demo-assembly

Assembling the video

Remotion owns the finished timeline. ffmpeg only prepares inputs (trimming, cropping, normalising). That split is what makes transitions, captions and overlays on top of the footage possible in a single deterministic render.

node scripts/render/render.mjs --project demo/<slug>             # render
node scripts/render/render.mjs --project demo/<slug> --studio    # open the editor

(demo/<slug> is the video's project folder, e.g. demo/project-overview — one per video.)

The first run copies remotion-template/ into demo/<slug>/remotion/ and installs it. It is never overwritten afterwards, so any edits the user makes there survive.

How props reach the renderer

build-props.mjs turns demo.json into props and does three things the renderer should not have to:

1. Verifies every artifact exists. A missing clip becomes a clear error rather than a black rectangle.
2. Inlines the event log and word timings so components never touch the disk.
3. Reconciles durations. Generated video comes back at whatever length the model chose; if a clip is shorter than its scene, the scene is trimmed to the clip rather than freezing on the last frame.

Media resolves through --public-dir <project>, so staticFile("clips/x.mp4") points straight at the pipeline's output. Nothing is copied.

What the renderer does with the event log

DemoClip reads the capture's event log and eases the frame toward each click just before it lands, holding briefly, then easing out. This is the single biggest difference between "screen recording" and "product video" — it leads the viewer's eye across a 1920px UI.

Zoom is deliberately gentle (1.18×). The footage is already rasterised, so pushing harder makes text mushy and reads as a jump cut. Disable per scene with "zoomToClick": false.

Callouts with "anchor": "lastClick" place themselves next to the most recent click — where the viewer is already looking.

Audio

• Narration is laid out on an absolute timeline, not inside scenes, because transitions overlap scenes and two overlapping voices are instantly audible.
• Presenter clips are muted; the original narration track plays instead. The avatar's baked-in audio is a re-encode of the same thing.
• A music bed ducks under narration automatically (music.duckDb, default −14 dB). Without ducking it sounds like a stock template.