AI Director Workflow: From Script to Storyboard to Screen

A practical AI director workflow for turning a script into a shot list and storyboard, with prompt recipes, continuity checks, and troubleshooting.

Most videos do not fail at the render stage. They fail earlier — when the idea lives only in someone's head, the script is half-formed, and nobody can say what shot fourteen is supposed to look like. A generative model will happily hand you a gorgeous clip that has nothing to do with the story you meant to tell. What it will not hand you is a plan.

That is the gap an AI director workflow fills. Not a button that makes films, but a repeatable process: shape a script, break it into shots, lock a look, approve still frames, and only then let anything move. Below is that process end to end, including the documents, prompt patterns, and continuity checks that keep a sequence from falling apart.

Why the Bottleneck Moved From Rendering to Directing

Professional video has always been a chain of narrowing decisions: what story, what scene, what shot, what frame. Generative tools collapsed the last link in that chain. A frame that once required a camera, a crew, and a location can now be produced from a sentence in under a minute.

That speed does not remove the old bottleneck. It exposes it. If your script is vague, you generate vague footage faster. If your shot plan is missing, you collect a pile of beautiful clips that refuse to cut together. If your visual language is undefined, every shot looks like it wandered in from a different film.

Experienced AI video creators tend to settle into a ratio that surprises newcomers: roughly two thirds of the time goes into preparation and one third into generation. That is not because generation is hard. It is because a locked script and an approved board make generation almost mechanical, while an unlocked one turns it into a slot machine.

The practical consequences are easy to see:

Fewer wasted generations. A shot you can describe precisely usually takes one or two attempts, not twenty.
Real continuity. Characters, wardrobe, props, and lighting stay consistent when they are defined in a document before they are defined in pixels.
Faster revisions. When a client asks for "darker and colder," you change one line of the style bible, not a hundred prompts.
Better decisions. Comparing frames side by side is how you notice that your climax has no reaction shot.

What an AI Director Workflow Actually Handles

It helps to separate the work into three layers. Most confusion in AI filmmaking comes from trying to solve a layer-three problem with a layer-one tool.

Layer one: story

Structure, beats, scene order, emotional arc. Language models are genuinely useful brainstorming partners here — they can propose three ways to escalate a conflict, compress a scene that runs long, or suggest the ending you have been circling for an hour. They should never be the final authority. Structure is taste, and taste is what you bring.

Layer two: the shot plan

Coverage. How a scene splits into individual shots, with size, angle, movement, and duration. Shot grammar is largely rule-based, which makes it a good fit for automation. A two-person dialogue scene almost always wants a master, a two-shot, and singles. Generate a first pass, then edit it like a producer: delete, merge, reorder.

Layer three: the frame

Composition, lens feel, lighting, palette, wardrobe, atmosphere. Here the model's job is to offer options and your job is to choose. The choosing is the directing.

What should stay human: final story judgment, tone, performance nuance, and anything involving a real brand, a real person, or licensed property. Those need a person in the loop, both creatively and legally.

A quick sanity check before you move on: if you cannot describe your video in one sentence and sketch its five key frames on paper, you are not ready to generate. You are ready to plan.

Step One: Write a Script a Camera Could Photograph

A shootable script contains only what can be seen and heard. If a line cannot be photographed, it does not belong in the script; it belongs in your notes.

Start from a one-sentence logline

Format: a specific protagonist must take a specific action before a specific consequence, in a specific world.

Weak: "A woman learns about the power of friendship."

Shootable: "A night-shift lighthouse keeper must relight a dead beacon before a storm drives a supply ship onto the rocks."

The second version hands you the location, the time of day, the key prop, the antagonist, and the final image. You can storyboard it immediately.

Structure in beats, not pages

For a sixty to ninety second piece you rarely need more than five to seven beats:

Establish — the world and the protagonist's ordinary state.
Disruption — the thing that changes.
Decision — the protagonist commits.
Escalation — the obstacle gets worse.
Reversal — the plan fails, or the cost appears.
Climax — the decisive action.
Resolution — the new normal, in one image.

Each beat becomes one to three scenes. Each scene becomes three to eight shots. That arithmetic keeps a short film tight and a product spot coherent.

Treat dialogue as a last resort

Synthesized dialogue and lip sync remain the least predictable part of AI video. Where you can, write scenes that work silently: a decision made by a hand closing a door, a threat delivered through frosted glass, a reunion shown through a dropped bag. Silent beats are cheaper, more cinematic, and far more reliable.

For narrated pieces — explainers, documentaries, ads — the voiceover script is the timeline. Write it with timestamps first, then build shots to match, not the other way around.

Step Two: Break the Script Into a Shot List

A shot list is the highest-leverage document in the whole workflow. It converts language into production instructions, and it converts a mood into something you can check off.

Use one consistent table

#	Scene	Size	Angle	Movement	Description	Duration	Notes
1	Lighthouse, dusk	Extreme wide	Level	Slow push	Dead beacon, storm on the horizon	4s	Establishing, cold palette
2	Lamp room	Medium	Low	Static	Keeper strikes a match	3s	Warm key, dust in the air
3	Insert	Close-up	Level	Handheld drift	The match dies between her fingers	2s	No face in frame
4	Exterior	Wide	High	Slow tilt	Waves breach the breakwater	3s	Scale, danger

Four columns of that table are enough to generate every frame, and enough for an editor to assemble a rough cut.

Reuse coverage patterns instead of inventing grammar

Dialogue: master, over-the-shoulder, singles, then a cutaway for breathing room.
Action: wide geography, tighter detail, impact insert, reaction.
Reveal: obstructed frame, partial reveal, full reveal, character reaction.
Montage: three to five short shots that share one motion direction.
Product: hero frame, detail macro, in-use shot, result shot, logo frame.

If shot grammar is new to you, starting from a finished structure beats staring at a blank page. Browsing Orelon's templates is a fast way to see how a complete sequence is broken down before you build your own.

Flag the shots you know will be difficult

Mark anything involving hands doing precise work, crowds, on-screen text, animals, reflections, or a character speaking in close-up. Those are the shots where generation is least predictable. Either simplify them in the script — a hand entering frame is easier than a hand tying a knot — or budget extra attempts in advance.

Step Three: Lock a Style Bible Before You Generate a Frame

The style bible is a half-page document that fixes the look so every later prompt can reference it instead of re-describing it. It is the difference between a sequence and a collection.

The seven lines that matter

Palette: three colors plus one accent. "Desaturated slate blue, wet asphalt grey, a single sodium-orange accent."
Light: direction, quality, source. "Low sun from frame left, hard shadows, practical lamps visible in frame."
Lens: focal length and depth. "35mm equivalent, shallow depth of field, mild barrel distortion."
Texture: grain, contrast, cleanliness. "Fine 35mm grain, lifted blacks, soft highlight roll-off."
Camera behavior: "Slow deliberate pushes, no whip pans, tripod or gimbal only."
World rules: period, technology level, weather, wardrobe palette.
Negative list: what must never appear — modern signage, glossy plastic, heavy lens flares, specific clothing.

Generate reference frames, not only descriptions

Words drift between models and even between prompts on the same model. Reference images do not. Build three to five anchor frames that define the look — one wide, one medium, one close-up, one night exterior, one interior — and reuse them throughout the project. Orelon's image generator suits this kind of iterative frame design: generate variations, compare them side by side, refine, then lock the one you want as a visual reference.

Step Four: Storyboard Stills Before You Storyboard Motion

Here is the workflow decision that saves the most time: generate stills before you generate motion.

Motion multiplies everything — time, inconsistency, and the cost of a mistake. A still frame takes a fraction of the effort and lets you judge composition, casting, wardrobe, and lighting with a clear eye. Only when the frames work as a sequence should you animate them.

The image-first storyboard loop

Generate the establishing frame for a scene.
Approve composition and light before moving on.
Generate the scene's remaining shots using the approved frame plus the style bible.
Lay the frames out in order and read them like a comic strip.
Fix any frame that breaks the read: a missing eyeline, a wardrobe jump, a sudden palette shift.
Only then move to motion.

Step four is where projects get rescued. A row of stills tells you in thirty seconds whether your story reads. A row of animated clips hides that information behind movement, and redoing it is far more expensive.

Continuity checks on the board

Eyeline: characters look in consistent directions across a conversation.
Screen direction: anyone who exits frame right should enter the next shot from frame left.
Wardrobe and props: the same jacket, the same mug, the same amount of mud.
Time of day: light direction and color temperature agree across shots in one scene.
Scale: a close-up of a hand and a wide of a building should feel like the same world.
Motion logic: if the camera pushes in during shot six, does a pull-out in shot seven feel motivated?

Step Five: Animate Approved Frames and Cut a Previs Edit

Animated storyboards — previs — are the bridge between stills and a finished piece. The goal is not final quality. The goal is to prove the edit works.

Describe motion, not just subject

A strong image-to-video prompt has three parts: what is in frame, what moves, and how the camera behaves. Note that the camera instruction is as specific as the subject instruction. "Cinematic" is not a camera instruction.

Lighthouse lamp room, dawn light through salt-streaked glass. The keeper's hand lifts a match to the wick; the flame catches and steadies. Slow push in, 35mm, shallow focus, fine grain, desaturated with one warm accent.

Keep clips short, then assemble

Generate three to five second clips. Short clips are easier to control, easier to regenerate, and they cut together better than a single long take. Build the edit around shot rhythm:

Wide shots hold longer, four to six seconds.
Medium shots, three to four seconds.
Close-ups and inserts, one and a half to three seconds.
Cut on action to hide transitions cleanly.

Orelon's video creation workspace lets you move from approved frames to motion without leaving the project, so the style bible, references, and shot order stay in one place.

Cut to a temporary track

Lay a scratch music bed and rough sound effects before you polish visuals. Most rhythm problems that feel visual are actually audio problems. If a cut feels wrong, try moving it two frames earlier before regenerating the shot.

Prompt Patterns, Mistakes, and Where to Automate

Once the style bible exists, prompt writing becomes assembly rather than invention. The reliable pattern is: subject, action, environment, light, lens, style, negative.

Establishing shot

Abandoned coastal road at dusk, cracked asphalt, a single ruined guardrail. Wind-driven grass. Low sun from frame left, long shadows, wet air. 24mm wide, deep focus, muted teal and grey with rust accents. Fine grain, anamorphic softness. No cars, no people, no signage.

Character close-up

Weathered keeper in her sixties, wool coat, close-up at eye level. She exhales slowly; a strand of hair moves. Hard practical light from frame right, deep shadow on the left cheek. 85mm, shallow depth of field, natural skin texture. Muted palette, no makeup gloss, no modern clothing.

Action insert

Hands striking a match against a brass striker, macro scale. The match flares and steadies; ember glow spreads across the fingertips. Single warm source, everything else near black. 100mm macro, extreme shallow focus, fine grain. No visible face, no rings, no logos.

Save the prompts that work as reusable blocks — "night exterior," "dialogue close-up," "product hero" — and your next project starts half finished. Further tested structures are organized in Orelon's prompt library by scene type.

Mistakes that cost the most time

Starting with motion. Animating before the stills work multiplies every mistake. Fix the frame first.
Rewriting the story inside a prompt. If your prompt is a paragraph of plot, the model will ignore half of it. Prompts describe frames; scripts describe stories.
Changing style mid-project. A new palette in shot twelve reads as an error, not a choice. Lock the bible and version it deliberately.
Over-covering. Twelve shots of the same beat is indecision, not coverage. Three strong shots beat twelve average ones.
Ignoring screen direction. Reversed movement makes viewers feel lost without knowing why.
Chasing realism over clarity. A slightly stylized look that reads instantly beats a photorealistic frame nobody can parse.
Skipping silence. Constant music and constant motion exhaust an audience. Let a shot breathe.
Treating generation as free. Every attempt costs time, so design shots likely to succeed: fewer characters, simpler hands, clearer light.
Storyboarding in your head. If the sequence is not written down, you will rediscover the same problems at a much higher price.

Automate the options, own the selection

Automate decisions that are rule-based: shot lists from scene descriptions, coverage patterns, palette variations, aspect-ratio versions, thumbnails, alternate cuts. Direct by hand those that are taste-based: the ending, the performance beat that carries the theme, the frame that becomes the poster, the order of the final two shots.

A useful rule: generate twenty variations, but choose deliberately and be able to say why. That single habit is the difference between work that looks generated and work that looks directed.

FAQ: Scripts, Storyboards, and Continuity

Do I need a full screenplay before storyboarding? No. A logline, five to seven beats, and a clear final image are enough. Anything more is a bonus, not a prerequisite.

How many shots should a one-minute video have? Usually twelve to twenty. Fewer feels static; more feels frantic unless the piece is intentionally fast-cut.

How do I keep a character consistent across shots? Lock a reference frame for the character, describe wardrobe and features identically every time, keep lighting direction constant within a scene, and regenerate from the approved frame rather than from the text prompt alone.

Should prompts be written per shot or per scene? Per shot, always. Scene-level prompts produce averages, and averages look like nothing in particular.

What if the generated board does not match what I imagined? That is normal and useful. A board is a negotiation between what you want and what reads clearly. Adjust the shot, not just the wording.

Can I storyboard without any image generation? Yes. Sketches, photo references, and stock frames all work. The point is to see the sequence before committing to motion.

How long should preproduction take? For a sixty-second piece, one to three hours of script and shot planning typically saves five to ten hours of regeneration and re-editing.

How do I handle aspect ratios and cutdowns? Design for the widest frame you will deliver, then reframe for vertical or square versions with the subject centered and key action inside a safe area. Generating once and reframing beats generating twice.

When should I stop iterating on a frame? When further changes no longer change the story. If a frame is ninety percent right and the sequence reads, move on. Perfection at board stage is a form of procrastination.

Bring Your Next Sequence to Life in Orelon

The director's job has not changed: decide what the audience sees, in what order, and why. What changed is how quickly those decisions can be tested and revised. Write the logline, build the shot list, lock the style bible, approve the frames, and only then let the camera move.

Orelon is built for that rhythm — cinematic ideas in motion, with image and video generation sharing one workspace so your board and your final cut stay in sync. Start with a still, prove the sequence, then bring it to life in the Orelon video generator. Your first frame is the only decision that cannot be automated.