AI Shot Design Workflow: Storyboards, Camera Moves, Edits

Shot design is where most AI video projects quietly fail

Most disappointing AI video is not disappointing because the model is weak. It is disappointing because nobody decided what the shot was supposed to accomplish. A prompt like cinematic drone shot of a city at dusk can produce a technically impressive clip that carries zero story information: no clear subject, no spatial logic, no relationship between foreground and background, no reason for the camera to move in that direction. The clip looks expensive and communicates nothing.

Shot design prevents that. It is the practice of answering a short list of questions before generation: What is the subject? Where is the camera relative to that subject? What does the audience know at the end of the shot that they did not know at the start? How does this shot hand off to the next one?

Traditional productions answer those questions with a director, a cinematographer, a storyboard artist and a location scout. Solo creators and small teams rarely have all four. That is the gap AI-assisted planning closes. Not by replacing craft, but by making the pre-production conversation fast enough to actually happen at all.

This article covers a full workflow: script breakdown, shot listing, previsualization, camera language, consistency management, generation settings, editing and the mistakes that cost the most time. It is deliberately tool-agnostic, because the underlying thinking survives every model update.

Why pre-production is the highest-leverage place to use AI

Iteration is cheap in text and expensive in pixels. Rewriting a shot description takes ten seconds. Regenerating a thirty-second sequence, fixing continuity, re-timing audio and re-editing takes an afternoon. Every decision you settle before the first generation is a decision you do not have to remake in post.

AI assistants are genuinely strong at three pre-production jobs:

Volume. They can draft forty variations of a shot description in the time it takes you to type three. Breadth of options is exactly what early planning needs.
Translation. They can convert a plain-language intention such as she should feel trapped into concrete camera, lens and lighting language.
Bookkeeping. They can keep a running list of shots, props, locations and continuity notes, and flag when shot 14 contradicts the note you wrote for shot 3.

They are weak at two jobs you must keep for yourself:

Taste. Choosing which of the forty variations is right for the tone of the piece.
Structure. Deciding what the sequence is actually about, and therefore which shots deserve screen time.

A useful rule: use AI to expand and organize options, use your own judgment to collapse them. If you let the tool collapse options for you, you get average work, because the tool does not know what you are trying to say.

A quick decision test

Before generating anything, ask whether you can describe the shot in one sentence that includes a subject, an action and a change. Wide shot: two figures walk toward the gate, then stop when the light goes out. If you cannot, the shot is not ready. A generation tool will fill the gap with generic motion, and generic motion reads as filler in the edit.

Step 1: Break the script into beats, not pages

Pages are a formatting convention. Beats are the unit of storytelling, and they are what you should be planning around.

A beat is a change

A beat is the smallest unit of story in which something changes: a decision, a reveal, a reversal, a shift in power. A four-page scene might contain three beats or eleven depending on how much turns. Writing beats first gives you a natural map for coverage, because each beat usually needs one establishing shot, one or two coverage shots, and one reaction.

A practical method:

Read the scene out loud and mark every moment where you would naturally pause.
Label each pause with what changed. She decides to lie. He notices the door is open.
Under each label, write one sentence describing the shot that would carry that change most efficiently.

You will often discover that a scene you imagined as twelve shots is really five, plus one insert. Fewer, better-chosen shots are almost always stronger and much cheaper to generate.

Turn beats into coverage

Standard coverage logic still applies in AI-assisted work: master, two-shot, single, reverse, insert, cutaway. What changes is that you can previsualize all of them before committing. Use the assistant to draft a coverage list per beat, then delete aggressively. If two shots carry the same information, keep only the one with better composition.

A useful constraint: for every beat, ask which single frame could represent it in a still photo. That frame is your anchor shot. Build coverage around it rather than generating a symmetric set of angles you will never use.

Step 2: Build a shot list that survives generation

A shot list is only useful if it contains the fields that generation actually requires. Vague lists produce vague clips.

Required fields for every shot

Shot number and beat. So you can reorder in the edit without confusion.
Subject and action. Who is doing what, in present tense.
Shot size. Wide, medium, close, extreme close, insert.
Camera height and angle. Eye level, low, high, over-the-shoulder, top-down.
Camera movement. Static, push in, pull out, pan, tilt, track, arc, handheld, crane.
Lens feel. Wide, normal, long, macro. Describe the distortion and depth compression you want, not the focal length, unless your tool supports it.
Lighting and time of day. Practical sources, direction of key light, contrast ratio.
Duration. Target length before you start generating, not after.
Continuity notes. Wardrobe, props, screen direction, weather, injuries, anything that must match.

Generating with these fields in front of you cuts wasted renders dramatically, because most failed generations fail on an unspecified field rather than on an unspecified style.

How many shots do you actually need

A rough starting ratio for narrative work: one shot per beat plus one coverage shot per beat, plus inserts for any physical action that needs to be legible. A three-minute sequence typically lands between 25 and 60 shots. If your list is 200 shots for three minutes, you are writing a shot list for an editing style, not for a scene. Cut it.

Short-form vertical work inverts the ratio. In a 30-second piece, 8 to 14 shots is usually plenty, and the first shot has roughly two seconds to establish subject and stakes before the viewer scrolls.

Step 3: Previsualize cheaply and often

Previsualization is not a luxury step for large productions. It is the fastest way to discover that the sequence does not work while it is still cheap to change.

Storyboards as a thinking tool

You do not need illustration skill. A storyboard can be ten rectangles with stick figures, arrows for camera movement, and one line of text each. What matters is that you can see the sequence of sizes and the direction of movement across the whole scene at once. Problems that are invisible shot by shot become obvious in a strip of twelve frames: three wide shots in a row, movement that flips direction every cut, no close-ups in the emotional peak.

AI assistants help here in three ways:

Generating a first-pass frame for each shot from your description, so you have something to react to.
Keeping the strip consistent in aspect ratio and style so the sequence reads as one scene.
Drafting an animatic order and marking suggested durations based on dialogue length and action complexity.

Camera rehearsals before generation

Before you generate final footage, generate low-resolution tests that answer one question each. Does the push-in land on the right beat? Does the arc reveal the second character at the right moment? Does a wide shot hold the tension, or does it feel distant?

Treat these tests as rehearsals, not as footage. The goal is to make your decisions irreversible in pre-production so that generation becomes execution rather than exploration.

Step 4: Choose camera language on purpose

Camera language is the most abused part of AI video, because generative tools reward motion: moving cameras look impressive even when the movement means nothing. Freeze that instinct and choose movement for a reason.

Movement vocabulary and what each move communicates

Static. Observation, stability, deadpan comedy, tension that comes from what is happening inside the frame. Underused and powerful.
Push in. Increasing intimacy, realization, dread. Best reserved for beats where something changes internally.
Pull out. Isolation, revelation of context, endings. A pull-out is a statement.
Track with subject. Momentum, companionship, momentum toward a goal.
Arc or orbit. Emphasis, heroism, unease when the arc is too slow.
Handheld. Immediacy, documentary truth, chaos.
Crane or drone. Scale, geography, transitions between story locations.

When you ask an assistant for movement suggestions, ask it to justify the movement in terms of the beat. Any suggested move without a story reason is decoration.

Lenses and perspective decisions

Lens choice is really a choice about the viewer's relationship to the subject. Wide lenses exaggerate depth and distance, which is why they suit geography, interiors and comedy of space. Long lenses compress depth and isolate faces, which is why they suit intimacy and surveillance-like observation. Macro shots make texture into subject matter and are excellent for inserts.

Two practical rules:

Do not change the perspective logic mid-scene. If a scene is shot on wide lenses, a random long-lens shot reads as a mistake unless the cut is intentional.
Use one deliberate perspective shift per scene. A single push from wide to long perspective at the emotional peak lands harder than constant variation.

Step 5: Keep characters, wardrobe and geography consistent

Consistency is the number one reason AI sequences feel broken. Faces drift, hair length changes, a coat swaps colour, a doorway moves from the left wall to the right. Audiences tolerate rough texture far better than they tolerate a character changing identity between cuts.

Build a continuity document and reuse it verbatim:

Character block. Age range, build, hair, facial hair, distinguishing features, clothing with colour and material, accessories. Write it once, paste it into every prompt featuring that character.
Location block. Architecture, wall colours, window placement, furniture, time of day, weather, light direction.
Screen direction map. Which way characters move and where the camera sits relative to a fixed landmark. This prevents the classic mistake of a character walking left, then right, then left across consecutive shots.
State log. Injuries, wet clothes, missing props, food on a plate, anything that must escalate or stay constant.

Then do consistency checks in batches rather than one at a time: generate a contact sheet of a character across all shots and look at them side by side. Drift is easy to see next to an identical neighbour and almost invisible shot by shot.

When to fix drift in generation vs in post

If a wardrobe colour is slightly off, colour correction in the edit is faster than regeneration. If facial structure has changed, regenerate. If only background details are wrong and the shot is a tight close-up, ignore it. Spend your regeneration budget on identity, not on trivia.

Step 6: Match generation settings to shot type

Different shots fail for different reasons, so tune your approach per shot type rather than using one global setting.

Establishing wide shots. Prioritize stability and clean geometry. Keep motion simple and let the frame hold long enough to establish space.
Dialogue coverage. Prioritize face stability and eye-line. Keep movement minimal or static; listeners and speakers must sit in consistent positions across shots.
Action shots. Shorten duration per shot and cut faster. Real speed is often better conveyed in the edit than inside a single generated clip.
Inserts. Prioritize texture and detail. Macro-style prompts with a single action and a shallow depth of field read very well.
Transitions. Design them last, after the sequence works with hard cuts. If a scene needs a fancy transition to make sense, it usually needs a shot instead.

Two more decisions matter more than any slider: duration and number of attempts. Set a target duration and stick to it so your edit is not hostage to whatever the model produced. And budget a fixed number of attempts per shot — three is a common practical limit — then either accept the best take or rewrite the shot description. Endless retries on a badly specified shot is the single biggest time sink in AI-assisted production.

Step 7: Edit, sound and finish

The edit is where shot design is either confirmed or exposed. Assemble a rough cut with no effects at all, in the order you planned, at planned durations. Watch it once without stopping. Whatever feels long or confusing is a design problem, not a rendering problem.

Practical finishing order:

Picture lock on structure. Do not grade or add effects before the sequence works with plain cuts.
Sound first, effects second. Ambience, movement sounds and music do more for perceived quality than most visual effects. A weak shot with strong sound design reads as intentional; a strong shot with no sound reads as a test render.
Continuity pass. Check screen direction, wardrobe, props and light direction against your state log.
Grade for cohesion. Unifying colour and contrast across generated clips is the fastest way to make heterogeneous footage feel like one film.
Final polish. Grain, subtle vignetting and consistent sharpness hide model-to-model differences better than any single filter.

If you cut in DaVinci Resolve, Blender or any mainstream editor, the workflow is identical. The tool does not matter; the order does.

Common mistakes and how to fix them

Writing prompts instead of shot descriptions. Prompts describe appearance, shot descriptions describe intent. Fix: convert every prompt into subject, action, size, angle, movement, lens, light.

Too many shots, all the same size. New creators default to medium shots. Fix: force variety — at least one wide and one close-up per beat.

Movement everywhere. Fix: make at least half your shots static and reserve movement for beats that change.

Skipping the state log. Fix: keep a running continuity document open while you write prompts, and check it before each generation session.

Judging single clips instead of sequences. Fix: assemble a rough cut before deciding a shot failed. Many clips that look weak alone work perfectly in context.

Chasing models instead of shots. Fix: pick one or two generation tools, learn their behaviour thoroughly, and invest your remaining time in planning. New tools rarely fix a design problem.

FAQ

Do I need storyboards if I am generating the footage anyway?

Yes, but low fidelity is fine. The purpose is not to show anyone; it is to check sequence logic — shot sizes, movement direction and pacing — before you commit time to generation.

How long should each AI-generated shot be?

Plan between three and six seconds for most narrative coverage, one to two seconds for inserts and cutting beats, and up to eight seconds for an establishing shot that needs to breathe. Decide before generation, then edit to your plan rather than to the clip.

What is the minimum viable planning document?

A table with columns for shot number, beat, subject and action, shot size, movement, duration and continuity notes. That single table eliminates most wasted generations.

How do I stop characters from changing between shots?

Use a fixed character block pasted verbatim into every prompt, keep lighting and lens logic consistent within a scene, and review character shots side by side in batches rather than one at a time.

Should I generate everything myself or mix live-action plates with generated footage?

Mixing works well when generated shots are treated as inserts, establishing shots or impossible-to-film moments, and live-action carries faces and dialogue. Keep a consistent grade and grain across both, and the seam disappears faster than you expect.

What is the best way to learn shot design quickly?

Recreate a two-minute scene you admire, shot for shot, with your own descriptions. You will learn pacing, coverage ratios and movement logic faster from imitation than from any tutorial, because you have to justify every cut to yourself.