Video in. Spotted sound out.
Input — the brief
A clip and one line.
Drag in the video — or skip it and work from a brief alone. One sentence says what the piece is: “A huge sea monster emerges from the ocean, epic and massive.” Three services take it from there: Gemini watches, DeepSeek thinks, ElevenLabs makes the sound.
No guide track. The picture is the guide.
Watch — the eyes
Every half-second, the picture is read and written down.
The video is paused twice a second and described: “water bulges”, “creature breaks surface”. Every note carries an exact timecode (MM:SS:FF) and becomes the PICTURE EVENTS list the whole job hangs off — plus a scene-by-scene shot list for context. If the eyes fail for any reason, the run quietly continues from the brief alone.
Anything faster than half a second — a footstep — slips between the pauses. Feet get their own engine later.
The supervisor
One quick pass decides who works and what each team aims for.
Atmos always runs. Spots always runs when there's a picture. Design is only hired when the piece calls for invented sounds — a dry, realistic brief gets no design.
The three teams write the plan
The heart of it — each team writes its part of the list.
Atmos
The air · the room · the place
Two to four long beds. Never tied to a moment — they sit under everything.
Spots
The picture's moments
Groups the frame notes into real physical events. One moment, one timecode, its layers: impact, vibration and splinter are three layers of one event, locked together so they can't drift. Continuous things — a helicopter rotor, rain, water running off a body — become long beds, never trimmed to fit a moment.
Design
The only team allowed to invent
Stingers, transitions, creature vocals — hired only when the supervisor says the cut earns it.
Three checks, every run
The plan is corrected before it costs a single generation.
Vocabulary
Every description is rewritten in the professional language of a 173,000-sound library — so “bubbles” comes out as “deep boiling churn of huge air pockets, massive water displacement.”
Scale
If the subject is big, its own sounds must be written big: mass words are required, smallness words (fizz, thin, gentle) and machine-hiss words are banned from breath and vocals. Failures are rewritten automatically; anything still wrong wears a red warning. The world around the subject keeps its own dynamics — a calm ocean before the monster stays calm.
Housekeeping
Self-described continuous sounds become beds. Duplicates collapse into one. An unspotted one-shot is dropped rather than lied about. Design can't re-cover what Spots already serves.
Generation — ElevenLabs
Each description and its length becomes a real clip.
“Starts immediately, no fade-in” rides on every prompt, so the attack sits at the front. Clips are made several at a time; any failure is recorded against its sound and shown as a red pill — a silent failure can't hide.
The judge — the careful pass
Keep. Move. Drop.
A more careful mind reads the brief, the picture and every moment, then votes on each one. Moves fix timecodes. Drops remove what doesn't serve. If it can't answer, it does nothing — it never deletes on a hunch.
It's a text critic, not an ear — its verdicts are about the plan. The human plays the audio.
Delivery
Playable pills, re-rolls per sound, and exports a mixer can use.
Hover any pill to see its exact prompt and the library references that shaped it. Re-roll one sound at a time. Export numbered folders of mp3s — or spotted stems: 48k WAVs, each trimmed so its attack lands exactly on its frame, with a cue sheet (timecode · sound · duration · why).