one command, five robots, zero hand-editing Β· Synexar Β· July 6, 2026
Seed clean synthetic patients on staging β unique names, tidy schedule, zero real data.
PHI check: zero foreign rows
A Playwright robot drives the real app in a headless browser, recording 1080p while it clicks, types and scrolls.
Every action drops a timestamp mark β the cut map.
footage = reusable asset library
The script file β ElevenLabs narrators (Brian / Jessica). Each line's length becomes its scene's length.
Music beds generated once, reused & extended.
swap voice = 1 env variable
Brand cards = HTML animations rendered frame-by-frame (deterministic, buttery).
Veo 3.1 makes the cinematic shots β corridors, sunrises, porch swings (~$3 per 8s).
light brand, refined lockup
FFmpeg glues it: crossfades, music ducks under the voice, broadcast loudness.
QA = look at pixels (frame grids) + listen at head/mid/tail.
deliver: chat + Azure links
The narration drives. A scene lasts exactly as long as its sentence (footage freezes politely if the voice runs long).
Fixed visual beats drive: 8s cinema shots, 5s product money-shots, kinetic statement cards. The voice is placed into the beats (long lines politely push the next one β never collide).
this exact loop ran ~10 times over two days: demo β 4 demos β unified tour β brand v2.5 β commercials β 90s β 60s final pair
~$31 of Veo shots + pennies of narration.
4 demos, a 6:30 tour (two voices), 6 commercials β 2 finals.
The Jessica 60 was pure configuration: same script, 3 new shots, one config entry β first-try assembly.
Every cut rebuilds from the committed kit + staging. The skill file carries every rule & scar β the next session starts where this one ended.
β drawn by Samantha Β· tap any box to flag it, notes go in the tray β β