The VS Code of video editing, built for agents.
Example: long-form talking head
A single speaker to camera — an explainer, a podcast solo, a course lesson. The recording is long and raw: restarts, flubbed lines, and dead air between thoughts. The edit cuts it down to the clean story, back-to-back, never mid-word. The transcript drives the A-roll; the waveform separates silence from performance.
1. Brief
Write the brief first. It records the source files, the target duration (or "keep everything good"), and any editorial instruction. Absent a target length or explicit direction, make no editorial cuts: keep every good part, drop only silences and double takes, and keep the story chronological.
2. Analysis
The transcript is the spine, so transcribe everything with speech.
- Transcribe every video and any external audio.
transcribewrites word-level start/end times to a JSON file — the times you cut on. Run it on each camera take and on any separate recording (a lav or interface track you will sync to). - Render the waveform to read the gaps.
waveformmarks silence in red and returns it as second ranges. A transcript marks only speech, so a gap in it can be silence or something to keep — a laugh, a breath. The waveform tells them apart: red is dead air to cut; a gap that still shows signal is performance to keep. - Resolve double takes visually. The transcript shows repeated lines; usually the last take is the keeper. When takes read equally well,
grabthe first word of each and keep the sharpest — motion blur means a bad take.
3. Lay out the A-roll
Drop every double take and silence, and put the good parts in chronological order.
import { For } from "solid-js";
const raw = "/Recordings/take.mp4";
// Kept stretches, in order — source range only.
const takes = [
{ src: raw, sourceIn: 48.2, sourceOut: 53.9 }, // opener — third take, sourceIn on the first word
{ src: raw, sourceIn: 61.7, sourceOut: 69.4 }, // next sentence — sourceIn opens ~0.4 s early for a breath
{ src: raw, sourceIn: 74.0, sourceOut: 78.1 }, // mid-sentence pickup — flubbed clause dropped
{ src: raw, sourceIn: 95.2, sourceOut: 102.8 }, // closer — trailing air left in sourceOut
];
// Chain them back-to-back: each start is the sum of the durations before it.
let cursor = 0;
const aRoll = takes.map((t) => {
const start = cursor;
cursor += t.sourceOut - t.sourceIn;
return { ...t, start };
});
export default function Project() {
return (
<stage camera={[0.25, 0, 0, 0.25, 235, 70]}>
<scene name="Talking head" width={1080} height={1920} fill="black" active>
<sequence name="A-roll">
<For each={aRoll}>
{(clip) => (
<video src={clip.src} width={1080} height={1920}
start={clip.start} sourceIn={clip.sourceIn} sourceOut={clip.sourceOut} />
)}
</For>
</sequence>
</scene>
</stage>
);
}
The silence you keep between sentences lives inside the clips, not in a timeline gap — a gap in a sequence freezes or blanks the frame. Carry a pause by extending a clip's sourceOut (or opening the next sourceIn) into the real silence, and keep the timeline gapless.
Save the file — the app recompiles and re-renders it — and get the spine right before layering anything on top.
4. Cut points
- Open tight. Set the first
sourceInon the first word — dead air at the top kills retention. - Between sentences, leave 0.3–0.5 s of silence. Enough to breathe, not to drag. Tune it with each clip's
sourceOut/ nextsourceIn. - Mid-sentence, butt the words together. When you drop a clause inside a sentence, close the gap as far as it goes — never mid-word. Grab the frames around the join to confirm the word is whole.
- End on up to 0.8 s. Leave a little trailing air in the last
sourceOutso the video doesn't stop abruptly.
5. Verification
Run check on the scene first: on a long timeline of many clips, a sourceOut/start that don't meet leaves a gap no sampled frame would land on, and the check names those spans outright.
Then the cut points, which matter most. Capture the first frame of every clip with capture at each clip's start, and check the cut didn't land on a bad frame — motion blur is the usual tell; nudge sourceIn a few frames to a settled one. Reconcile against the brief, and re-check neighbouring cuts after any structural change.
6. Visual hook (optional)
A hook makes the video readable at a glance. Aim for a frame you could lift out and use as a thumbnail.
- Just a title. One line of large Inter at medium weight, in a single colour. That is the whole hook: no sub-header, eyebrow, or secondary line, and no background, box, panel, or fill — the text sits directly on the A-roll.
- Placement. Beside the speaker's head or below it, never covering the head. A popular strategy is to place it A-roll aware.
- Duration. Keep it on screen as long as the viewer needs to read it.
One line of copy is a job for the native <text> tag:
<text
x={80} y={1240} width={920} height={260}
start={0} end={3}
fontFamily="Inter" fontSize={84} fontWeight={500} color="#fff"
textAlign="left" textBaseline="top"
>
The one editing trick
</text>
7. Captions (optional)
Add captions last, after every cut and overlay is verified. src names the recording's subtitle file (.srt, .vtt, or a transcript .json). Use the whisper preset aligned to the bottom, and start it where the hook ends — set both start and sourceIn to that timestamp so the transcript stays aligned (see reference/jsx/captions.md in the project's docs). Keep captions off overlays and off the important part of the A-roll.
{/* Hook holds until 00:03; captions begin there. */}
<captions src="talk.srt" preset="whisper" verticalAlign="bottom" start={3} sourceIn={3} />