sound_design_video v1.0.0

name: "sound_design_video" description: "Design sound for video: choose and layer sound effects, record DIY foley, build seamless ambience loops, and source royalty-free SFX libraries. Use when a video edit needs sound effects, foley, ambience beds, or transitions under dialogue. Pairs with audio-mixing, which owns the dialogue and music mix chain."

Sound Design for Video

Purpose

Take a locked or near-locked cut and give it a complete sound world: picked and layered SFX, DIY foley synced to picture, an ambience bed that glues scenes together, and transitions that move the audience between them. Delivers source-picked, license-clean audio elements ready for the mix.

Workflow

1. Spot the cut. Watch it straight through and mark every moment that wants sound, on a simple cue sheet: hard sync hits (doors, impacts, clicks), movement (footsteps, cloth, props), location glue (room tone, city, nature), and transitions (scene changes, reveals, builds). Two kinds of cues: literal sounds the camera shows, and designed sounds the audience only feels (whooshes, risers, tension beds).
2. Source or record. For each cue, choose the fastest legitimate route:

• Library SFX: search references/sfx-libraries.md for the licence fit. Prefer CC0 or commercial-cleared sources for anything commercial. Keep an AUDIO_CREDITS file with file name, author, source URL, licence.
• DIY foley: record it yourself with a phone or handheld, close-mic'd in a quiet room. The foley field guide (references/foley-field-guide.md) lists the prop per sound and the recording rules. One cohesive movement pass (footsteps + cloth + prop bumps together) beats a stack of isolated one-shots.
• AI generation: acceptable for designed sounds (risers, whooshes, textures) where no library file fits. Log it as generated in AUDIO_CREDITS.

1. Build the ambience bed. One continuous bed per scene, not per shot. Layer 2-3 elements (e.g. room tone + street walla + distant traffic) at low levels and crossfade them when the location changes. For a repeating bed, turn any long recording into a seamless loop: bin/make_loop.py -i source.wav -o bed_loop.wav --length 30 --overlap 2 --start 10. Ambience sits in the background: it should be felt, not heard. Rule of thumb: if the audience can name the bed, it is too loud.
2. Cut and layer. Lay dialogue and foley first, ambience under those, designed SFX last. Layer designed hits from 2-3 parts: transient (the attack that sells sync), body (the texture), tail (the decay that sells space). Match sound perspective to the shot: close shot = dry and detailed, wide shot = quieter with some room reverb. Cut designed sounds to the frame; the attack of a sync hit should land on the visual impact within 1-2 frames or the scene feels wrong.
3. Sit everything under dialogue. Dialogue owns the mix. Reference targets from the sibling audio-mixing skill: ambience around -30 to -35 dB (felt, not heard), foley around -18 to -20 dB below dialogue, designed SFX a step below dialogue but above music. When effects muddy speech, do not just turn them down: high-pass foley and ambience to clear rumble under the voice, and carve a small dip in their 2-5 kHz range where consonants live. Then hand the stems to audio-mixing for the dialogue chain and final loudness normalization.
4. Check on small speakers. Listen to the full cut on a phone speaker. Dialogue must stay intelligible, beds must not rumble, sync hits must still read. Fix balance there, not on studio monitors.

Output Contract

• A cue sheet (scene -> cues -> source/foley file used).
• All element files: named consistently (s03_foley_footsteps_gravel.wav), 48 kHz, peaks at or below -6 dBFS so the mix stage has headroom.
• AUDIO_CREDITS.md: every non-recorded file with author, source URL, and licence.
• Stems or a session ready for the audio-mixing skill: dialogue, foley, ambience, designed SFX on separate tracks.

Operating Rules

1. Never invent licences, attributions, or library facts. Every source file gets a verified licence; if it cannot be verified, it does not ship.
2. Ambience is felt, not heard. If the audience notices the bed, pull it down.
3. Sync hits land within 1-2 frames of the visual impact or the scene reads as fake.
4. Effects never fight dialogue for 2-5 kHz. Carve space with EQ before reaching for the fader.
5. When the scene changes, state what ends as well as what begins. No threat sounds leaking into a safe location.
6. Loudnorm is the final process and belongs to the audio-mixing skill. This skill delivers premixed stems, never the mastered file.