From @rajeevkugan
My Claude SFX setup.
Claude can animate anything now. The sound effects it adds? Random, late and way too loud. Every sound in my reels is placed by Claude, and this is the setup that makes it work. Three steps.
Part 1
Three steps
01
Index your own library
Point Claude at your sound folder, not random downloads. It tags every file (whoosh, hit, pop, click, riser) and measures exactly where each sound peaks, in seconds. That number is the whole trick.
Today: put 30 to 50 sounds you like in one folder and have Claude write an index.json with id, type, length and peak time.
02
Write a cue sheet
Claude reads the animation (the cuts, the words, the things that pop in) and writes a cue sheet: one sound per beat, each with a time, a sound id and a volume. Different sound at every cut, extras only on hero moments.
Today: ask for the cue sheet as a table first, check it, then let it build.
03
Sync the peak, not the start
This is the big one. A whoosh builds up before it hits, so starting it on the frame makes it sound late. Start each sound at beat time − peak time so its loudest moment lands on the frame.
Today: add the peak offset to every cue and listen to the difference.
+
Mix to −14 LUFS
Keep the effects under your voice with one master SFX level, so a "too loud" note is a one-number change. Then loudness-normalise the whole mix to −14 LUFS with a limiter, so it doesn't clip or get turned down on Instagram.
Today: run the final file through the ffmpeg line in Part 2.
Part 2
Your starter SFX skill
Save this as SKILL.md in a folder like skills/my-sfx/. In Claude Code that's ~/.claude/skills/my-sfx/SKILL.md. In any other chat, paste it in with your request. Change the paths to your own library.
---
name: my-sfx
description: Add sound effects to my videos from my own library, synced by peak.
---
# My SFX
## Library
- Sounds live in sfx/. Index is sfx/index.json:
{id, file, kind, len, peak} (peak = seconds from
the start of the file to its loudest moment)
- Re-index when I add sounds. Only use indexed sounds.
## Cue sheet (write it first, show me, then build)
- One row per beat: {at, id, vol, why}
- A different sound at every cut. Extras only on hero
moments. No sound on every word.
- About 20 to 25 cues per 30 seconds, max.
## Placement
- Start each sound at at - peak so the PEAK lands
on the beat. Never start a sound on the beat.
## Levels
- One master SFX_GAIN scales every cue (start at 0.3).
- Voice always on top. If an effect fights the voice,
lower it, don't move it.
## Final mix
- Loudness-normalise to -14 LUFS, then limit peaks
to about -1.5 dB. Check the file, not the preview.ffmpeg -i in.mp4 -c:v copy -af "loudnorm=I=-14:TP=-1.5:LRA=7,alimiter=limit=0.84:level=false" -ar 48000 -c:a aac -b:a 256k out.mp4
Want it done for you?
My full setup goes deeper: voice tuning, the motion system, sound design, and automation that tests every reel before it posts. If you want scroll-stopping video for your brand, that's what we do at FROZN Media. DM me @rajeevkugan.
DM on Instagram ↗Skill file shown is a starter. Folder paths and tool names depend on how you run Claude.
