background-music
Lays a background music bed UNDER the voice on a near-final cut — a FLAT constant bed by default (no ducking, no fade-in, −18 dB, short tail fade-out only). Ducking + fade-in are opt-in. OPTIONAL, for
SKILL.md.claude/skills/background-music/SKILL.md
GitHubVideo Studio: this skill lives in
engines/legacy-editor. Before running anything from it,cd engines/legacy-editorand use that folder's copy (engines/legacy-editor/.claude/skills/background-music/), which has its scripts and reference files. Every path below is relative toengines/legacy-editor. This copy only lets Claude find the skill from the Video Studio folder.
Background Music — Flat Bed Under the Voice
Optional step 6. Takes a near-final cut and lays a music bed under the voice.
The default is a FLAT bed: the music sits at one constant level for the whole video — no sidechain ducking, no fade-in — with only a short tail fade-out at the end. The voice arrives peak-limited at −6 dBFS (≈ −21 dB mean) from the static rough-cut chain, so a quiet fixed bed (−18 dB) rides cleanly under it. A peak limiter guards the sum; there is no loudnorm on the flat path (single-pass loudnorm on a finished mix pumps and re-introduces the very level variation we're avoiding).
Why flat is the house default: ducking (music dipping under the voice and swelling back up in the gaps) made past mixes get audibly louder and quieter in parts. A flat bed has a measured loudness range of 0.0 LU — it does not move. That's the default. Ducking and fade-in are opt-in only.
Most useful on short-form explainers, but works on any format. It's opt-in — skip it unless asked.
When to run
After Captions (5), before Final render (7). Long-form skips captions, so for long-form it runs right after the graphics/second-pass cut. Always operate on the latest rendered cut, never on raw.
Where the music comes from
Drop a track into projects/<job>/audio/. Two ways to get one:
- Your own / licensed track — drop an
.mp3/.wav/.m4aintoaudio/. Use the newest file there. - Generate one (royalty-free, no copyright strikes — on brand) via the vendored
media-useskill:node .claude/skills/media-use/scripts/resolve.mjs --type bgm --intent "<mood>" --project projects/<job>(run its sign-in Preflight first permedia-use/audio/references/bgm.md; there is NOnpx hyperframes bgmcommand). Prompt the mood to match the reel, write the result intoaudio/.
If audio/ is empty and no track is pointed at, ask one line: "Drop a music track in audio/ or want me to generate one — what vibe?"
Run it
.claude/skills/background-music/scripts/mix-music.sh \
<near-final-video.mp4> \
projects/<job>/audio/<track>.mp3 \
projects/<job>/outputs/<job>-music.mp4 \
[bed_db] [duck] [fadein]
- Output is non-destructive —
<job>-music.mp4, never clobbering the no-music cut. That music-mixed file is what Final render (7) exports as the deliverable. bed_db— music gain. Default −18 (quiet, clearly background). More negative = quieter:-24≈ barely-there,-16≈ noticeable,-12≈ prominent. Stay negative — this is a bed, not a duet. Balance is taste; re-run with a differentbed_dbto tune.duck—off(default) = FLAT constant bed, no ducking.on= opt-in sidechain auto-duck (music dips under voice, swells in the gaps) and re-normalizes the whole mix to −14 LUFS. Only turn it on when you specifically want the bed to breathe with the talking.fadein— music fade-in seconds at the start. Default0= no fade-in (the bed is just there). Pass e.g.2for a 2-second fade-in. Opt-in.
Default (flat) is just three args past the output — no flags needed:
mix-music.sh cut.mp4 audio/track.mp3 out-music.mp4
Tuning notes
- The script loops the track to cover the full video and fades the tail out over the last ~1.2s (always on — a clean ending, not a level move).
- Voice intelligibility is the priority. If the bed feels too present, drop
bed_dbfurther (e.g. −22). If it's buried and you want it more felt, raise toward −16. Keep it flat — don't reach for ducking just to make it louder; lower the bed or raise it as a whole. - Don't default to ducking. It's there for the rare case you want the bed to swell in the gaps. The flat bed is the house sound.
- Verify after: play it, or
ffprobe/ebur128the output. On the flat path the bed sits at a constant level (the mix's loudness range tracks the voice, not the music); on the duck path the mix is re-normalized to −14 LUFS.