AI Video Scripting Workflow: From Story Idea to Shot List

Why Scripting Still Decides Whether AI Video Works

Every few months a new generation model arrives that renders faces, hands, and camera moves a little more convincingly than the last. What rarely improves is the thing that actually makes a video watchable: the writing underneath it. A model can produce a beautiful four-second shot of a woman walking through rain, but it cannot decide that she is walking away from a decision, that the rain should begin on the second beat, or that the next shot needs to be a tight insert of her hand letting go of a key.

That gap is where most AI video projects stall. Teams generate dozens of gorgeous clips that never assemble into a story, because the script was a single paragraph of atmosphere rather than a production document. The fix is rarely a better model. It is a disciplined pre-production workflow that treats prompts as screenplay pages: structured, specific, and consistent from first shot to last.

This guide walks through a complete scripting workflow for AI video — logline to shot list, prompt grammar, continuity control, camera language, generation-method choices, and the review pass that catches problems before you render anything. It is deliberately tool-agnostic. The same structure works whether you generate with a hosted text-to-video service, a local diffusion pipeline, or a hybrid edit that mixes synthetic shots with real footage.

The Pre-Production Pipeline: From Logline to Locked Shot List

AI video projects fail most often at the handoff between idea and render. Skipping steps does not save time; it moves the cost downstream into reshoots, which in generative work means re-prompting clips you already paid attention to. Build the five artifacts below in order and you will cut your render count dramatically.

Stage 1 — Logline and dramatic question

A logline is one sentence: protagonist, goal, obstacle, stakes. "A night-shift paramedic must decide whether to report the colleague who saved her daughter's life." Notice there is no visual information yet. That is intentional. Loglines are for you, not for the model. They tell you which shots are essential and which are decorative.

Pair the logline with a dramatic question — the thing the audience is waiting to have answered. Every scene should either complicate that question or delay its answer.

Stage 2 — Beat sheet and structure

Before writing scenes, sketch the beats. For short-form video, six to ten beats is usually enough: hook, setup, first complication, escalation, turn, decision, resolution, button. For a thirty-second piece, beats may last only two seconds each. For a five-minute brand film, each beat might hold three or four shots.

Write beats as active statements with a visible action attached: "She recognizes the bracelet." Not: "Emotional moment about memory." Actions are what the camera can photograph, and what a video model can actually render.

Stage 3 — Scene breakdown

Convert beats into scenes with three pieces of information:

• Location and time — "hospital loading dock, 3 a.m., sodium lights, wet asphalt."
• Who is present — name every character, including background figures you want to stay consistent.
• What changes — the value shift in the scene. Someone gains trust, loses leverage, reveals a lie.

This is also the moment to decide what you will not show. Constraints are a creative tool in generative video, because anything you imply off-screen never has to be rendered consistently.

Stage 4 — Shot list with prompt-ready rows

The shot list is the working document. Each row should contain: shot number, duration, framing, subject, action, environment, lighting, camera movement, audio note, and a short prompt draft. Nine columns sounds heavy until you realize it prevents the single most common failure mode — a clip that looks right but belongs to a different film than the clip beside it.

Lock the shot list before you generate a single frame. Reordering a spreadsheet is cheap; reordering rendered clips is not.

Writing Prompts That Read Like Screenplay Pages

A prompt is not a keyword soup. It is a compressed scene description with a consistent internal grammar. The most reliable structure uses five slots, always in the same order, so you can debug one slot at a time when a clip comes back wrong.

The five-slot shot description

1. Subject — who or what, with stable identifying details (age range, wardrobe, hair, prop).
2. Action — one clear physical verb in present tense. "Lifts the phone," not "is thinking about calling."
3. Environment — location, weather, time of day, background activity.
4. Camera — shot size, angle, lens feel, movement.
5. Light and grade — source of light, color temperature, contrast, film look.

Example: "Mid-thirties paramedic in a navy jacket lifts a battered phone from a metal tray; hospital loading dock at 3 a.m., wet asphalt, distant ambulance; medium close-up, slight low angle, 40mm feel, slow push-in; sodium-vapor key light from the left, cool blue fill, high contrast, subtle grain."

That is nineteen seconds of thinking and four seconds of screen time. Repeat it forty times and you have a film with a coherent visual identity.

Dialogue, voice, and audio direction

If your chosen model supports audio, decide early whether dialogue is generated, dubbed, or recorded separately. Generating speech inside a video model is convenient but limits control over performance. A practical compromise: generate silent video with clean lip movement, then layer voice in the edit. Write the dialogue lines in the shot list even if you plan to record them later, because line length determines how long the shot must be.

For music and sound design, note the intent per scene — rising tension, hollow ambience, silence before a reveal — rather than naming tracks. Sound is half the cut.

Negative constraints and continuity notes

Most models accept some form of exclusion list. Keep it short and literal: no text overlays, no extra fingers, no lens flares, no slow-motion unless requested, no cuts inside the clip. Long negative lists degrade output quality because they consume prompt attention that should be describing the scene.

Keeping Characters and Sets Consistent Across Dozens of Clips

Consistency is the hardest problem in AI video, and it is solved with reference assets rather than adjectives. Repeating "the same woman with green eyes" in forty prompts produces forty different women.

A working method:

1. Build a character sheet. Generate or photograph a clean, well-lit reference still of each character in neutral pose. Save it once, name it clearly, and never regenerate it casually.
2. Define a wardrobe lock. Describe clothing in exactly the same words every time, in the same order. "Charcoal wool coat, cream scarf, silver watch on left wrist." Reworded descriptions drift.
3. Create a location plate. One wide establishing frame per location, reused as a style and lighting reference.
4. Use image-to-video or reference-conditioned generation for any shot featuring a recurring character. Text-only prompts are for one-off inserts, environments, and transitions.
5. Log the seed or reference ID for every approved clip, so a close variant can be regenerated instead of reinvented.

Treat this as a continuity bible. On a ten-minute project it is the difference between a coherent film and a slideshow of attractive strangers.

Camera Direction: Cinematography Language That Models Understand

Cinematography vocabulary improves output when it is concrete and worsens it when it is abstract. "Cinematic" is nearly meaningless to a model. "Wide anamorphic frame with horizontal flares, low overcast light, muted teal grade" is actionable.

Framing and shot size

Use standard sizes: extreme wide, wide, medium, medium close-up, close-up, extreme close-up. Add a lens feel when it matters — 24mm for environmental context, 50mm for neutral observation, 85mm for shallow portraits, 100mm macro for texture inserts.

Movement

Describe movement as a physical action with a direction and speed: slow push-in, lateral truck left, handheld follow, crane up revealing the rooftop, static locked-off shot. Avoid combining two movements in one clip unless the model is known to handle compound motion cleanly.

Lighting

Name the source and the quality. "Practical desk lamp as key, dark background, warm 3200K pool of light" is better than "moody lighting." For night exteriors, specify the light source you want visible in frame; otherwise models invent random glow.

Coverage strategy

Plan coverage deliberately: one wide, one medium, two close-ups, and one insert per scene is a reliable minimum. Generate the wide first. If it looks wrong, nothing downstream will save it.

Matching the Shot to the Right Generation Method

Not every shot deserves the same technique. Choosing method per shot reduces cost, time, and frustration.

Shot type	Recommended method	Why
Establishing environment	Text-to-video	No recurring subject to protect
Character dialogue	Image-to-video from a reference still	Preserves facial identity
Product close-up	Image-to-video with a styled plate	Precise lighting and label control
Transition or texture	Text-to-video, short duration	Cheap, forgiving, easy to trim
Precise motion beat	First-and-last-frame or motion-guided	Locks start and end positions
Style exploration	Style reference or low-resolution draft	Fast iteration before committing

Decision criteria, in priority order: does the shot contain a recurring character, does it contain a branded or legible object, does it require a specific start or end frame, and how much screen time does it hold? The longer the shot and the more recognizable the subject, the more control you need to buy with references.

A Worked Example: Sixty Seconds for a Fictional Product

Suppose you are scripting a one-minute film for a fictional commuter backpack. Beats: (1) the bag on a crowded train platform, (2) the strap failing, (3) the owner catching it, (4) a quiet repair-shop moment, (5) resolution at sunrise, (6) product button shot.

Scene breakdown gives you two locations, three characters seen briefly, and one hero prop. The shot list might hold twelve rows: three for the platform, two for the strap failure, two for the catch, three for the workshop, one sunrise wide, one product insert.

Because the bag appears in nine of twelve shots, it becomes a reference asset. You generate one clean studio plate of the bag, one worn variant, and one close-up of the strap stitching. Every subsequent prompt includes the reference and identical wardrobe-style description language: "charcoal canvas commuter backpack, tan leather base, brass zipper, slim silhouette."

Characters appear in only four shots, so you can afford reference stills for the hero and a locked description for two background commuters — no reference needed, as long as their wardrobe words stay identical across prompts.

Duration budgeting matters as much as prompting. Twelve shots across sixty seconds averages five seconds each, which is generous. If the model handles only four-second clips reliably, you need fifteen or sixteen shots, or you need to let some draws hold longer in the edit and accept slight motion.

Finally, the edit. Assemble rough cuts with temp music before refining any single clip. A shot that felt weak in isolation often works perfectly under rhythm.

Common Mistakes and How to Fix Them

Starting with visuals instead of structure. If you cannot summarize the story in one sentence, the twenty prompts you are about to write will not cohere. Fix: write the logline first, always.

Overstuffed prompts. Cramming six actions into one clip produces mush. Fix: one action per clip, stitched in the edit.

Vague verbs. "Experiences joy" renders as nothing. "Laughs, tears in eyes, hand over mouth" renders as a shot.

Renaming characters mid-project. "The woman," then "the medic," then "Sarah" fragments identity. Fix: pick one canonical description string and copy-paste it.

Ignoring aspect ratio and frame rate early. Vertical social cuts and widescreen cuts need different framing decisions. Fix: decide distribution format before the shot list.

Skipping audio planning. Silent drafts hide pacing problems until late. Fix: temp audio on the first assembly.

Rendering every shot at maximum quality. Fix: draft at low resolution, approve the composition, then re-render final picks.

Assuming the model will cut for you. Models generate continuous moments, not edits. Fix: plan the cuts in the shot list and own them in post.

A Review Checklist Before You Generate

Run this pass on your shot list before the first render. It takes fifteen minutes and saves hours.

• Does every shot contain one clear action?
• Are recurring characters and props identified by reference asset or canonical description?
• Is the camera movement specified for each shot, and does it vary across the scene?
• Is the lighting source named rather than described as a mood?
• Do adjacent shots differ enough in size and angle to cut cleanly together?
• Is total runtime within ten percent of the target?
• Is audio intent noted per scene?
• Are negative constraints short and literal?
• Is there a designated hero shot that the whole sequence serves?

If any answer is no, fix the document, not the prompt.

Frequently Asked Questions

Do I need film training to script AI video well?
No, but you need vocabulary. Learning twenty cinematography terms — shot size, angle, movement, key light, color temperature — covers most of what models respond to.

How long should an AI-generated clip be?
Match clip length to what your chosen model handles reliably, then let the edit create length. Many strong sequences are built from three- to five-second fragments with a few longer holds.

Should the script be written for humans or for the model?
Both, in layers. The logline, beats, and scenes are for collaborators. The shot list rows are for generation. Keep them in one document so they stay in sync.

How do I handle dialogue-heavy scenes?
Generate clean visual coverage with lip movement, then record or synthesize dialogue separately. This gives you control over performance and makes revisions far cheaper.

What if a shot never looks right after several attempts?
Change the approach rather than the wording. Convert it to a close-up insert, split it into two simpler clips, or replace it with a reaction shot. Coverage solves what prompting cannot.

How many prompts should I keep in flight?
Draft three to five variations per shot at low resolution, pick one, then commit. Endless variation hunting is the most expensive habit in AI production.

Can I mix generated clips with real footage?
Yes, and it often produces the strongest results. Use generated shots for environments, inserts, and impossible angles, and real footage for performance and product accuracy. Match grade and grain in post.

The practical conclusion is simple. Better models will keep arriving, and they will keep improving the pixels. The structure — logline, beats, scenes, shot list, reference assets, one action per clip, deliberate camera language, and a disciplined review before rendering — is what turns those pixels into a story someone actually finishes watching.