Generating clips is the easy part

Generating a good four-second clip stopped being hard some time ago. Type a prompt, wait, get something usable — and if not, roll again.

Turning eight of those clips into two minutes somebody watches to the end is a completely different job, and it has not got easier at all. It is the same job it was twenty years ago: cutting.

The round trip is the real cost

Here is the workflow most people end up with, because it is what the tools encourage.

Generate in a browser tab. Download. Open an editor. Import. Arrange. Discover that shot three does not cut against shot four — wrong pace, wrong energy, the motion ends badly. Go back to the tab. Regenerate. Download. Import. Find it in a folder full of files called video_final_2.mp4. Drop it in. Re-trim it, because the new clip is a different length. Re-sync everything after it.

Now do that eleven more times.

The per-clip generation price is what everyone compares. This round trip is what actually costs you the afternoon, and it is invisible until you are in it.

The fix is not a better prompt. It is that the generator and the timeline should be the same window. When they are, fixing shot three is a re-roll in place: the clip is replaced where it sits, and the captions, the audio and every cut after it stay exactly where they were.

Four things no model release will fix

Pacing. How long a shot holds, when you cut, where the video breathes. No generator has an opinion about this, and it is most of what separates watchable from not. It is trimming, and trimming happens on a timeline.

Continuity of edit. Two individually beautiful shots that do not cut together are still two shots that do not cut together. Fixing it means trimming one, reordering, or bridging with a transition — all edit decisions.

Audio. Generated video is silent or comes with sound you did not choose. Voiceover, music underneath, and ducking so the words stay legible is most of the perceived quality of a finished video, and none of it is generation.

The 80%-right clip. The single most common output. A generator's only answer is to roll again — which loses the take you liked, because generation is not deterministic. An editor's answers are: trim to the good two seconds, interpolate to smooth it, upscale it, cut around the bad moment, put a title over it. Four of those five keep the take.

What "generate in the timeline" actually changes

Three things, in rough order of how much time they save.

Re-roll one shot, not the video. The clip is a clip on a lane. Replace it and everything downstream holds. This is the whole argument in one sentence.

Generate variants and pick. Because generation is a draw from a distribution, the honest way to use it is to take several draws at once and choose. Firing off a few variants of the same prompt and picking the best is strictly better than rolling one at a time and hoping — and it is only practical when the results land next to the timeline instead of in a downloads folder.

Everything lands in one library. No filenames, no folders, no wondering which of video_final_2.mp4 and video_final_2 (1).mp4 was the good one.

The workflow that comes out of it

Once generation is a timeline operation rather than an errand, the sensible order flips:

1. Block the whole thing out first, with placeholders or rough generations. Get the shape right — how many shots, how long, what order.
2. Lay the audio down early. The narration or the music decides the rhythm, and the visuals get cut to it.
3. Then generate for real, shot by shot, into the slots you already know the length of. You will generate far fewer clips, because you are no longer producing footage speculatively.
4. Fix rather than re-roll wherever the clip is close.
5. Polish last — colour, captions, titles.

That is just editing. It is the order every editor has used forever, and the only thing AI changes is where the footage comes from.

The uncomfortable version

The generation part of AI video is becoming a commodity. A dozen organisations train the models, most products call the same handful of endpoints, and the quality gap between them narrows with every release.

What does not commoditise is the hour after the clip arrives. That hour is editing, it has always been editing, and no model release is coming to do it for you.

Which is a good thing, incidentally. It means the skill is still the same skill — knowing what to cut, when to hold, what the audio is doing — and that transfers to every tool that will exist five years from now. Learning the timeline is worth more than learning any generator's prompt syntax.