A good script has always been the foundation of a good video. But the script you write for a human film crew and the script you write for an AI video model are two very different documents. Human directors can read subtext, interpret mood, and fill in the gaps when a scene description is loose. A generative model cannot. It takes your words literally, and whatever you leave ambiguous, it will decide on its own, usually not in the way you intended.
This is why the same idea can produce either a polished, coherent video or a sequence of beautiful but unrelated clips. The difference is rarely the model. It is almost always the script, or more precisely, the way the script has been structured for machine execution. This guide explains how to transform a traditional screenplay into an AI-ready script: breaking scenes into shot units, writing descriptions the model can actually follow, building character reference data, and speaking the language of cinematography in a way the model understands.
Why Traditional Scripts Fail with AI Video Models
Traditional scripts are written for interpreters. When a screenplay says "a tense conversation in a dimly lit office," a human director knows what that means emotionally and visually. They can decide the camera angle, the blocking, the light quality, and the pacing. A generative model sees a vague instruction and fills the gaps with whatever its training data suggests first.
The result is a series of "average" choices. The office looks generic, the lighting looks neutral, and the tension you imagined never appears, because you never actually described it. Traditional scripts also fail on consistency. A human crew remembers that the character was wearing a blue jacket in scene one. A model generates each shot independently, so unless the jacket is specified in every single prompt, it will change color, style, and fit from shot to shot.
The core problem is translation. You are writing in a language of implication, but the model only understands a language of specification. The solution is not to write worse scripts; it is to write scripts in a format that preserves your creative intent while giving the model everything it needs to execute it.
The Shot Unit: Your New Basic Building Block
The single most important change you can make is to stop thinking in scenes and start thinking in shot units. A shot unit is the smallest self-contained description that can be handed to a video model as one generation request. Instead of one long paragraph describing a whole conversation, you write several shot units, one for each distinct camera setup and action beat.
This restructuring has three benefits. First, each generation request is simple enough for the model to execute precisely, which improves adherence. Second, it gives you natural checkpoints for consistency: every shot unit can reference the same character description, so drift has fewer places to hide. Third, it makes the editing workflow obvious, because each shot unit maps directly to a clip in your timeline.
Anatomy of a Shot Unit
A complete shot unit contains four elements. The subject block identifies who or what is in frame and repeats the canonical character description. The action block states what happens, written in present tense and kept to one clear action. The camera block specifies the framing, angle, and movement. The environment block describes the location and lighting in enough detail to be reconstructed.
Here is a minimal example: subject "the same woman in the red coat from the reference set," action "she turns and looks at the camera," camera "medium close-up, eye level, slow push-in," environment "rainy city street at dusk, neon reflections on wet asphalt." Every element is explicit. Nothing is left to interpretation.
Writing Descriptions the Model Can Execute
The quality of your prompts depends on how precisely you can describe the visual reality you want. Think of yourself as writing for a brilliant but literal-minded assistant who has never seen the world.
Subject, Action, Camera, Lighting
Start with the subject and repeat its stable identity every time. If you have reference images, say so explicitly in the prompt; if not, give a compact physical description that never changes between shots.
The action comes next, in present tense and singular. "She walks toward the door" is executable. "She reflects on the conversation while moving hesitantly" is not, because the emotion is invisible to the model. If the emotion matters, show it through action: "she pauses, looks down, then walks slowly toward the door."
Camera language is your most powerful tool for cinematic feel. Learn a small vocabulary: wide shot, medium shot, close-up, extreme close-up, low angle, high angle, bird's-eye view, over-the-shoulder, tracking shot, pan, tilt, push-in, pull-back. Each of these maps to a visual result, and using them consistently makes your output look directed rather than accidental.
Lighting deserves its own sentence. Instead of "moody," write "single hard light from the left, deep shadows on the right side of the face, cool color temperature." The model will honor the concrete description far better than the adjective.
Building Character Reference Data with Multiple Images
Words alone will never fully stabilize a character. The most reliable way to keep a face, a costume, or a product consistent is to provide visual references, and the most robust form is a set of multiple images of the same subject.
A single reference image anchors appearance in one pose and one light. Multiple images, taken from different angles and in different expressions, let the model separate the durable identity from the incidental appearance. When you combine this reference set with your canonical text description, you create a strong constraint that the model carries into every shot unit.
Build the reference set before writing the script. Define the character visually first, shoot or collect four to six consistent images, and correct their color so they agree. Then, as you write each shot unit, reference the set and repeat the same textual description. The combination of stable imagery and stable language is what defeats character drift.
Using Cinematic Directives to Sound Like a Director
Cinematic language is more than camera names. Directors talk about depth of field, lens distortion, color grading, motion blur, composition rules, and pacing. All of this can be communicated to a video model, but only if you translate it into explicit instructions.
Describe depth of field with words like "shallow depth of field, background softly blurred" or "deep focus, everything sharp." Describe composition with concrete terms: "subject centered," "rule of thirds, subject on the left third," "symmetrical composition." Describe motion quality: "slow, fluid camera movement" versus "handheld, slightly shaky." Describe the feel of the grade: "warm skin tones, teal shadows," "high contrast, desaturated."
The goal is to build a shared vocabulary between your creative intent and the model's execution. Over time, you will develop a personal set of cinematic directives that reliably produce the look you want, and you can reuse them across projects like a director reusing their signature style.
Choosing the Right Model for Each Scene
Not every scene needs the same model. Different models have different strengths: some excel at photorealistic humans, others at stylized animation, others at fast motion or complex physics. A professional workflow treats the model library as a toolkit and assigns each shot unit to the tool best suited for it.
This means your script should include a technical note per shot unit, indicating the target model and the reason. A shot requiring a realistic human face under difficult lighting might go to a photorealistic model, while a dream sequence with surreal physics might go to a more stylized one. This per-shot assignment is the difference between using one average tool for everything and assembling a film from the best of each.
Keep a log of what works. Note which models deliver which looks, how long each takes, and what their failure modes are. Over a few projects, this log becomes a reliable decision guide that saves time and generation budget.
From Script to Edit: The QA Loop
A script is not finished when the last shot unit is written. It is finished when the edit works. The practical workflow is a loop: write, generate, review, revise.
Generate one test for each new character and each unfamiliar setting before producing the full sequence. Check adherence, consistency, and quality against the shot unit. If the output drifts from the description, revise the prompt, not the model. If the output is faithful but ugly, adjust the cinematic directives. Only after the test passes should you generate the remaining shots.
Then assemble a rough cut and watch it as a sequence, not as separate clips. Continuity problems that are invisible in individual shots become obvious in the cut: a lighting mismatch, a costume change, a jump in camera language. Go back to the affected shot units, fix them, and regenerate only what the edit actually needs.
Before and After: One Scene, Two Scripts
To make the difference concrete, here is the same scene written two ways.
The traditional version: "Marina walks into the café, nervous about the meeting. The afternoon light streams through the window."
The AI-ready version, as three shot units: "Shot 1: Marina, the woman in the grey coat from reference set A, enters the café and pauses at the door. Wide shot, eye level, static. Warm afternoon light through a large window, dust visible in the light beam. Shot 2: Close-up of Marina's face as she scans the room. Shallow depth of field, background blurred. Same warm light, soft shadows. Shot 3: Marina walks toward the camera between two tables, then stops. Medium shot, slight low angle, slow push-in. Café interior, wood tones, warm light, empty background."
The second version tells the model exactly what to generate, gives the editor clean cut points, and keeps the character stable across all three shots. That is the entire difference between a script and an execution plan.
Adapting Shot Units to Platforms and Aspect Ratios
The same scene can demand completely different shot units depending on where it will be published. A vertical short for a social feed, a horizontal video for a streaming platform, and a square cut for a carousel all place different demands on framing, composition, and pacing.
For vertical formats, the subject should fill the frame and composition should favor the center. Wide establishing shots lose most of their value when the sides are cropped away, so design shot units that work within the vertical frame: close-ups, medium shots, and shots where the action stays in the central band. Camera movement matters too: a horizontal pan wastes screen space in a vertical video, while a push-in or a vertical tilt uses it well.
For horizontal formats, you have the opposite problem: empty space on the sides. Shot units for widescreen should plan for depth, using background elements and layered composition so the frame feels intentional at both edges. For square formats, keep the essential subject within the central square and treat the top and bottom as safe zones.
Pacing also changes by platform. A shot unit destined for a fast-cut social video should describe a single clear beat that can live for two or three seconds. A shot unit for a longer documentary-style piece can afford slower movement and longer holds. Write the aspect ratio and the intended cut length into the technical note for each shot unit, and your generations will line up with the edit instead of fighting it.
This platform awareness is part of what separates generic AI content from content that feels native to the feed where it appears. The model does not know where the video will live, so you have to decide that before you write the prompt.
Frequently Asked Questions
Do I need to write shot units for every video?
For short clips, a single well-structured prompt can be enough. For anything with multiple scenes, characters, or a story, shot units are the safest way to maintain control.
How long should a shot unit be?
As long as needed to specify subject, action, camera, and environment, and no longer. Usually two to four sentences. If you need more, split the shot.
What if the model ignores part of my prompt?
Simplify. Models lose fidelity on overloaded prompts. Keep the essentials and move secondary details to a separate shot unit.
Can reference images replace text descriptions?
They complement each other. References anchor identity; text carries action, camera, and mood. Use both.
How do I keep a consistent style across a whole project?
Define a style block: color grade, lighting character, lens feel, and composition rules. Reuse it verbatim in every shot unit, the same way you reuse the character description.
