Why AI Video Pre-Production Changed the Job, Not the Craft
Generative video tools have collapsed the distance between an idea and a moving image. A scene that once required a location scout, a lighting package, and a permit can now be tested before lunch. But the constraint did not disappear — it moved. The bottleneck is no longer capture. It is clarity.
Video models generate exactly what you describe, and they generate ambiguity just as faithfully as they generate intention. If your script says "she feels lost," a model has no idea what to render. If your shot card says "medium close-up, handheld, overcast window light, subject centered slightly left, shallow depth of field, slow drift right," the model has everything it needs. Scripting and shot design have therefore become the highest-leverage skills in an AI production pipeline — more important than which model you subscribe to.
This guide walks through a complete workflow: shaping a script that can be shot, breaking it into beats, converting beats into shot cards, choosing a model per shot type, timing cuts to emotional rhythm, and reviewing generations without drowning in versions. It is written for solo creators, small studios, and marketing teams who need repeatable output rather than one lucky clip.
The Five-Stage Workflow From Script to Shot List
Treat pre-production as a pipeline with checkpoints, not a single creative leap. Each stage below produces a document you can hand to another person — or feed to another tool.
Stage 1: Lock the story spine
Before writing a single shot, write one sentence per act. Not per scene — per act. Something like: "A courier discovers the package is alive. She hides it from her employer. She chooses the package over her job." Three sentences, three turns. If you cannot summarize the spine this tightly, no amount of shot polish will fix the story in the edit.
The spine also protects you from a common AI production trap: generating beautiful footage that does not accumulate. A viewer forgives rough texture. They do not forgive a scene that means nothing to the next scene.
Stage 2: Break the script into beats
A beat is a unit of change — something shifts in information, power, or emotion. A two-minute short usually has eight to fourteen beats. Mark them in the script with a simple dash note in the margin: she notices the stain, he stops lying, the light dies.
Beats are your real production units. They map to shots more reliably than paragraphs do, because a paragraph can contain three beats and one beat can span four shots.
Stage 3: Turn beats into shot cards
Each beat becomes one to four shot cards. A shot card is a compact record containing:
- Shot ID — S03-02, so notes stay traceable
- Beat reference — which beat it serves
- Shot size — wide, medium, close, insert
- Action — what physically happens, in present tense
- Camera — angle, movement, height, lens feel
- Light and palette — time of day, source, color bias
- Duration — target seconds in the edit
- Continuity notes — wardrobe, props, screen direction
- Prompt draft — the sentence you will actually paste into a generator
Filling nine fields sounds bureaucratic. In practice it takes ninety seconds per shot and saves you an hour of regenerating clips that cannot be edited together.
Stage 4: Assign camera language deliberately
Do not let the model choose your coverage. Decide in advance which beats deserve a wide, which deserve a close-up, and where the camera should move. A useful default: open scenes wide to establish geography, move to mediums for dialogue, and reserve close-ups for reversals. Movement should have a motivation — a reveal, a follow, a withdrawal.
Stage 5: Generate, review, reconcile
Generate in small batches per beat rather than per full script. Review against the shot card, not against a vague feeling. Reconcile only what fails a specific criterion: wrong framing, broken continuity, unreadable action. This keeps your review loop honest and fast.
Writing Scripts That Translate Into Shots
Most scripts written for humans fail as prompts because they rely on interiority. A script for AI video keeps the interiority for the actor's intent but always adds an observable anchor.
Concrete beats abstract
Replace "she is anxious" with "she checks the door twice in ten seconds." Replace "the city feels oppressive" with "heat shimmer above the asphalt, crowds moving frame-right, no sky visible." The internal state still guides your choices, but the observable detail is what the model can render and what the audience can read.
Write for cuts, not coverage
Traditional coverage assumes you can shoot everything and assemble later. With generative video, every shot costs time and compute, so write in a way that implies a cut. If two ideas live in one sentence, split them into two shots. Short, single-idea shots edit better and generate more reliably than long continuous action.
Track continuity in a dedicated column
Continuity failures — a jacket changing color, a phone switching hands, a room rearranging itself — are the most common reason AI sequences feel fake. Give continuity its own column in your shot sheet and fill it obsessively: wardrobe state, prop position, time of day, weather, screen direction of movement, and which characters are on which side of frame.
Keep dialogue short and gesture-led
Long spoken lines are difficult to render convincingly. Write dialogue in short exchanges and give the model something to show: a pause, a turn away, a hand on a table. If a line must be long, plan to cover it with a reaction shot or an insert rather than a locked close-up of a speaking face.
Shot Design Vocabulary That Makes Prompts Precise
You do not need a film degree, but you do need a shared vocabulary with the model. These four axes cover most of what matters.
The shot size ladder
Wide (geography and isolation), full (body language), medium (conversation), medium close-up (emotion with context), close-up (emotion alone), extreme close-up (detail as pressure), insert (object information). Naming the size explicitly prevents the model from defaulting to a mid-shot every time.
Camera movement
Static, pan, tilt, dolly in or out, tracking, crane, handheld drift, orbit. Add speed: slow, steady, urgent. Movement should answer a question — a slow push in asks what the character is hiding; a handheld drift suggests instability. Avoid stacking three movements in one prompt; models blur them into mush.
Lens, depth, and focus
"Shallow depth of field with the background falling into soft bokeh" reads very differently from "deep focus, everything sharp." Mention focal feel when it matters: wide-angle distortion for unease, long-lens compression for intimacy at a distance. Rack focus is a powerful storytelling tool but hard to control — use it sparingly.
Light and color
Specify the source and the mood: overcast window light, single practical lamp, hard noon sun with sharp shadows, neon spill from a sign frame-left. Then specify the palette bias: cool teal shadows, warm amber highlights, desaturated except for red. Consistent light language across a sequence is what makes separately generated shots feel like one film.
Matching Models to Shot Types
No single generator wins every shot. Think in terms of casting: assign each shot to the model most likely to nail it.
Realism, faces, and dialogue
For naturalistic human performance, prioritize models with strong facial coherence and stable skin texture. Test with the hardest shot in your script — a medium close-up of someone speaking under mixed light — rather than a flashy landscape. If the face holds for four seconds without warping, the model can carry your dialogue scenes.
Stylized and effects-heavy shots
For stylized looks, particle effects, fantasy environments, or animation-adjacent styles, look for models with strong prompt adherence and stylization controls. Photorealistic models often fight you when you want a graphic, illustrated feel; a style-forward model will get there in fewer attempts.
Multi-shot consistency and character locking
When a character must persist across many shots, test the model's reference-image and multi-image conditioning. Feeding two or three reference frames — a front-facing portrait, a three-quarter view, and a full-body shot — dramatically improves identity stability. Plan your reference set before you generate anything; retrofitting consistency later is expensive.
Fast iteration models for coverage
Keep one fast, inexpensive model in your toolkit purely for blocking. Use it to test pacing, framing, and edit rhythm at low fidelity, then regenerate the shots that survive the cut on a higher-quality model. This two-tier approach is the single biggest time saver in an AI video pipeline.
A Worked Shot Card Example
Take a simple beat: a courier realizes the package is breathing.
S05-01 — Insert. The package on the passenger seat, lid slightly ajar. Static, slight handheld. Late afternoon light through a windshield, warm amber with dust in the air. 2.5 seconds. Continuity: peeling sticker on the lid top-left, seatbelt across the package. Prompt draft: Close insert shot of a cardboard package on a car seat, lid slightly open, dust motes drifting in warm late-afternoon light through the windshield, subtle handheld camera, shallow depth of field.
S05-02 — Medium close-up. The courier's eyes flick down to the seat, then up to the road. Slow dolly in. Same light, same direction. 3 seconds. Continuity: same jacket as S02, hair tucked behind right ear. Prompt draft: Medium close-up of a young courier driving, glancing down at the passenger seat, slow dolly in, warm side light, natural skin texture, background softly blurred.
S05-03 — Wide. The car continues along an empty road, camera static and high. 4 seconds. Continuity: car moves frame-left to right, road surface dry. Prompt draft: High wide static shot of a small car traveling along an empty rural road at golden hour, long shadows, minimal composition.
Three cards, three shots, roughly ten seconds of screen time, and enough specification that any competent generator can produce something editable. That is the scale you should aim for: small, specific, and continuous.
Emotional Pacing: Timing Cuts to the Beat
Shot duration is an emotional instrument. Fast cutting raises tension; long takes create weight, unease, or contemplation. Most AI-generated sequences feel monotonous because every shot runs the same length.
A practical approach is to score your edit on paper before generating. Assign each shot a target duration and vary it in a pattern: 3 seconds, 3, 2, 1.5, 1, then a 6-second hold at the emotional peak. The contrast does the work. Also mark where sound will carry the transition — a cut on a breath, a door, or a musical hit lands far harder than a cut on silence.
Finally, respect the exit. A shot needs a beat of stillness or a clear action completion before a cut, or the edit feels like a slideshow. When you write action into a shot card, include how the action resolves, not just how it begins.
Common Mistakes and How to Fix Them
| Mistake | Why it hurts | Fix |
|---|---|---|
| One giant prompt per scene | Model averages everything into mush | Split into shot cards, one idea each |
| Inconsistent light description | Sequence feels assembled from different films | Lock a light and palette phrase and reuse it verbatim |
| No continuity column | Wardrobe and prop drift | Track wardrobe, props, screen direction per shot |
| Generating before scripting | Beautiful clips that cannot be edited | Finish the beat sheet first |
| Reviewing by vibe | Endless regeneration loops | Score each clip against the shot card fields |
| All shots the same length | Flat pacing | Assign target durations and vary them |
| Long dialogue in close-up | Facial warping, uncanny delivery | Break lines into short exchanges and reaction shots |
Two more quiet killers: letting the model choose camera movement, and forgetting to plan sound. Camera movement is storytelling, not decoration — decide it yourself. And because AI video has no natural room tone, you must build a sound bed from scratch; a sequence with a deliberate ambience track reads as far more professional than a technically cleaner one without.
Quality Control Without Endless Regeneration
Set a review rubric and apply it mechanically. Four questions: Does the framing match the card? Does the action read in one viewing? Does continuity hold against the previous shot? Does the tone match the sequence? If a clip passes three of four, keep it and fix the issue in the edit or with a small insert rather than regenerating the whole shot.
Version your files with a consistent naming pattern — project, sequence, shot ID, version. Keep a short notes log of what you prompted and what failed; patterns emerge quickly. Most creators find that two or three specific prompt mistakes cause the majority of their wasted generations, and those mistakes are only visible in a written log.
Finally, cut before you polish. Assemble a rough sequence at lower quality, watch it end to end, and only then invest in higher-fidelity passes for the shots that survive. Editing reveals which shots are load-bearing, and it is far cheaper to discover that on a draft than on a finished render.
Frequently Asked Questions
How many generations should I plan per shot?
Budget three to five attempts for straightforward shots and eight or more for shots with faces, hands, or complex action. If a shot needs more than ten attempts, the shot card is usually the problem, not the model — simplify the action or split the shot.
Do I need film vocabulary to get good results?
You need precision, not jargon. "Close-up, slow push in, warm window light" works as well as a technical description of focal length. What matters is naming the size, the movement, and the light every single time.
How do I keep a character consistent across many shots?
Build a reference set first: one front-facing portrait, one three-quarter view, one full-body frame, all in the same light. Use multi-image conditioning where available, and repeat the same descriptive phrase for the character in every prompt rather than inventing new adjectives.
What saves the most time overall?
The two-tier approach: block every shot on a fast model, edit the sequence, then regenerate only the surviving shots at high quality. It prevents expensive rendering of shots that never make the cut.
Can I write the script and design shots in one pass?
You can, and experienced creators often do. But if you are new, separate them. Writing for story and writing for shootability pull in different directions, and doing both at once usually produces a script that reads well but generates badly.
How long should a finished AI short be?
Short is a feature, not a compromise. Ninety seconds to three minutes is the sweet spot for most projects, because every additional shot multiplies continuity risk. Build a tight sequence you can finish rather than an ambitious one you abandon.
Where to Start Tomorrow
Pick a single scene — thirty seconds, two characters, one location. Write the spine in three sentences. Break it into beats. Build shot cards for each beat with size, movement, light, duration, and continuity filled in. Generate rough versions on a fast model, cut them together, and watch the result with sound off, then with sound on.
The first pass will be rough. The second will reveal which of your habits cost you time. By the third, you will have a personal pipeline: a reusable shot card format, a short list of models cast by shot type, and a review rubric that stops you from regenerating endlessly. That pipeline — not any single tool — is what turns scripting and shot design into a repeatable production skill.
