Canonical short-form-ad audio mix in one FFmpeg pass. VO loudnorm + 3.0× per-clip + 2.0× mix, music 0.13 base + apad+afade, sidechain compress 20:1 @ 0.01, climax line +20%, optional video-to-music duration sync. Replaces the reactive multi-round tuning that cost v03 6+ passes.
Install gooseworks: https://gooseworks.ai/creative-agent/install.md # then, in Claude Code, Cursor, or Codex: /gooseworks use the mix-master skill
Superseded: the video kit now does this with the audio-mix part, version 1.0.1 and the sound-layer part, version 1.0.4, in the parts folder of this repository. Its mix mode is the audio-mix step (voice, a ducked bed and timed effects) and its finish mode is the sound layer, which levels every video to -14 LUFS with the true peak at or below -1 dBTP. This atom stays, unchanged in behaviour, for skills outside the kit until they move; its scripts still run.
The v03 Ironman LEARNINGS describe 6+ tuning rounds because no default mix protocol existed: round 1 music too loud, round 2 VO too quiet, round 3 Whisper mis-transcribed, round 4 sidechain ratio finally right, round 5 music ended too early, round 6 climax sat at the same loudness as the rest.
This atom encodes the protocol that landed and ships it as the default. Operators tune by flag; they don't redesign the chain.
--video <path> — video file with no audio (or whose audio gets dropped) (required)--vo <path,path,...> — comma-separated list of per-scene VO mp3s in scene order (required)--vo-starts <ms,ms,...> — start time in milliseconds for each VO clip on the timeline (required, same length as --vo)--music <path> — music bed file (required)--sfx <path,path,...> — optional comma-separated SFX list--sfx-starts <ms,ms,...> — optional SFX start times--output <path> — output mp4 (required)--total-duration <s> — target total duration in seconds (required — video AND music are sync-fit to this)--vo-boost-line <N> — 1-indexed VO clip that is the climax; gets +20% additional volume (default unset)--vo-volume <float> — per-clip VO volume multiplier (default 3.0)--vo-mix-volume <float> — extra mix-bus volume on VO (default 2.0)--music-volume <float> — base music volume (default 0.13)--music-swell-volume <float> — peak music volume during apad/swell (default 0.21)--sidechain-ratio <float> — sidechain ratio, capped at 20 by FFmpeg (default 20)--sidechain-threshold <float> — sidechain threshold (default 0.01)--target-i <lufs> — VO loudnorm integrated target (default -16)--target-tp <dbtp> — VO loudnorm true-peak (default -1.5)--target-lra <lu> — VO loudnorm LRA (default 11)--vo and --vo-starts must be same length; --sfx and --sfx-starts must be same length; ffmpeg + ffprobe must be on PATH.ffprobe). If music shorter than --total-duration, append apad=pad_dur=2.0 then trim back to target. If music longer, trim with afade=out over last 1.5s.filter_complex_script file (long chains exceed shell arg limits):
[N:a]loudnorm=I=...:TP=...:LRA=...,adelay=<start>|<start>,volume=<vo_volume>[voN];--vo-boost-line N): volume becomes vo_volume * 1.2.amix=inputs=K:duration=longest,volume=<vo_mix_volume>[vo_pre];[vo_pre]asplit=2[vo_final][vo_sc];apad=pad_dur=2.0,atrim=0:<TOTAL>,afade=t=out:st=<TOTAL-1.5>:d=1.5,volume=<music_volume>[music_base];[music_base][vo_sc]sidechaincompress=threshold=<thresh>:ratio=<ratio>:attack=20:release=600[music_ducked];adelay + volume then mix into [sfx_bus].amix=inputs=<2 or 3>:duration=longest[a_out].-c:v copy).[a_out], write manifest.json and verification.md.<output> — mixed mp4 with VO + music + SFX.manifest.json — chain parameters, vo timing, climax line, integrated LUFS after.verification.md — pass/fail report (integrated LUFS within ±0.5 of expected target, no clipping).<target-i> (post-protocol; should be approximately -14 social-ready after the boosts).<target-tp> (no clipping).<total-duration> ± 100ms.afade=out — last 100ms RMS < -30 dB.--vo-boost-line was set, the corresponding clip's volume= value is vo_volume * 1.2.apad insufficient — emit a warning; video may end before music. Operator should re-source music or shorten --total-duration.create-clips Phase 4 catches it.-c:v copy to preserve quality.scripts/mix.sh \
--video acme/ads/video-01/edits/master-no-audio.mp4 \
--vo "acme/ads/video-01/audio/vo-scene-01.mp3,...,acme/ads/video-01/audio/vo-scene-14.mp3" \
--vo-starts "0,2000,3500,...,28000" \
--music acme/ads/video-01/audio/music-cue-01.wav \
--output acme/ads/video-01/edits/master-final.mp4 \
--total-duration 30.0 \
--vo-boost-line 14Consumed by: video-orchestrator/edit-video Phase 3–5, auto-fix-from-review-notes dispatcher (when vo-intelligibility or music-vo-balance polish-notes items reference mix-master).
--video <path> — video file with no audio (or whose audio gets dropped) (required)--vo <path,path,...> — comma-separated list of per-scene VO mp3s in scene order (required)--vo-starts <ms,ms,...> — start time in milliseconds for each VO clip on the timeline (required, same length as --vo)--music <path> — music bed file (required)--sfx <path,path,...> — optional comma-separated SFX listResearch supported advertising promises from a specific buyer's context, current product facts, customer language and observed ads. Turn them into static, video or copy ideas; match video ideas to actual recipes and save a product-scoped angle bank for write-video-ad-script. Product-led discovery can use verified facts without reviews. Reference persistence and reach are observations, not proof of performance.
Analyze ad campaign performance data (Google, Meta, LinkedIn) to identify what's working, what's wasting budget, and specific cut/scale/test recommendations. Runs statistical analysis, funnel diagnostics, and multi-channel budget reallocation with specific dollar-amount shift recommendations and scenario modeling.
For paid lead-gen and participant-recruitment ads, replaces vanity CPA with true CAC per qualified lead by joining ad-platform data with downstream funnel events, surfaces tracking gaps, and classifies every creative into Scale / Keep / Investigate / Cut.