Never use a clip's native audio as the score. Video models generate music per clip, so every cut is a key change. Mute it and build score, beds, and voice separately.
Never stream-copy H3 clips into a concat. Their timebase corrupts the timeline when mixed with re-encodes: 247 seconds of segments became a 454-second file. Re-encode every segment first.
amix quietly renormalizes. It ate the narration. Sum with amerge and pan instead.
Single-pass loudnorm is dynamic. It pumps quiet sections up. Measure, then apply one linear gain.
One giant filtergraph is undebuggable. Sequential file passes let you listen to every intermediate.
A single-frame PNG plus a fade is invisible. Loop it into a stream before the alpha fade, and set an end action or the encode never finishes.
Fal's ElevenLabs SFX endpoint is broken. It hardcodes a deprecated model. Call ElevenLabs directly.
Social previews show frame zero. A fade from black kills click-through. The social cut opens mid-action.
Off-model creatures break the spell. Ewoks drifted into lemurs and gremlins until the prompt described them explicitly and crowds were capped at five.
Restyle passes have hard limits. Gemini's video edit rejects inputs over eight seconds; split with overlap and dissolve, or skip the pass.
Clips are the expensive part. Fix pacing, mix, and titles downstream. A trim is free; a regeneration is two dollars a clip.
