agentleFS
OpenRootdevOpenRootdev/sfx-tool/skills/sound-design/SKILL.md

sound-design

Skill0 starsChanged 3 days agoMITView on GitHub

Sound design and music for videos made by creative agents - what makes a video sound professional, and how to get there in launch films, explainers, tutorials, demos, social clips, showcases, trailers, 3D and motion graphics. Covers matching sound to the picture, choosing and cutting music, hooks, silence and contrast, filter sweeps, risers, hits, layering, ducking and voice pockets, keeping SFX in the music's key and tempo, stems and loudness. Includes a plan-to-mix renderer (sfx-tool mix), a music analyser (sfx-tool analyse) and a video spotter (sfx-tool spot) for scoring a finished video or learning from a reference video. Use whenever a video (HyperFrames, Remotion, ffmpeg, any editor) needs music, sound effects or a final mix, when a finished video needs better sound, when audio feels flat, crowded or cheap, or when the user asks for sound design, a soundtrack, SFX timing or "make it feel professional", even if they don't say "sound design".

13 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 5 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 11 edited versions we read keep these.

  • 5

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Intention: Every sound is there for a reason the viewer would feel: to make a moment land, to set a place, to carry a transition, to add …

    line 29
  2. Distinctive:

    Musicality: When there is music, sound effects agree with it: hits on the beat grid (or deliberately off it), tonal sounds in the music's …

    line 45
  3. Distinctive:

    Keep these two apart. Facts are worth following every time.

    line 52
  4. Distinctive:

    Do the sounds come from one world, and does each fit what it scores?

    line 156
  5. Distinctive:

    If there is music, are hits on its grid and tonal sounds in its key?

    line 157

“Common” is measured: kept by at least 6 in 10 of the 11 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Sound design for agent-made video

Sound is the other half of a video. Picture tells people what is happening; sound tells them how to feel about it, where to look and whether to keep watching. People rarely notice good sound, but they notice right away when it is missing or cheap.

This skill describes what professional sound has and the techniques that get you there, with when each one helps and when it gets in the way. None of the techniques is required. A quiet tutorial and a loud launch film share the same qualities but use almost none of the same moves. Decide from the video in front of you, and be able to say why each sound is there.

You are usually the creative agent making the whole video, so picture and sound can be designed together: the music can shape the animation as much as the animation shapes the sound. When the picture is already finished (the user's own edit, a screen recording, a render you cannot change), the sound fits around it instead: references/finished-video.md.

You cannot hear. You can measure, plan precisely and tell the user exactly where to listen. Say "measured", not "sounds great", and give timestamps for the moments that need a human ear.

What professional sound has

These hold across genres. Use them to judge your own mix.

  1. Intention. Every sound is there for a reason the viewer would feel: to make a moment land, to set a place, to carry a transition, to add weight. A sound with no reason is noise, however nice it is.
  2. Contrast over time. Loud means something only after quiet; full only after thin. Amateur mixes sit at one level from start to end. Professional ones have a shape: an opening, tension, a release, an ending. The shape can follow the story's cause and effect: thin while something is set up or asked for, full when the result arrives. Short held breaths before big moments are one of the strongest tools there is (the brain expects a beat, and the hit that follows resolves it).
  3. Clarity. People can follow about two separate sound streams at once (Walter Murch's "two and a half"). The most important thing at each moment (usually a voice, otherwise the hero moment) must be clearly audible, and everything else makes room for it.
  4. Coherence. The sounds come from one world: the same space (all close and dry, or all big and roomy), the same material language, and a small set of sounds reused and varied rather than a new library sound for each moment. Mismatched sounds are what make a video feel assembled rather than designed.
  5. Belonging to the picture. A sound tells the viewer what a thing is made of, how big it is and how it moves: glass sounds glassy, big things are low and long, fast things are bright and short, and a sound lasts as long as the motion it scores. When the brand or hero has a signature sound, its return pays off.
  6. Musicality. When there is music, sound effects agree with it: hits on the beat grid (or deliberately off it), tonal sounds in the music's key, cuts on bar lines, an ending that lands on a downbeat and rings out.
  7. Clean delivery. The right loudness for the platform, no clipping, no clicks, no accidental silence, nothing chopped mid-phrase.

Facts and taste

Keep these two apart. Facts are worth following every time; taste depends on the video.

Facts (measurable, and breaking them sounds broken):

  • Online platforms normalise to about −14 LUFS integrated; keep true peak at or below −1 dBTP. Broadcast uses −23 LUFS (EBU R128). references/mixing.md
  • Sound early by more than ~45 ms looks wrong; up to ~100 ms late is tolerated. Put hits on the contact frame or one frame late.
  • Speech must stay intelligible; anything competing with it in 2–4 kHz, or within ~10 dB of it, hurts.
  • Music needs a licence that covers the use.

Taste (every number in the references is a starting point): the genre, the palette, how many sounds, how much reverb, how deep a filter sweep goes, whether there is music at all. Derive these from the brief, brand and audience each time. Don't carry a favourite style from one video into the next.

Silence is taste too, with conditions. A near-silence in the middle of a video works when the picture explains it (a held frame, a pause before a reveal, a question before its answer), when it is short (around a second or two), and when it ends on the beat with something that lands. Silence nobody chose reads as a broken player. List the ones you mean in the plan's gaps so the mix report knows.

What usually matters, by kind of video

VideoUsually matters mostUsually leave out
Tutorial, talking head, narrationvoice clarity, music ducked low and steady, a few soft UI soundsbig hits, risers, sweeps under speech
Calm explainer, brand storyone warm palette, music with space in it, gentle transitions, a clear endinghard impacts, busy drums
Product launch, feature filmmusic-first edit, a build and a release, a signature sound for the product, a confident endingsounds on every cut, wall-to-wall energy
UI or product demosmall, tight interface sounds matched to each interaction, quiet bedcinematic whooshes, sound on every pixel moving
Showcase, product reel, motion graphics (music-led, no voice)a hook that repeats and sticks, a tempo that suits the pace of the picture, the arrangement following the story (thin while setting up, full on the results), one or two breaks the picture explainsa sound on every visual move, a busy melody fighting the picture
Social short (5–30 s)an immediate hook in the first second, tight sync to the beat, a loop-friendly or punchy endlong intros, slow builds
Trailer, game, 3D or motion pieceworld-building ambience, a strong dynamic arc, silence before the reveal, physical weightconstant layering that flattens the arc
Data or chart videosounds that track the data (rising pitch with a rising line), restraintsounds that compete with the numbers

Many videos sit between rows; take what fits. Some need no sound effects at all, only music shaped well.

The toolbox

What each move does to the viewer, so you can decide whether you need it. Numbers and recipes are in references/filters-and-moves.md.

MoveWhat it doesHelps whenGets in the way when
Low-pass on the musicdistance, "behind glass", anticipationintros, flashbacks, a subject not yet arrivedthe music needs to be clear
Rising high-passthins the music, builds tensionthe bars before a drop or revealthere is no release after it
A short gap (held breath)makes the next hit landright before the biggest moment, once or twice, when the picture explains itused often, or nothing on screen gives a reason
Riser, swell, reversepulls the viewer into a momentleading into a cut or reveal, ending exactly on itit ends early or late, or every transition has one
Hit, sub, impactweight and finalitythe reveal, the logo, the last frameeverything is an impact
Layering by frequencyone event that sounds big: a low layer for weight, a body, a bright top for detailbig moves, reveals, a logo landingsmall UI moments, or every sound is huge
Duckingmakes room for what mattersunder voice, under a hero soundit pumps audibly with no reason
Voice pocket (band ducking)the music dips only where the voice lives (about 1–4 kHz) and stays full around itnarration over full musicno voice, or the music is already sparse
Arrangement as storythe music thins and fills with cause and effecta setup and its payoff, an input and its resulta calm piece that wants one steady bed
A hook (a short riff that repeats)makes a piece memorable and lets the picture move freely over a steady grooveshowcases, motion graphics, socialunder narration, or in calm and serious pieces
Getting into and out of a songno jolt where music starts, ends or changesstarts, endings, a second song, the return after a break(rarely)
A metaphor sound in place of the literal onesays how something feels rather than what it isa story with a message, a mood piecedemos and tutorials, where the literal sound is information
Reverb and spaceputs everything in one placeglueing a palette, ending tailsit blurs fast sounds or speech
Signature sound (leitmotif)identity, payoffa brand or hero that returnsthere is nothing to return to
Foley, UI and typing soundsmakes on-screen actions physicalinteractions, typing, objects touchingthey cover the music or the voice
Stemsuse only the parts of a song you needunder voice, intros, endings(rarely)

How to match sounds to what is on screen (material, size, colour, motion, transformations, counterpoint): references/creative-matching.md. Choosing, cutting and keying music: references/music.md.

A way of working

This order works well; change it when the video calls for it. Details: references/workflow.md.

  1. Sonic identity. One line from this brief: palette, music direction, tempo and key, e.g. "warm felt piano and paper foley, 84 BPM, D major" for a calm explainer. Then a sound map: the video's key elements and the sound each one gets.
  2. Music first or picture first. Edits cut to music (launches, montages, social) work best with the track chosen first: sfx-tool analyse music.mp3 --fps 30 --beats-out beats.json gives tempo, beat grid and key to time the picture to. When a voice or footage sets the timing, lock that first and fit the music to it. When the picture is finished, take its moments from its source if you made it (exact, small changes included); without source, sfx-tool spot video.mp4 lists its cuts, moves and holds to score against. Lock the few structural moments (the first result, the return after a break, the ending) to bar lines; the rest of the motion can move freely over the groove.
  3. Spot it. A cue sheet: each moment, what the viewer should feel, and the move (if any) that does it. Most moments need nothing; usually 3–6 matter.
  4. Place sounds on the final picture, shaped like the motion they belong to, then write the plan and render: sfx-tool mix plan.json -o mix.wav --report report.json (references/plan-format.md).
  5. Measure, mux and hand over with a listen list: 3–5 timestamps and what to listen for.

Tools

sfx-tool (Node 18+, no other dependencies): npx github:OpenRootdev/sfx-tool <command>, or node bin/sfx-tool.js <command> in a clone. Any other DAW, library or tool works with the same ideas.

NeedUse
Tempo, beat grid, key, notes for SFX, energy per secondsfx-tool analyse <music> (estimates: half/double tempo is possible; relative major/minor share notes)
Small tuned SFX: taps, thocks, knocks, pops, chimessfx-tool new / Studio (sfx-tool studio --open); the sfx-tool skill
Typing / text-appear trackstyping clips in the plan, or sfx-tool track
A finished or reference video: cuts, moves, holds; its soundtrack's bands, breaks, tempo, groove; how they line upsfx-tool spot video.mp4 --sheet sheet.png --events-out events.json (needs ffmpeg)
Mix: clips, automation, filters, ducking and voice pockets, reverb, swells, loudnesssfx-tool mix plan.json -o mix.wav
Stems (drums, bass, vocals, other) from any songDemucs, local and free: python -m demucs -n htdemucs song.mp3
Music bed when none is supplieda licensed library track, the user's own, or local generation (MusicGen); check the licence
Mux and final loudness checkffmpeg -i video.mp4 -i mix.wav -map 0:v -map 1:a -c:v copy -c:a aac -b:a 256k out.mp4; -af ebur128=peak=true -f null -

Before you hand it over

Ask yourself:

  • Can I say why each sound is there? Does anything play just because it could?
  • Does the soundtrack have a shape: an opening, a peak, an ending that resolves?
  • At every moment, is the most important thing clearly audible?
  • Do the sounds come from one world, and does each fit what it scores?
  • If there is music, are hits on its grid and tonal sounds in its key?
  • Facts: loudness within 1 LU of the target, true peak ≤ −1 dBTP, no silence nobody chose (intended ones listed in gaps), report warnings fixed or explained, voice never masked.
  • The listen list for the user, and which calls are taste they should judge (is the beat too busy, does the palette fit the brand).

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.