After training hundreds of creators, the same failure point shows up before anyone opens a video model.
Every week, in the masterclass I run for AI Studios, someone shares a shot that took an entire evening to build. The lighting is right. The camera move is smooth. The character does not fall apart mid-motion. And the shot still does not work, because nobody in the room, including the person who made it, can say what happens right before it or right after it.
The biggest mistake I see after two years of teaching AI filmmaking to hundreds of people, through a paid community and corporate training sessions at large French studios, is not a settings problem or a model problem. It is that beginners start generating before they know what story they are telling shot to shot. They open Kling, Runway, or Veo with a mental image of the final scene and no written record of how that scene breaks down. Then they wonder why the pieces do not fit together.
This is not a hot take. It is a pattern I have now watched repeat with enough people, in enough sessions, that I stopped assuming it was a coincidence.
What Is the Actual Mistake, If It’s Not the AI Model?
The mistake is skipping the shot list. Not skipping it because they do not know what a shot list is, most have heard the term, but because AI video generation feels fast enough that planning seems like a waste of the speed they just paid for.
Here is the trap: a single AI-generated clip can now look finished in under a minute. That speed convinces people that filmmaking itself has gotten fast. It has not. Only the rendering step got fast. The thinking step, deciding what each shot needs to show, where the camera sits, what the previous shot handed off, still takes exactly as long as it always did. Skip it, and you are not saving time, you are moving it to the edit, where it costs more.
I made this exact mistake on an early scene for Lost Garden, my AI-animated series. I generated eleven strong individual shots before I sat down to cut them together, then found that four of them could not follow each other. The tell-tale signs of skipped planning showed up right there in the timeline:
- Screen direction flipped between two shots that were supposed to be the same conversation
- The key light jumped sides for no reason a viewer could justify
- A prop that mattered to the scene simply disappeared for one shot and came back in the next
I rebuilt three of the eleven from scratch. The generating had been the cheap part.
Why Does This Happen to Almost Everyone?
A few things keep showing up in my sessions, whether the person is a hobbyist working alone or part of a video team at a company I am training.
The interfaces themselves push people the wrong way: every AI video tool is built to get you from prompt to clip as fast as possible, and none of them ask what that clip is supposed to connect to. Speed looks like progress even when it isn’t.
There is also a habit carried over from working alone. A shot list, a beat sheet, a continuity note, these existed in traditional production because a director had to communicate with a crew. Working solo, people assume they can hold all of that in their head instead. Past six or seven shots, almost nobody can.
And film school does not really prepare you for this version of the problem. Shot lists there are built for a human crew, a camera department, a script supervisor. That format has no obvious AI equivalent, so a lot of beginners have simply never seen what a shot list for prompting a model looks like. They default to skipping it entirely.
A shot list built for a human crew tells people what to do. A shot list built for an AI model has to tell the model what NOT to reinvent between one clip and the next.
How Do You Fix It Before Generating a Single Clip?
The fix is not complicated, which is exactly why people underestimate it. Before opening any video model, write down, for every shot in the sequence:
- The camera angle and any movement, stated as an instruction, not a vibe (“low angle, static” beats “cool shot”)
- What carries over unchanged from the previous shot: light direction, character position, time of day, key prop
- The one thing this shot has to accomplish in the story, in a single sentence
That third point is the one people skip most, and it is the one that saves the most rework. If you cannot say in one sentence why a shot exists, the model cannot either, and it will fill the gap with something generic.
The three lines worth writing before you open a video model. Diagram by Frank Houbre / ScreenWeaver.
This is the exact gap I built ScreenWeaver to close: a workspace that keeps the script, the shot breakdown, and the storyboard connected, so the planning happens before the generation credits get spent, not after. It is not a shortcut around pre-production. It makes skipping pre-production harder to do by accident.
Does Planning First Actually Slow You Down?
This is the honest objection, and it deserves a real answer instead of a slogan. Planning does cost time up front. What it buys back is regeneration time, which is where most beginners actually lose their week.
A 2026 industry review from ProductionHub found that producers using AI inside pre-production workflows, script breakdowns, storyboarding, shot planning, are already seeing productivity gains in the 5 to 10 percent range. The same piece is blunt about the limits: previz and storyboards ”still require heavy human revision,” even with AI in the loop. In other words, the part that saves you money is the planning stage done well, not the generation stage done fast.
I have felt both sides of that math directly. On my anime short Voidborne, made solo, the shots I had fully planned before generating cost me one or two attempts each. The shots I improvised cost four, five, sometimes seven regenerations before I had something usable.
Same model, same prompting skill, wildly different cost. The only variable was whether I knew what the shot needed before I asked for it.
Voidborne went on to win recognition at more than a dozen international AI film festivals, including a placement at the Chroma Awards, which describes itself as the world’s largest AI film and media competition, backed by ElevenLabs, Google, and Adobe. Nobody watching the finished film can tell which shots came from a plan and which came from four rerolls. But I can tell you which ones cost me a night’s sleep.
What Actually Changes Once You Fix This
Once planning comes first, the shift is noticeable fast, even for people who have been generating AI video for months. Regeneration count drops, because the model is being asked a precise question instead of an open one. Editing gets faster, since continuity was decided at the planning stage instead of discovered halfway through the cut.
Regeneration attempts per shot on Lost Garden and Voidborne: planned vs improvised.
The change people bring up to me most, weeks after a masterclass and completely unprompted, is that the work starts feeling like directing again instead of gambling. None of this needs a bigger budget, a better model, or sharper prompting skills. It needs writing down, before you generate anything, what each shot is for. That habit is boring next to watching a model render a beautiful frame. It is also the entire gap between a sequence and a pile of clips.
Do I need a full screenplay before I can plan shots for AI video?
No. A short beat list, one line per beat covering location, who is in it, and what changes, is enough to break into shots. A full screenplay helps for longer projects, but is not required to fix this specific mistake.
Which AI video tools does this apply to?
All of them. Runway, Kling, Veo, Sora, and similar models all generate a single shot at a time with no built-in memory of the shot before it. The planning step exists to compensate for that, regardless of which model you use.
Runway, one of the AI video models this planning habit applies to.
Kling AI, another model this planning habit applies to.
How long should shot planning take for a short film?
For a two to four minute short, plan on roughly the same time you would spend on the generation itself, not more. The goal is a one-sentence purpose and three continuity notes per shot, not a full technical script.
If you are staring at a beautiful AI-generated shot right now, wondering why it still does not sit right next to the one before it, that is the mistake. Write down what the shot is for before you generate the next one. It is the least exciting fix in AI filmmaking, and the one that actually works.
