sound

Sound for a launch film

Sound is half of "premium". On the Auni film the first critique opened with it: "the sound effects kinda suck which i think ruins the premium feeling". The fix was not cleverer code. It was a real SFX pack, a real song, and a sound on every movement. This skill covers the song, the effects and the mix. Commands are the film repo's (npm run ...), made by ${CLAUDE_PLUGIN_ROOT}/kit/scripts/new-film.sh.

1. The ten rules

1. Real SFX are the single biggest jump in perceived quality. Generated sounds are for timing only. Auni v1's sounds were synthesised from code; the user's verdict above came first in the first round of notes.
2. Every movement makes a sound. The user: "any movement or click or mouse move or transition or loading spinner or anything should have an associated sound effect... sound effects secretly apply a polished premium feel. Dont be afraid to really use a lot of sound effects." The final film had 352 cues in 39.4s, about 9 a second.
3. A sound matches something on screen. "sound effects should try match some visual effect": a shine sound on a gloss sweep, a riser into a glow. No sound without a cause, and big moments get a visible flourish for the sound to sit on.
4. Pick the song before you animate. The film is cut to the song's grid, not the other way round. Fast is about 150 BPM, medium 120, slow 100.
5. Count everything in beats. Beat 0 is the drop. beat_s turns beats into seconds. Picture and sound read the same beat names, so they can't drift apart.
6. The drop is the reveal. Tension before it (a quiet bar holding the question), the product's first appearance exactly on it.
7. Place each sound by its loudest point, not its start. A whoosh peaks on the cut; a riser ends on the drop. The kit measures every file's peak once (public/sfx/stats.json).
8. Write cues next to the picture, in the scene's own sounds(add), on the same beats and the same helper functions the motion uses.
9. Measures shortlist, ears choose. You can't hear. The best-measured track of 465 (+71 dB bass jump) was "not hype enough". Check the licence before anyone falls in love with a song.
10. Measure the file you deliver. Encoding to AAC raised Auni's true peak by 4 dB. Loudness is read off the final mp4, never the wav.

2. The pipeline

 refs -> identify (Shazam) -> licence -> candidates -> scan drops -> auditions
                                                                        |
                                                    USER PICKS BY EAR <-+
                                                            |
 lock grid (BPM, drop, phase) -> music.json -> scenes in beats -> sounds(add)
                                                                        |
 measure the delivered mp4 <- draft/final <- mix.wav <- splice (song()) <+

3. The steps

• Study the references. npm run refs -- <video> out/refs/<name> for each reference film. Read analysis.json and audio_vs_cuts.png. Expect cuts on sound hits more than on beats (the reference studio: 26 of 28 cuts on a transient, 14 of 28 on a beat). Shazam runs at 3 points, which catches films that switch songs.
• Fetch clean copies. npm run music:fetch -- "<artist> <title> [instrumental]". It refuses a file Shazam names differently. Prefer instrumentals: vocals fight on-screen text.
• Licence first. Before shortlisting, write each song's route in the table from music.md. A major-label song needs sync plus master licences: five figures, weeks, and it may be refused.
• Find the drops. npm run music:scan -- <folder> ranks a folder on three measures. npm run music:grid -- <file> --drops lists every drop in one song, with its build.
• Audition. npm run music:audition -- <file> <drop_s> [--film-drop 9] cuts a 20s clip with the drop where the film's drop is, and refuses clips with no real bass jump. Send YouTube links timed about 9s before each drop (the script prints them) in your final message. Links written above a question box were never seen: "where ? i cant see them here".
• If the user delegates the pick ("your pick bro. use your best judgement"), build on your best guess but call it provisional. Auni's delegated pick was rejected on the next listen. Swapping is a re-map of beats, not a rebuild.
• Lock the grid. npm run music:grid -- public/music/<file> --bpm <n>. It fixes the tempo, finds the drop from the bass, fits the phase to the kicks, and fails above 30 ms of error. It prints a bar table and writes src/music.json. Then set licence in it.
• Map sections to scenes. Quiet bar holds the question, drop is the reveal, break is the hero shot, re-drop is the logo. See music.md section 6.
• Check the seams. npm run music:seams scores every phrase jump. The engine's song() splices on a scene cut when scenes run long or short, and npm run check throws a message with both fixes when the lateness isn't whole phrases.
• Write sounds(add) next to each scene's picture, using roles (v("whoosh", i)), not file names. The event-to-sound table is sfx-vocabulary.md.
• Mix and check. npm run mix prints cues, length, LUFS and true peak. npm run audio:check runs the no-ears bar and writes out/cue-sheet.png. Read the PNG.

4. What you can and can't hear

You can't judge taste. You can prove a lot without ears:

Question	How to check it
Did a refactor change the sound?	Diff out/sound.json cue lists before and after (Auni: "236 236 identical")
Is it loud enough, and safe?	npm run final prints LUFS and true peak of the mp4
Is the splice invisible?	npm run music:seams: 0.9 or more, under 3 dB step
Does every motion have a sound?	npm run audio:check lists named beats with no cue within 2 frames
Is a cue buried?	audio:check flags a cue within 40 ms of a louder one on the same side
Where are the gaps and pile-ups?	Cues per second, printed by audio:check, and the cue sheet PNG

Say the limit out loud to the user: "My numbers can only tell you a track has a drop, not that it sounds expensive."

5. Bring your own audio

The kit ships generated stand-ins so timing works on day one. They are not good enough to show anyone. Full steps are in sourcing.md.

• SFX pack: copy files into public/sfx/, run npm run sfx:stats, map each file to a role in public/sfx/roles.json, run npm run check (it fails on any cue whose file is missing).
• Song: put it in public/music/, run npm run music:grid -- public/music/<file> --bpm <n>, set licence in src/music.json, run npm run check. The splice guard tells you how many beats to add or cut.
• Never commit licensed audio to a public repo, and never put any audio into the plugin itself. The kit's .gitignore already keeps public/music and public/sfx audio out.

6. Reference files

File	What it holds	Read it when
music.md	Tempo bands, reference songs, clean copies, licences, scanning, drop maps, auditions, the grid lock with numbers, music.json, mapping sections to scenes, the phrase splice	Choosing, auditioning, locking or swapping a song
sfx-vocabulary.md	Visual event to sound, by kit role, with timing and level; layering, density, panning, the levels guide	Writing or reviewing any scene's sounds(add)
mix.md	The mix steps, the measured loudness table, every failure and how the kit prevents it, the no-ears bar	Running or debugging the mix, before the final
sourcing.md	Where SFX and music come from and their licences, installing your own pack and song, what never ships	Day one of a real film; any licence question

Related: launch-film:make owns the grid and budget (../make/SKILL.md); launch-film:review checks sound coverage before any cut is shown (../review/SKILL.md); launch-film:assets covers sounds for 3D moves (../assets/SKILL.md).