Why Pre-Production Decides Whether an AI Video Works
AI video generation has collapsed the distance between an idea and a moving image. You type a sentence, wait a moment, and footage appears. That speed creates a quiet trap: because generating a shot feels cheap, planning starts to feel optional. In practice the opposite is true. When generation is fast, the scarce resource stops being rendering power and becomes clarity of intent.
Every shot you generate is an interpretation of what you wrote down. If your prompt says "a woman walks into a cafe," the model chooses her age, her clothes, the time of day, the city, and the emotional temperature of the room. Those choices are wonderful when you make them deliberately and expensive when you make them by accident across forty shots that are supposed to belong to the same film.
The cost of skipping pre-production shows up in three predictable places.
Continuity drift. Faces, jackets, wall colors, and light direction change between shots. The audience may not name the problem, but they feel it as a lack of trust in the world you built.
Re-generation loops. You keep prompting the same beat with slightly different wording, hoping the model stumbles into your intention. Ten attempts later you still do not have the shot, and you have lost the thread of the story.
Unusable coverage. You have beautiful isolated clips that cannot be cut together because nothing matches in scale, direction, or pacing.
Pre-production is not bureaucracy. It is the process of converting a feeling into decisions: which shots exist, in what order, in what visual register, and with what must remain constant. The good news is that modern AI tools make that process faster than a traditional shot list ever could, and they can help you see a scene before you commit to generating it.
What AI-Assisted Scene Design Actually Covers
Scene design is often misunderstood as background art. In an AI workflow it is broader: it is the set of decisions that make a generated frame look intentional rather than plausible. Those decisions stack into four layers, and each layer has its own failure mode.
The four layers of a shot
World. Environment, era, geography, weather, time of day, architecture, texture density, and the general level of realism or stylization. Failure mode: the world changes between shots, so the location reads as generic.
Subject. Who or what is on screen, wardrobe, hair, age, props they carry, posture, and emotional state. Failure mode: the subject becomes a different person with a similar description.
Camera. Framing, lens feel, height, distance, movement, and depth of field. Failure mode: every shot is the same medium-wide angle, so the edit has no rhythm.
Light. Key direction, contrast ratio, color temperature, practical sources, and atmosphere such as haze or dust. Failure mode: light flips sides between cuts, which reads as a mistake even to viewers who know nothing about cinematography.
What an assistant can and cannot decide
An AI planning assistant is genuinely strong at volume and structure. It can expand a one-line premise into a beat sheet, propose twenty shot variations, suggest a color arc, or point out that you have no establishing shot before your close-up. It is weaker at taste and specificity. It does not know that your brand forbids a certain palette, that your actor has a scar on the left cheek, or that the client wants the product logo readable in shot four.
Treat the assistant as a first-pass director of photography and a tireless continuity clerk. You remain the director. The division of labor is simple: let it generate options, and you make irreversible choices.
Storyboards as decision records
A storyboard in this workflow is not a drawing exercise. It is a record of decisions you have already made, so that later prompts can be short and unambiguous. If your storyboard card specifies "nocturnal street, sodium practical light from camera left, 35mm equivalent, subject mid-frame walking away," your generation prompt can be four lines instead of four paragraphs. That is the real productivity gain: fewer words per prompt because more thinking happened earlier.
Building a Storyboard From a Script, Step by Step
Step 1 — Break the script into beats
Read the script and mark every moment where something changes: a new intention, a new location, a new piece of information, or a shift in emotion. Those are beats. A thirty-second piece usually holds six to ten beats; a three-minute piece holds twenty to thirty. Number them. Everything downstream references these numbers, which makes revision conversations dramatically shorter.
Step 2 — Give every beat a shot purpose
For each beat, write one sentence describing what the audience must understand or feel. Then choose the shot size that delivers it most efficiently. Need geography? Wide. Need reaction? Close. Need relationship between two people? Two-shot or over-the-shoulder. This step prevents the most common AI video problem: shots that look impressive but carry no narrative load.
Step 3 — Write shot cards before you write prompts
A shot card is a small block of structured text with fixed fields:
- Beat number and purpose
- Shot size and angle
- Subject description (identity, wardrobe, action)
- Environment description (location, time, weather, key set dressing)
- Light description (direction, quality, color)
- Camera movement (static, pan, dolly, handheld, crane)
- Duration and aspect ratio
- Continuity notes (what must match adjacent shots)
Shot cards are the single highest-leverage artifact in the whole workflow. They are cheap to edit and they make prompt writing mechanical, which is exactly what you want when you are producing thirty shots in an afternoon.
Step 4 — Generate keyframes, not sequences
Generate still images for the first frame of each shot first. Stills are faster, cheaper, and easier to judge. Line them up in a contact sheet and read them as a sequence. If the story does not work as a series of stills, it will not work as motion, because motion adds energy but rarely adds meaning.
Only when the still sequence reads clearly should you animate. And when you animate, prefer short clips — three to five seconds — that you can trim, over long clips you must rescue in the edit.
Step 5 — Lock your reference set
Before the first animated shot, assemble a reference set: a character sheet per recurring subject, a location sheet per recurring environment, and a short note on the grade. Save these as images you can attach to prompts. Most visual inconsistency in AI video is not a model limitation; it is the absence of a stable reference.
Prompt Architecture for Scenes and Shots
The six-slot prompt skeleton
A reliable shot prompt answers six questions in order. Order matters because most models weight earlier tokens more heavily.
- Subject and action — who is doing what, in the present tense.
- Environment — where, when, weather, era, notable set dressing.
- Light — direction, quality, color, atmosphere.
- Camera — shot size, angle, lens feel, movement.
- Style and rendering — realism level, film stock feel, palette, grain, aspect ratio.
- Continuity anchors — the specific details that must match the previous shot.
A filled example: "A ceramicist in her forties lifts a bowl from a pottery wheel, hands wet with clay. Small studio at dusk, window on the right, shelves of unfinished pots behind her. Warm low sunlight rakes across the room from camera right, soft falloff into shadow. Medium close-up, chest height, 50mm equivalent, shallow depth of field, static camera. Naturalistic documentary look, muted earth palette, fine grain, 16:9. Same apron, same hair tie, same window position as the previous shot."
Notice the last sentence. Continuity anchors are boring and they are the difference between a sequence and a slideshow.
Negative constraints that actually help
Negative prompts are most useful when they target artifacts you keep seeing, not abstract quality words. Instead of "bad quality, blurry," specify the problem: extra fingers, text on signs, warped faces in the background, floating objects, mismatched eye direction, lens flare across the subject's face. Keep the list short and specific. A long generic negative list dilutes itself.
Iterating without drifting
When a shot is wrong, change one slot at a time. If you rewrite all six slots and the result improves, you have learned nothing and you cannot reproduce it. If you change only the light slot and the shot improves, you now know that lighting was the variable, and you can apply that knowledge to the next twenty shots.
Keep a running log: shot number, prompt version, one-line note on what changed and what improved. This sounds tedious for a five-shot project and becomes essential for a fifty-shot one.
Keeping Characters and Sets Consistent Across Shots
Character sheets beat adjectives
"A woman in her thirties with brown hair" will produce a different person every time. A character sheet — a single reference image plus a locked description of face shape, hair, wardrobe, and accessories — produces recognition. Build one sheet per recurring subject and attach it to every shot that includes them. If your tool supports multiple reference images, use two per character: one front-facing and one three-quarter view.
Wardrobe, props, and set dressing locks
Audiences track props more than they realize. A mug that is blue in shot two and white in shot nine breaks the scene even if nothing else changes. Write a short continuity list before generating: wardrobe per character, hero props, and three or four distinctive set dressing elements per location. Then treat that list as a contract.
Color and grade continuity
Decide the palette arc of the piece up front. A common structure is warm opening, neutral middle, cool resolution — or the reverse. Write the palette for each scene in words ("sand, rust, pale blue sky") and apply the same grade in post so that the generated shots sit in one world. If you plan to grade heavily in post, keep generation contrast moderate so you retain room to move.
Chaining frames for motion continuity
Many video models accept a first frame and a last frame. This is one of the most practical tools available for AI video: generate your keyframe for shot A, generate the keyframe for shot B, then use the last frame of A as the first frame of B and let the model interpolate the movement. You get continuous motion across a cut without guessing at camera paths. Reserve this technique for moments where continuity is narratively important — a handoff, a follow, a transformation — because it constrains the model and reduces spontaneous invention.
Composition and Camera Language for Generated Footage
Coverage strategy that survives the edit
Shoot (or rather, generate) for the edit. A safe default coverage pattern per scene is: one wide establishing shot, one medium shot that carries the action, one close-up for reaction or detail, and one insert for texture. That is four shots per beat-group, which edits cleanly and gives you flexibility if a shot fails.
Movement the models handle well
Short, single-direction movement is dramatically more reliable than complex choreography. Reliable: slow push in, slow pull out, lateral truck, gentle handheld drift, orbiting arc under ninety degrees, subject walking toward or away from camera. Risky: whip pans, fast dolly zooms, complex multi-axis movement, camera passing through objects, and any shot where the camera and subject both move in unpredictable ways.
Design your sequence so the risky moves carry the fewest narrative requirements. If a fancy move fails, the scene should still make sense without it.
Lens and depth cues
Lens language is one of the fastest ways to add production value. Wide lenses exaggerate space and make environments feel large; longer lenses compress space and isolate subjects. Specify an equivalent focal length in your prompt and keep it consistent within a scene so cuts feel intentional. Shallow depth of field also helps generated footage: it hides background detail that the model may render inconsistently.
Aspect ratio and safe areas
Decide aspect ratio before generating, not after. Vertical, square, and widescreen compositions require different framing and different amounts of headroom. If you need multiple formats, generate the widest and reframe with intent, or plan two separate compositions for hero shots. Also leave room for captions and UI overlays if the video will live in a feed.
Choosing Tools and Assembling a Practical Stack
Planning and storyboard layers
You need three capabilities: script breakdown into beats, a structured shot card board, and still-image prototyping. Some suites combine all three; many creators assemble them from a notes app, a simple table, and an image generator. The tool matters less than the field structure. If your shot cards have consistent fields, any spreadsheet will do.
Generation layer
Most workflows mix model types: one for photoreal humans, one for stylized or animated looks, one for environments. Rather than chasing a single model for everything, test each model on your specific subject — the same character, the same location, the same lighting — and keep notes on which performs best. Practical criteria: reference image support, first-and-last frame support, maximum clip length, motion stability, and output resolution.
| Criteria | Why it matters |
|---|---|
| Reference image support | Determines character and location consistency |
| First/last frame control | Enables motion chaining across cuts |
| Clip length | Short clips force editing discipline, long clips hide problems |
| Motion stability | Fewer warping artifacts mean fewer re-generations |
| Resolution and aspect | Decides your delivery formats |
| Iteration speed | Directly affects how many shots you can explore |
Audio and finishing layer
Sound carries more perceived quality than most creators expect. A simple stack works: a music bed chosen for tempo, a small set of recorded or synthesized sound effects for transitions and physical actions, and a clean voice track. Then grade, add grain or texture, and export. Generated footage often benefits from a light grain layer because it unifies shots with slightly different rendering characteristics.
A Worked Example: 30-Second Product Teaser
Consider a short teaser for a desk lamp. Eight beats, roughly four shots per beat-group compressed into twelve total shots.
Beats. (1) Dark room, single warm glow. (2) Hand reaches toward the switch. (3) Light blooms and reveals the desk. (4) Close-up: the articulated arm adjusting. (5) Wide: the full workspace now lit. (6) Subject sits and begins working. (7) Detail: light falling across notebook pages. (8) Pull back to the lit room, logo area empty for a title card.
Shot cards. For beat three, a shot card might read: medium shot, camera left of desk, subject hands only, warm key from lamp, ambient fill very low, 35mm equivalent, slow push in, three seconds, 16:9, continuity — same desk surface and mug as beat one.
Prompt. "Close medium shot of a wooden desk in a dark room, a matte black articulated desk lamp turning on, warm tungsten pool of light expanding across the surface, a ceramic mug and a closed notebook visible, low ambient blue fill. Camera slightly left of the lamp, 35mm equivalent, slow push in. Photoreal, soft contrast, subtle grain, 16:9. Same desk surface and mug as previous shot."
Pipeline. Generate keyframes for all twelve shots as stills. Review them as a contact sheet, then fix only the two or three that break the sequence. Animate in three-to-five-second clips, chaining frames where the hand or the lamp must connect across a cut. Edit to a temp music bed, add a switch click and a soft room tone, grade warm-to-neutral, add grain, export.
Total planning time for a piece like this is often under an hour, and it saves multiples of that in failed generations.
Common Mistakes, Fixes, and a Pre-Generation Checklist
| Mistake | Fix |
|---|---|
| Prompting full scenes instead of individual shots | Break into beats and write one shot card per beat |
| Changing many prompt slots at once | Change one variable per iteration and log it |
| No character reference | Build a character sheet image before animating |
| Every shot the same size and angle | Plan coverage: wide, medium, close, insert |
| Relying on long clips | Generate short clips and cut them |
| Inconsistent light direction | State key direction in every prompt for that scene |
| Grading late and heavily | Keep generation contrast moderate; grade with headroom |
| Ignoring sound | Choose music tempo before the edit, not after |
| No continuity list | Write wardrobe, props, and set dressing locks per scene |
| Generating before the stills read well | Approve the contact sheet first |
Pre-generation checklist. Story broken into numbered beats. Every beat has a stated purpose. Shot cards complete for every shot. Character and location references saved and attached. Palette arc defined in words. Aspect ratio decided. Light direction consistent per scene. Continuity list written. Stills approved as a sequence. Clip lengths capped. Music direction chosen. Export settings confirmed.
FAQ
How many shots should a one-minute AI video have?
For a one-minute piece, twelve to twenty shots is a comfortable range, depending on pacing. Fast montage sections can run two seconds per shot; dialogue or contemplative sections can hold six to eight seconds. The number matters less than whether each shot has a distinct job.
Do I need to draw storyboards by hand?
No. The value of a storyboard is the decision record, not the illustration. Structured shot cards plus generated keyframe stills give you the same benefit with far less friction. Hand sketching remains useful for quickly exploring composition ideas before you spend time on generation.
Why do my characters keep changing between shots?
Usually because you are describing them with adjectives rather than anchoring them with references. Build one image per character, lock wardrobe and hair details in text, and attach the reference to every prompt. Also check that your prompts are not accidentally introducing new details, such as a hat that appeared in one prompt and quietly disappeared in the next.
How long should each generated clip be?
Three to five seconds is a practical sweet spot. Longer clips tend to accumulate motion artifacts and force you to accept a take that only works at the end. Shorter clips give you more control in the edit and reduce the cost of a failed generation.
Should I generate video or stills first?
Stills first, always. Stills are the fastest way to test composition, color, and character consistency, and a sequence of stills tells you whether the story reads before motion adds noise. Only animate once the stills hold together.
What is the biggest quality upgrade for the least effort?
Sound design and consistent lighting direction. Two extra minutes writing light direction into each prompt, plus a music bed with a deliberate tempo, will improve perceived production value more than any model upgrade.
How do I keep a project consistent when several people work on it?
Keep the shot card sheet, the reference folder, and the prompt log in one shared place with a single owner for continuity. If two people generate shots from the same beat without the same references, you will get two different films.
How do I handle revisions from a client?
Change the shot card first, then the prompt. Revising cards is cheap and keeps the change localized; revising prompts directly usually causes collateral drift in shots the client did not ask you to touch.
