What it can do, and what it still gets wrong

Before writing a single prompt, you need to know what generated video gets right and what it gets wrong. Without that, you burn thirty generations on a shot that will never see the light of day, and you think you're doing it wrong when nobody can pull it off.

The first thing to understand: you don't count in minutes

The tools produce shots of a few seconds. Not films. A one-minute video is fifteen to twenty-five separately generated shots, then assembled. This isn't a passing limitation we'll soon work around: the longer a shot runs, the more chances the model has to lose the thread, and the cost rises with duration.

So the question to ask yourself at the start is never "how long will my video be", but "how many shots do I need to produce". That number is what tells you whether the project fits in one evening or takes a week.

What comes out very well, right from the first try

• Atmospheres: landscapes, sky, fog, city at night, underwater.
• Materials: smoke, rippling liquid, fabric, particles, fire.
• Slow movements: a camera gliding forward, a scene that breathes.
• Isolated objects: a product spinning, a close-up detail.
• The abstract: shapes, transitions, animated backgrounds.

What still fails, and is best avoided

• Text written in the image: on most tools, letters warp as soon as anything moves. Text gets added in editing, on top. Mind you, that's no longer true everywhere: see the box below.
• Hands doing something precise: grasping, writing, counting. Still hands pass, hands in action twist. Here too, the gap between tools has become enormous.
• The same person from one shot to the next: this is the hardest problem, and it has its own full lesson.
• Chained physics: pouring liquid into a glass, breaking an object, a bouncing ball. The start is fine, the rest invents.
• Crowds: faces in the background melt into one another.

What "it fails" means in 2026: it depends on the tool

This list long held true for all generators. That's no longer the case, and the gap between two tools is now bigger than the gap between two prompts.

A test run on 8 August 2026, the same prompt sent to three models: a close-up of a hand writing a sentence with a pen. It's the textbook case of a shot that fails.

• The first produced illegible scribbles, and the pen changed shape mid-shot.
• The second wrote two roughly readable words, but the hand wasn't progressing: the text was already there and wasn't moving forward.
• The third wrote "to be continued…" in cursive, correctly spelled, the letters forming under the nib, the tip exactly at the end of the word in every frame.

Practical takeaway, and it matters more than the list: don't decide a shot is impossible, decide it's impossible with the tool you have on hand. Before giving up on an idea, try it on a second model. What was undoable six months ago gets done today, and not by everyone at the same time.

The method that saves the most time

Write the shot list before generating anything. One line per shot, in plain language, without worrying about the prompt: "shot 1, the city seen from above at dawn", "shot 2, the apartment window", "shot 3, the steaming cup".

That list costs ten minutes and changes everything. You can see at a glance how many shots you're about to produce, which ones fall into the "still fails" column, and which you can swap for an idea that works on the first try. Most abandoned projects are abandoned because people started generating before knowing where they were going.

Do this right now: take a video idea you have in mind and write its shot list. Count them. If you go past twenty, cut before you start.