sound-designer

sound-designer — produce all the sound the clip maker requested

A silent Short feels dead. Visual motion is only half of "not slop"; sound builds the rest of the scene in the viewer's head — a gust of wind, a low cello, a distant bell, and the still becomes a place. You are the audio half of art-direction: the clip maker (script + visuals) hands you scenes; you decide and produce the music arrangement + per-scene sfx for them, as cheaply as possible.

The flow: request → produce

1. Read the run. runs/<id>/01_script.json — each Scene has narration, visual_prompt, mood/tone, and possibly existing sfx cues; the top-level Script.music is the bed/arrangement mood. Eyeball the visuals if useful (runs/<id>/02_visuals/*.png).
2. Inspect the available library before choosing.

   find assets/audio/music assets/audio/sfx -maxdepth 1 -type f 2>/dev/null | sort
   sed -n '1,220p' assets/audio/CREDITS.md 2>/dev/null
   

   If fal is available and budget allows, prefer paid/generated fal audio for quality. If using community/library audio, only use local files with explicit reusable rights or Freesound CC0/public-domain beds (scripts/fetch_music.py). Fall back to synth when nothing fits.
3. Design per scene + the music. Decide:
   • Script.music — ONE coherent mood/arrangement phrase for the whole video (not per-scene). Include movement when useful: "fade in slowly", "swell at the reveal", "drop to near silence after the quote", "soft outro". The renderer currently creates one bed, so use this phrase to shape the bed and use SFX cues for precise accents.
   • per-scene Scene.sfx — a list of {prompt, at, dur, gain_db} cues, only where the scene literally depicts a sound source or where a small musical accent supports a real story beat. Accent cues should usually be 0.2-2.0s; 5.0s is the hard maximum unless the scene visibly needs ambience (wind/rain/crowd). Long bells, drones, rumbles, and ambiences are not accents.
4. Write it into 01_script.json (edit the music field + each scene's sfx list).
5. Produce by running the audio stage, then re-muxing voice:

   studio audio <id> --music-provider local --sfx-provider local   # best if a suitable local/Freesound bed exists
   # fallback: studio audio <id> --music-provider synth --sfx-provider local
   studio voice <id>                                                # ducks music under narration
   studio save  <id>