AI Director Assistant: Script Writing and Shot Design Guide

Why an AI Director Assistant Changes Pre-Production

Most AI video projects fail before the first render. The generator is rarely the weak link. What breaks a project is the absence of a director's plan: no beat structure, no coverage logic, no continuity rules, and no shared vocabulary between the writer and the person typing prompts.

An AI director assistant closes that gap by putting three jobs in one workspace: story development, shot planning, and continuity management. Instead of juggling a script document, a spreadsheet, a folder of reference images, and a notes app, you move a scene beat directly into a shot description, and that shot description directly into a generation prompt. Each step inherits the last one's context.

The practical payoff is speed without drift. A 90-second narrative short typically needs 40 to 70 generated clips once you account for coverage, inserts, and pickups. If every clip is prompted from scratch, your character's jacket changes color, your location's light direction flips, and your pacing collapses in the edit. If every clip descends from a documented shot plan, the edit assembles itself and reshoots become targeted instead of total.

This guide walks through the full pipeline: writing a script that a generative model can actually serve, converting that script into a shot list, designing prompts from camera language, building continuity locks, and running batches without losing track of versions.

What an AI Director Assistant Actually Does

It is not an autopilot that writes your film. Treat it as a structured assistant with three distinct layers of responsibility.

Layer one: story structure

The assistant helps you pressure-test narrative shape before you commit to visuals. You feed it a logline, a character pair, and a target runtime, and it proposes scene beats with rough durations. You keep the ones that create conflict and cut the ones that only explain.

Layer two: coverage and shot translation

The assistant converts prose into shot-sized units. A paragraph like "Mira hesitates at the door, then walks into the empty office" becomes four shots: a wide establishing frame of the corridor, a medium shot of her hand on the handle, a close-up of her face, and a tracking shot following her through the doorway. This is the step most creators skip, and it is the step that determines whether your final cut feels like a film or a slideshow.

Layer three: continuity governance

Every element that must stay identical across shots gets a locked description: character appearance, wardrobe, props, location layout, time of day, and color grade. The assistant surfaces these as reusable blocks so you never retype a description inconsistently. This is where AI planning tools earn their keep, because consistency errors are the most expensive mistakes in generative video.

Where the human stays in charge

The assistant proposes; you decide. Taste, subtext, performance nuance, and the emotional logic of a cut remain your job. The most common failure mode is accepting generated structure wholesale, which produces competent, forgettable scenes that hit every beat and move nobody.

Building the Script Layer: From Logline to Scene Beats

A script written for AI production is not a lesser script. It is a more disciplined one. Two constraints shape it: runtime economy and visual specificity.

Logline and character card first

Before writing scenes, write one sentence that contains a protagonist, an obstacle, and a stake. Then write a short card for each main character with four fields: want, fear, physical description, and a single behavioral tic. That tic is gold. It gives you something concrete to direct in close-ups.

Example: Mira, 30s, night-shift building inspector. Wants to prove the building is unsafe. Fears being dismissed again. Wears a faded navy work jacket. Tic: taps her pen twice before speaking.

Every shot that features Mira now has a physical anchor and a repeatable action, which means your prompts carry performance information instead of just appearance.

Beat sheets sized to runtime

Work backward from runtime. A 90-second short supports roughly 6 to 9 beats. A 3-minute piece supports 12 to 16. If your beat sheet has more beats than that, you are writing a feature inside a short, and the edit will feel rushed no matter how good the clips look.

A workable short structure:

• Cold open (0:00–0:08): one strong image that poses a question.
• Setup (0:08–0:25): who wants what, and what stands in the way.
• Complication (0:25–0:50): the plan fails or the secret surfaces.
• Turn (0:50–1:10): the protagonist chooses something irreversible.
• Resolution (1:10–1:30): the consequence, held for a beat longer than comfortable.

Write each beat as one or two sentences of action. Dialogue comes later, sparse.

Dialogue that survives generation

Generated lip sync has improved, but long speeches still look unnatural. Write for short lines, off-screen delivery, and reaction shots. A two-second line over a close-up of someone listening reads better than ten seconds of a mouth moving.

Practical rules:

• Keep spoken lines under twelve words where possible.
• Prefer overlaps and interruptions over monologues.
• Convert exposition into action or an object.
• Let the last line of a scene be an image, not a sentence.

Translating a Screenplay into a Shot Plan

This is the technical heart of the workflow. A shot plan is a table, and each row is a generation job.

Anatomy of a usable shot row

Each row should carry: shot number, scene number, size (wide/medium/close), angle, movement, subject action, duration, and continuity keys. Duration matters more than people think. Four to six seconds is a comfortable default for generated clips; anything longer invites distorted motion and forces you to hide flaws in the edit.

A filled row looks like this:

• Shot 12, Scene 4: medium shot, eye level, slow push in, Mira enters the stairwell and stops, 5s, keys: MIRA-NAVY-JACKET, STAIRWELL-NIGHT, GRADE-COOL.

That row is now a prompt blueprint, a scheduling unit, and a continuity checklist in one line.

Coverage patterns that always edit well

You do not need creative coverage on every scene. Use reliable patterns and save invention for the moments that matter.

• Master + two singles: a wide of the space, then a medium on each person in the conversation. Cuts cleanly and hides continuity gaps.
• Insert chain: three to five tight inserts of hands, objects, screens, or feet. Inserts are cheap to generate, easy to fix, and buy you pacing control.
• Reveal build: wide, then detail, then the reaction. Use this when a discovery drives the scene.
• Chase rhythm: alternate a moving wide with a static close-up every two seconds. The contrast creates speed without complex camera work.

Choosing sizes deliberately

A common beginner mistake is making everything a medium shot. Vary scale on purpose: if two consecutive shots are the same size, either cut one or change one. Wide shots establish geography and isolation. Close-ups carry emotion and hide background inconsistency. Medium shots do exposition. A scene that moves wide-to-close as tension rises will feel directed even if the individual clips are simple.

Prompt Design: Writing Instructions the Model Can Follow

Once you have shot rows, prompt writing becomes mechanical — in a good way. Consistency comes from repeating structure, not from repeating adjectives.

The five-slot prompt formula

Build every prompt from five ordered slots:

1. Subject and action — who does what, in present tense.
2. Shot specification — size, angle, movement, lens feel.
3. Environment and time — location, light direction, weather, era.
4. Look and grade — film stock feel, contrast, palette, grain.
5. Constraints — what must not change or appear.

Example: Mira, a woman in her thirties in a faded navy work jacket, pushes open a stairwell door and stops. Medium shot, eye level, slow dolly in, shallow depth of field. Empty concrete stairwell at night, single overhead fluorescent, hard shadows to the right. Cool desaturated grade, fine grain, 24fps motion feel. Keep jacket color and hairstyle consistent; no text, no logos, no extra people.

Same formula, different subject, for the next shot. The repetition is the point: models reward predictable phrasing.

Camera language worth memorizing

You do not need a film school vocabulary, but you do need about fifteen reliable terms. Learn these and use them the same way every time:

• Size: extreme wide, wide, full, medium, medium close, close-up, extreme close-up.
• Angle: eye level, low angle, high angle, overhead, over-the-shoulder, Dutch tilt.
• Movement: static, pan, tilt, dolly in, dolly out, tracking, crane up, handheld.
• Optics: wide lens distortion, normal, telephoto compression, shallow depth of field, deep focus.

Negative prompts and constraint hygiene

Negative constraints do real work. Standard exclusions for narrative work: extra fingers, warped faces, floating objects, sudden lighting shifts, text artifacts, watermarks, camera shake unless requested. Keep the list short and stable. Long negative lists dilute attention and sometimes introduce the very artifacts they name.

Continuity Systems: Locking Characters, Props, and Locations

Consistency is a systems problem, not a prompting problem. If the description of your lead character exists in twelve slightly different phrasings, you will get twelve slightly different people.

Reference sheets

Create one reference sheet per recurring element. A character sheet holds: a locked appearance sentence (reused verbatim), a wardrobe list per scene, a hair and makeup note, and a behavior tic. A prop sheet holds: material, color, size, wear level, and how it is held. A location sheet holds: layout sketch, light sources, wall colors, and floor texture.

Store these as reusable text blocks. When you prompt, paste the locked sentence rather than paraphrasing it. If your tool supports image references alongside text prompts, use a single approved still as the visual anchor and keep your text description frozen.

Wardrobe and grade discipline

Change one variable per scene, not five. If the scene transitions from day to night, keep the wardrobe and grade identical and change only the light. If the character changes jacket between scenes, keep the light and grade identical. Changing multiple variables at once makes it impossible to diagnose what broke continuity.

Lock your grade in writing: cool desaturated with lifted blacks beats moody. Vague words produce vague drift.

A simple continuity audit

Before rendering a batch, scan your shot rows for three things: every character shot references the same locked sentence, no scene mixes grade descriptors, and no prop appears in a shot before its introduction. This five-minute check prevents most reshoots.

Render Workflow: Batching, Testing, and Reviewing

Generating 60 clips is a logistics exercise. Order reduces waste.

Test before you batch

Render one shot per scene first — ideally the hardest one. If the shot with the most motion, the most characters, or the most specific lighting works, the rest of the scene usually will. Fix problems at the prompt level, not in post.

Batch by scene, not by shot type

It is tempting to batch all close-ups together. Resist. Batching by scene keeps environment and lighting context in mind, and it lets you compare takes within a dramatic unit. Batching by location is the second-best grouping.

Versioning and selection

Generate three to five takes per shot, then select. Naming matters: s04_sh12_take03 beats final_final2. Keep a selection column in your shot plan and record which take won and why. The reason is more valuable than the choice, because it teaches you what to change next time.

Editing notes to capture during review

• Clips that are technically fine but emotionally wrong.
• Shots that run long enough to trim for air.
• Anywhere the eye lands on a continuity error first.
• Moments where sound design can carry a weak frame.

Ten Mistakes That Slow Down AI Video Projects

1. Prompting before writing. No shot list means no coverage logic and a chaotic edit.
2. Changing descriptors casually. Every paraphrase is a new character.
3. Uniform shot sizes. Everything becomes a medium shot and the film flattens.
4. Overlong clips. Four to six seconds keeps motion plausible.
5. Too many beats for the runtime. Crowded shorts feel rushed regardless of image quality.
6. Dialogue-heavy scenes. Short lines and reaction shots read better.
7. Vague lighting language. Say where the light comes from.
8. Batching everything at once. You inherit 60 clips of the same mistake.
9. No naming convention. Version chaos costs more time than rendering.
10. Ignoring audio. Foley, room tone, and music mask small visual flaws and lift strong ones.

A Worked Example: 90 Seconds, Four Scenes, Nineteen Shots

Take the Mira logline. Scene 4, the stairwell discovery, needs five shots and about sixteen seconds of screen time.

• Shot 12: medium, slow push in, Mira stops inside the stairwell door. 5s.
• Shot 13: close-up, static, her hand on a cracked concrete step. 3s.
• Shot 14: wide from above, overhead, she crouches and looks up the shaft. 4s.
• Shot 15: over-the-shoulder, static, a rusted support beam with daylight bleeding through. 3s.
• Shot 16: extreme close-up, static, her pen tapping twice against her palm. 2s.

Notice what the plan accomplishes. The pen is planted as a payoff. The scale moves from medium to extreme close, tightening as realization lands. Every shot has one action, which means every prompt has one action. The whole scene is buildable in one sitting, and if shot 14 fails, you know exactly which descriptor to adjust.

Apply the same discipline to the other three scenes, and you have a complete film plan of nineteen shots that share consistent character, light, and grade logic.

Choosing Tools and Building a Stack That Scales

You will likely use several tools rather than one: a writing workspace, a shot planning table, an image generator for keyframes, a video generator for motion, and an editor. Decide based on five criteria.

• Text-to-storyboard support: can it hold a shot list and attach reference images?
• Consistency features: character references, style locks, or seed control.
• Clip length and resolution: match your delivery target before you fall in love with a look.
• Iteration cost and speed: how fast can you test one shot?
• Export hygiene: filename conventions, metadata, and clean downloads matter more than they sound.

The right stack is the one where your shot plan lives next to your prompts. Splitting those two across different apps is the single biggest source of continuity drift.

FAQ

Do I need a finished screenplay before using an AI director assistant?
No. A logline, a character card, and a beat sheet are enough to start shot planning. Finish dialogue once you know how many close-ups you can afford.

How long should each generated clip be?
Four to six seconds is the sweet spot for narrative work. Longer clips are useful for establishing shots with little motion.

How do I stop characters from changing between shots?
Freeze one appearance sentence and reuse it verbatim, keep wardrobe changes isolated per scene, and use a single approved still as a visual reference when your tool supports it.

Is a shot list really necessary for a one-minute video?
Especially for a one-minute video. Short runtimes expose pacing problems immediately, and a shot list is the only reliable way to control pacing across many clips.

What is the fastest way to improve output quality?
Improve your prompt structure, not your prompt adjectives. Ordered slots — subject, shot, environment, look, constraints — raise quality faster than any single magic word.

Can I skip storyboards?
If your shot rows contain size, angle, movement, and duration, you already have a lightweight storyboard. Full frames are helpful for complex staging, not mandatory for simple coverage.

How many takes per shot should I generate?
Three to five. Fewer risks settling for a weak take; more creates a review backlog you will not finish.

Bringing It Together

An AI director assistant does not replace directing. It gives your decisions a place to live so they survive contact with dozens of generated clips. Write the logline and beats. Convert beats into shot rows with size, angle, movement, and duration. Build every prompt from the same five slots. Lock character, wardrobe, prop, and grade descriptions and paste them rather than paraphrasing. Test one shot per scene, batch by scene, and name files before you generate them.

Do that and the technology stops being a slot machine. You get a repeatable pipeline where a 90-second short is a plan you execute rather than a pile of clips you hope to assemble — and where the next project starts from a system instead of a blank page.