speaker in mono. Measure the loudness of the exported MP4 with the ebur128 filter in ffmpeg. Do not measure the WAV file.
6. Use ffmpeg to extract one frame before and one frame after each sound-effect trigger from the exported MP4, then plot the waveform for comparison. After that, choose five timecodes I can check with headphones. Do not claim that the sound turned out good.

Real-world example: A few days ago, while selecting music for a video, librosa mistakenly synchronized to weak beats. The entire rhythm grid was shifted by 0.22 seconds. The problem was discovered only after overlaying the grid on a waveform plot below 120 Hz.