Editor SFX Sparse

Requires

A running Scrollmark editor whose commands you can call: the SocialGPT editor chat, or scrollmark mcp from @scrollmark/cli for a coding agent. Search and import go through the editor's Freesound catalog. Without it, the when-to-use rules and the class table still guide sparse sound design by hand.

When

Load only when something is arriving (B-roll in, a title, a punch they asked to hit) or they asked for sound design. Speech-only recuts and interviews: do not place SFX.

Search and place

media_searchCatalog providerId: "freesound", kind: "audio". Pass sfxRole only for chip roles: hit whoosh riser reverse glitch foley ambience ui. Do not invent roles (pop click downer stinger drone drop explosion impact are not sfxRole). Extra query tokens AND; do not free-type "cinematic dark whoosh".

A sound the user names (mouse click, cork pop, ding, notification, camera shutter) is sfxRole: "ui" with the name as query ("cork pop"): the role adds no word of its own and only returns sounds of 0.3 s or more, long enough to hear over a voice.

Import a remoteId from that search and place it at the timeline second of the event in the same plan: media.importFromCatalog with "as": "$hit", then timeline.addMediaClip with mediaId: "$hit" and "as": "$sfx", then element.setParams with elementId: "$sfx" and params: { "volumePercent": 90 } (80–100; a one-shot under a voice needs 100, and louder is volume in dB). Keep hits short, but not shorter than 0.3 s: a 0.15 s click disappears under speech.

Check that it can be heard

You cannot listen, and a clip on the timeline is not a sound anyone hears. The plan result lists each short sound it placed under sounds, with a verdict (clear, quiet, buried, too short, unmeasured) and its marginDb over what plays under it; audio.measureSounds (through editor_read) measures any of them later. Fix every one that is not clear in the follow-up plan: too short → search again and swap in a longer sound; quiet or buried → make it louder than its file with volume in dB (+6 is about twice as loud, up to +20; volumePercent stops at 100, its file's own level), and if it still is not clear, a punchier sound, or move it a little off the spoken word into a pause. unmeasured means a clip under it could not be mixed (its unmeasured says which): you do not know whether it can be heard, so say that, and do not call it audible. Then say which sounds you changed and what the check reads now.

When the user says they can't hear a sound: measure it first (audio.measureSounds on those clips), tell them what it found ("the click is 0.16 s and 20 dB under your voice"), change what it shows is wrong, and measure again. Never reply that it is fixed after only moving clips or placing the same files again, and never give a cause the tool results do not show.

Do not: one whoosh per keep-range join; SFX on every word caption; a second bed via ambience or drone.

Pinned defaults

When the picture needs this class, search as below and import this remoteId if it appears. If it does not, search again with the extra tokens. Do not invent a URL.

Glitch is not a record scratch. Analog rewind is the reverse pin.

Agent picks (no default file)

Match the picture. Skip if it does not earn it. Tighten later.