From Script to Shot List: A Practical AI Video Workflow

Why the script-to-shot-list handoff breaks most AI video projects

A script describes what a story means. A shot list describes what a camera does. Those are different documents written for different readers, and the gap between them is where most production time quietly disappears.

In a traditional shoot, a director and a storyboard artist bridge that gap across several meetings, sketches, and location scouts. In an AI-assisted pipeline, the bridge has to be built in text, because every prompt is essentially a compressed instruction set for a system that has no memory of your intentions. Whatever you leave ambiguous, the model resolves on its own terms — and it will resolve it differently on the next shot.

Three failure modes show up again and again.

Ambiguity amplification. A script line like "she hesitates before answering" is emotionally clear to a human and operationally useless to a generator. What shot size? Is the hesitation in her face, her hands, or the pause before she speaks? Without a decision, you get a random interpretation, and randomness is expensive when you are paying in render time and review cycles.

Continuity drift. A character looks subtly different in shot 4 than in shot 2. A jacket gains a zipper. The kitchen counter shifts from warm oak to grey stone. Each of these is a small inconvenience at generation time and a serious problem at edit time, because audiences read discontinuity as amateurism even when they cannot name what is wrong.

Model mismatch. A tool that excels at sweeping cinematic camera motion may be mediocre at tight facial micro-expression. If your shot list does not distinguish between "drone reveal over a coastline" and "close-up of a tear forming," you will use one tool for both and get a compromised version of at least one.

A structured shot list is the cheapest fix available. It converts fuzzy intent into a table of explicit decisions, and that table becomes the input for every downstream step: prompting, model selection, review, and edit assembly.

What a production-ready shot list must contain

The word "shot list" gets used loosely. In practice, an AI-ready shot list is a spreadsheet-like document with one row per generated clip and a consistent set of columns. The point is not bureaucracy — it is that every column represents a decision you would otherwise make accidentally.

The minimum viable shot spec

At a bare minimum, each row should carry these fields:

Shot ID — a stable label such as S02-04. You will reference this constantly in review notes.
Scene and beat — which story unit this shot serves.
Duration — target length in seconds, chosen for the edit, not for the model's default.
Subject — who or what is on screen, named consistently.
Action — the single physical event in the shot.
Camera — shot size, angle, and movement.
Lighting — time of day, source, and mood.
Setting — location plus the two or three set details that must be visible.
Style — the visual register: documentary, commercial, animation, archival.
Audio intent — dialogue, ambience, or music bed.

Fields that save you later

Once the basics are in place, add columns that pay for themselves during revision:

Continuity anchors — the specific details that must not change between shots (wardrobe, hair, prop placement, weather).
Reference assets — filenames or identifiers for character sheets, location plates, or style frames.
Generator and version — which tool produced the take you kept.
Seed or reference ID — the value that lets you regenerate the same look.
Aspect ratio, frame rate, resolution — locked once, applied everywhere.
Dependencies — which earlier shot this one must match.
Status — idea, prompt drafted, generated, approved, final.

That last column sounds trivial. It is not. On a 40-shot project, tracking status in your head guarantees that two shots get forgotten until the final export.

When you use a tool that accepts a structured brief and returns a shot table automatically, you still want these columns. Automation changes who fills in the table, not what the table needs to contain.

Step 1: Parse the script into scenes, beats, and assets

The first real transformation is structural: prose in, units out.

Scene and beat segmentation

Split the script into scenes by location and time. Within each scene, mark beats — the smallest unit of change. A beat is usually a shift in information, emotion, or power: she recognizes the envelope; he decides not to open it; the neighbor knocks.

Beats are the right granularity because shots serve beats, not paragraphs. A three-sentence paragraph might be one beat or four. The test is simple: if the emotional temperature does not change, you probably have one beat.

Asset tagging: characters, locations, props

Run a second pass purely for nouns. Every named character, every location, every prop that recurs gets an identifier. Then define each one in a short, reusable description of 15 to 30 words.

For example:

CHAR-MAYA: woman in her early thirties, dark curly hair tied back, olive jacket, silver watch on left wrist.
LOC-KITCHEN: narrow galley kitchen, morning light from a single east window, pale wood cabinets, chipped white tiles.
PROP-ENVELOPE: cream envelope, no stamp, handwritten name in blue ink.

These definitions become your continuity contract. Every prompt that includes Maya pastes the same phrase. Boring repetition is exactly what you want here.

Deciding what stays human

The temptation is to automate the entire pass. Resist it in two places. First, emotional emphasis: which beat deserves a close-up is a storytelling judgment, not a parsing problem. Second, ambiguity that is deliberate: if the script withholds a face on purpose, the shot list should preserve that withholding rather than let a tool fill it in with the most predictable choice.

A practical compromise is to let automation produce a first draft of scenes, beats, and assets, then spend twenty minutes per ten pages correcting it yourself. That is dramatically faster than starting from a blank page and still keeps authorship where it matters.

Step 2: Convert beats into shots with deliberate coverage

Now translate beats into shots. This is where most beginners under-shoot.

Beat-to-shot ratios

As a working heuristic:

Dialogue beat: two to four shots (speaker, listener reaction, insert, wide).
Action beat: three to six shots, because motion needs setup, execution, and result.
Emotional beat: one to three shots, held longer, with less camera movement.
Transition beat: one shot, often a detail or a landscape.

For a 30-second piece, expect roughly 12 to 20 shots. For a 60-second piece, 25 to 40. If your count is far below that, your edit will feel slow and static; far above, and the viewer never settles.

Coverage patterns that prevent flatness

Borrow the patterns editors use:

Establishing → medium → close for entering a new place.
Shot / reverse shot for conversation, with matching framing on both sides.
Insert and reaction to compress time and show consequence.
Match cut on shape or motion to move between scenes without a dissolve.

The value of choosing a pattern in advance is that it forces variation. Without a pattern, generative workflows tend to produce the same medium shot at the same eye level again and again, because that is the safest output.

Ordering the table for generation

Once the table is filled, reorder it for production rather than for story. Generate all shots of a single character together, then all shots of a location, then all inserts. You will reuse the same reference image, the same seed, and the same prompt scaffolding, which improves consistency and cuts setup time.

Step 3: Turn each shot row into a precise generation prompt

A shot row is a decision record. A prompt is the instruction. The mapping between them should be mechanical.

The five-part prompt skeleton

Every prompt benefits from the same order of information:

Subject — the character or object using its continuity definition.
Action — one verb, one direction, one result.
Camera — shot size, angle, lens feel, movement.
Lighting and setting — source, quality, time of day, key set details.
Style — medium, grade, texture, era, reference register.

Example: "CHAR-MAYA lifts the cream envelope from the doormat, medium close-up at chest height, shallow depth of field, static camera with slight handheld float, overcast morning light through a hallway window, muted teal grade, 35mm film texture."

That is one sentence and roughly 45 words. It is enough.

Negative constraints and style locks

Add a short block of exclusions that applies to every shot in the project: no on-screen text, no visible logos, no extra fingers, no lens flare unless requested, no changing wardrobe. Keep the block identical across shots — consistency in negatives is as important as consistency in positives.

Prompt length and diminishing returns

Beyond roughly 80 to 100 words, additional description tends to dilute rather than refine. If a shot needs more control, the right move is usually to change the tool, add a reference image, or split the shot — not to write a longer sentence.

Step 4: Match the model to the shot, not the other way around

Every generative video tool has a personality. Some favor smooth cinematic camera moves; some are strong on human faces and small gestures; some handle stylized illustration and motion graphics better than photorealism. Work these strengths into the shot list itself, as a column.

A simple shot taxonomy

Group your shots into a handful of types and assign a preferred tool to each:

Establishing and landscape — tools with strong camera-motion control.
Character close-up and performance — tools with the best facial fidelity and stable identity.
Product and insert — tools that hold detail and geometry under slow movement.
Stylized and animated — tools with strong art-direction adherence.
Transitions and abstract — whichever tool gives you clean, controllable motion.

When to split a shot across two tools

Sometimes the best result comes from generating a base shot in one tool and refining it in another, or from generating a still frame in an image model and animating it. Plan for this in the shot list by adding a "pipeline" note. It matters at review time: a reviewer looking at a rough take needs to know whether they are judging a final render or an intermediate.

Step 5: Lock continuity with references and seeds

Continuity in AI video is not luck. It is a set of deliberate constraints.

Character sheets and reference frames

Create one canonical image per character — front, three-quarter, and profile if possible — and keep it in a single project folder. Use it as a reference in every shot featuring that character. Do the same for each recurring location, ideally captured from the same angle your shots will use.

Wardrobe, props, and time of day

Write down what must not change, and check it before generating: jacket color, hair tie, watch hand, envelope color, weather, light direction. Time of day is the most commonly lost detail, because a washed-out overcast look is the default for many models. If your story moves from morning to evening, encode that shift as an explicit shot-by-shot progression rather than expecting the model to infer it.

Seeds are your other lever. When a take is close but not perfect, regenerating with the same seed and a small prompt change usually lands better than starting from scratch. Record which seed produced each approved shot.

Step 6: Add audio, timing, and edit notes

Video is half the job. The shot list should anticipate sound so you are not fixing rhythm problems in the mix.

Voice and dialogue

For each line of dialogue, note the speaker, the intended pace, and whether the shot shows the mouth. Off-screen lines are dramatically cheaper and often better, because they avoid lip-sync problems entirely while keeping the performance in the viewer's imagination.

Music and cut points

Mark the shot that lands on the first beat of the music, the shot that carries the turn, and the shot that closes. If you are working to a track, list timings per shot so the edit follows the music instead of the music being stretched to fit the edit. A 30-second spot with cuts at 0:02, 0:06, 0:11, 0:17, and 0:24 will feel intentional; evenly spaced cuts feel like a slideshow.

A worked example: a 30-second product spot

Suppose the brief is a countertop coffee grinder. The script is four sentences long. Here is how it becomes a shot list of 14 shots.

Scene 1 — kitchen, early morning.

Wide establishing shot of a dark kitchen, single warm light over the counter. 2s.
Insert: hand reaching for the grinder lid. 1.5s.
Close-up: whole beans poured into the hopper. 1.5s.
Detail: dial turning from coarse to fine. 1.5s.

Scene 2 — the grind.

Medium close-up: the grinder running, subtle vibration, steam-free cool light. 2s.
Insert: grounds falling into a portafilter, macro, shallow depth of field. 2s.
Reaction: the character's face, small satisfied smile, warm light. 1.5s.

Scene 3 — the ritual.

Shot/reverse pair: hand tamping, then the tamper set down. 2.5s.
Espresso extracting, side angle, backlit. 3s.
Close-up: cup lifted, steam visible. 2s.

Scene 4 — payoff.

Character takes first sip at a window, morning light, wide. 3s.
Detail: empty cup on a saucer, grounds knocked out in the bin. 2s.
Product on the counter, clean hero angle, slight dolly in. 3s.
End card with product silhouette, no text in the generated frame. 2s.

Notice how almost every shot is under three seconds and each one answers a question the previous shot raised. Notice also that continuity anchors (same warm lamp, same mug, same jacket) appear in eight of the fourteen rows. That is not padding — it is what makes fourteen separate generations look like one film.

Pre-flight checklist, common mistakes, and FAQ

Pre-flight checklist

Before generating anything, confirm:

Every shot has a duration decided by the edit, not by the tool default.
Every recurring character and location has one canonical reference image.
Continuity anchors are written, not remembered.
Negatives are identical across all prompts.
Audio intent is defined per shot.
Status column is populated and current.

Before export, confirm: aspect ratio and frame rate identical across all clips, no shot longer than necessary, dialogue either on-screen with verified sync or moved off-screen, and music cuts aligned to planned timings.

Common mistakes

Writing prompts first and designing shots second. You end up with beautiful clips that do not cut together.
Changing the character description mid-project. Even small rewording shifts identity.
Generating in story order. You lose the consistency benefits of grouping by character and location.
Over-long prompts. More words stop meaning more control after a point.
Skipping the duration column. Every clip comes out the same length, and the edit has no breathing room.
Reviewing without a reference. Always compare a new take against the approved take of the same character in the previous scene.

FAQ

How many shots should a one-minute video have?
Between 25 and 40 for most commercial or narrative work. Fewer if the piece is deliberately slow and observational; more only if you are doing fast montage.

Do I need a storyboard as well as a shot list?
No, but reference frames help. A single still per shot — even a rough one — dramatically improves consistency, because you can supply it as an image reference rather than describing it in words.

Should I write the shot list before or after choosing tools?
Write the story first, then add a tool column. If you choose tools before you know what the shots need, you will shape the story around the software instead of the audience.

How do I handle a shot that keeps failing?
Change the problem, not the prompt. Split it into two simpler shots, replace it with an insert, or make it an off-screen moment covered by a reaction. Most "impossible" shots are actually two shots written as one.

What is the biggest time saver?
Grouping generation by character and location, and reusing the same seed and reference images. Consistency comes from repetition of constraints, and repetition is fastest when similar shots sit next to each other in the queue.

When is the shot list finished?
When every row has a duration, a camera instruction, a continuity anchor, and an audio note — and when you could hand it to someone else and they would know exactly what to generate. That is the real test of a good handoff.