A script is not a film. Between the page and the first usable frame sits a long chain of decisions: how many shots a scene needs, where the camera sits, which way a character faces, what the light is doing, and how the cut lands. Most AI video workflows skip that chain entirely and jump straight to prompting, which is why so many generations look stunning for three seconds and fall apart by the fourth shot.
An AI director assistant closes that gap. Instead of treating generation as a slot machine, you treat it as a shoot. You break the script into scenes, design the shots, lock the references, and only then spend compute on output. This guide walks through that process end to end: story bible, shot design, prompt grammar, continuity control, sound planning, and the mistakes that quietly ruin otherwise good projects.
Why an AI Director Assistant Changes Pre-Production
Traditional pre-production is expensive because every decision requires people, permits, and gear. A shot list costs time; a location scout costs money; a reshoot costs both. That economic pressure is why filmmaking developed such disciplined paperwork in the first place: storyboards, shot lists, continuity logs, call sheets. The paperwork exists to prevent waste.
With generative video, waste takes a different shape. Rendering is cheap relative to a film crew, but it is not free, and it is far from instant. A 10-second clip that almost works still costs you a queue slot, a review pass, and a decision. Multiply that across a 40-shot sequence and the real bottleneck becomes iteration discipline, not raw model capability.
An AI director assistant addresses the bottleneck by moving decisions upstream. It helps you define character identity, scene geography, and shot grammar before a single frame renders. When those three things are locked, generation becomes a production task rather than an experiment. You stop asking "does this look good?" and start asking "does this match the plan?" — a much faster question to answer.
The practical result is fewer abandoned sequences. Teams that plan shots before generating typically reuse more footage, re-render less, and finish edits with fewer "patch" shots that exist only to hide a continuity break.
Turn the Script Into Shootable Scenes
A screenplay written for humans is under-specified for a model. "Sarah waits nervously in the diner" gives a model almost nothing about framing, blocking, or duration. Your first job with an AI director assistant is translation: convert narrative intent into visual intent.
Break beats into shots
Read each scene and mark the beats — moments where something changes. A character enters. A line lands. A look shifts. An object is noticed. Each beat is a candidate for a shot, though not every beat deserves one. A useful rule: if the audience needs to notice something new, you need a new shot. If nothing changes, hold the shot.
Write each shot as a single sentence that combines subject, action, and framing: "Medium shot, Mara slides the envelope across the counter, camera slightly low, diner neon behind her." That sentence is your prompt skeleton and your editing note at the same time.
Assign shot intent, not just description
Every shot should have a job. Establish geography. Reveal emotion. Deliver information. Transition. Shots with no job get cut in the edit, so don't generate them. Marking intent also helps you choose duration: an establishing shot might need three seconds, a reaction shot one.
Write continuity notes as you go
This is the step most people skip. Keep a running log per scene: which side of the frame a character occupies, what they are holding, what they are wearing, whether a jacket is on or off, which window the light comes from. When you generate shot 14 three days after shot 13, that log is the only thing standing between a coherent scene and a costume change mid-conversation.
Build a Story Bible Before You Generate Anything
A story bible is a small set of canonical documents. It does not need to be elaborate, but it does need to exist outside your head, because models have no memory of your intent — only of the text and images you supply.
Character sheets that survive model changes
For each recurring character, collect: a front-facing portrait, a three-quarter view, a profile, and a full-body reference. Add a short written descriptor that avoids subjective adjectives. "Late thirties, square jaw, short dark hair with a grey streak at the left temple, olive jacket with brass buttons" beats "handsome, striking, cinematic" every time.
Keep the descriptor in a text file and paste it verbatim into every prompt that includes that character. Small paraphrases drift, and drift compounds across a sequence.
Location and palette locks
Define each location with two or three reference stills and a lighting rule. "Warehouse interior, single overhead sodium lamp, deep shadow on the left wall, wet concrete floor" is a lighting rule. It tells you where highlights and shadows belong, which is what keeps consecutive shots feeling like the same room.
Lock a palette too — three to five colors, written as plain words. A restrained palette makes visually different shots read as one film.
Prop and wardrobe inventory
List anything a character touches, wears, or carries, with one reference image each. Props are the most common source of continuity failure because they are small, and models happily reinvent small things. A locked prop sheet lets you reference the same object across shots instead of describing it from scratch.
Write Prompts Like a Director, Not a Search Query
Search queries retrieve; direction specifies. A prompt that reads like a wish list produces a lottery result. A prompt that reads like a shot order produces a take you can use.
The four-part shot prompt
Build every prompt from four blocks, in this order:
- Subject block — the canonical character descriptor plus wardrobe and prop state.
- Action block — one clear physical action in present tense. Avoid two actions; models blur them.
- Camera block — shot size, angle, lens feel, and movement.
- Environment block — location, time of day, lighting rule, atmosphere.
Keeping the order stable makes prompts easier to debug. When a result is wrong, you can usually trace it to one block instead of rewriting everything.
Camera vocabulary worth memorizing
Terms that reliably change output: extreme wide, wide, medium wide, medium, medium close-up, close-up, extreme close-up; low angle, eye level, high angle, dutch tilt; slow push in, pull back, orbit, handheld follow, static lock-off. Pair one size with one movement and stop there. Stacking "slow push in while orbiting with a handheld drift" gives you mush.
Lens language helps too. Shallow depth of field, wide-angle distortion, telephoto compression — each signals a different visual tradition, and models respond to those cues.
Constraints do real work
Negative phrasing is weaker than positive specification. Instead of "no extra people," write "empty street, no pedestrians visible" or better, specify the frame: "two figures only, centered, wide empty sidewalk on both sides." Describing the desired state outperforms forbidding the unwanted one.
Finally, add a technical tail when the tool supports it: aspect ratio, frame rate feel, grain level, and whether motion should be smooth or slightly handheld. Consistency here is what lets you cut shots from different sessions together.
Keep Visual Consistency Across Shots
Consistency is not one technique; it is a stack of small controls applied in the same order every time.
Reference frames beat descriptions
If your tool supports image-to-video or frame referencing, use it. Generate a still of the exact composition you want — correct character, correct framing, correct light — and animate from that still. Descriptions approximate; references pin.
When generating the still, work in passes: composition first, then character accuracy, then lighting. Trying to fix all three at once usually regresses one of them.
Wardrobe, time of day, and blocking
Three continuity dimensions cause most visible breaks:
- Wardrobe state. Jacket on or off, sleeves rolled, tie loosened. Track it per shot in the log.
- Time of day. Morning light and dusk light are different films. If a scene spans an hour, decide whether the light changes or stays fixed, and write it down.
- Blocking. Who stands where relative to the door, the window, the counter. Blocking errors read as spatial confusion, which audiences notice instantly even when they cannot name it.
Group scenes and hands
Crowds are hardest. Reduce count where possible: two characters in focus with soft background figures usually reads better than five rendered faces. For hands — holding, reaching, writing — keep the action simple, frame closer, and avoid complex finger articulation. If a shot needs a delicate hand action, consider generating it as a still and animating subtly, or cutting away before the action completes.
Plan Sound and the Edit Before You Render
Audio decisions change visual decisions, not the other way around. A scene cut to dialogue has different shot lengths than a scene cut to score.
Build a scratch track first
Before generating anything, build a rough audio spine: temp dialogue read on your phone, a music bed, and rough timing. Twenty minutes of work here saves hours of re-rendering because shots will already be the right length.
Cut for rhythm, not coverage
Editors of AI-generated footage fall into the coverage trap — generating many angles and cutting between them because they exist. Better to decide the rhythm first: long hold, two quick cuts, long hold. Then generate only the shots that rhythm requires. Fewer, longer shots also hide consistency gaps, because the audience has less time to compare.
Match audio continuity
Room tone is the invisible thread. If your tool offers ambience control, keep it consistent per location; if not, add a continuous ambience layer in the edit under all shots from that location. Sudden silence between shots is more jarring than a slightly wrong-looking frame.
A Repeatable Workflow From Idea to Final Cut
- Write or adapt the script at scene level, with clear beats.
- Build the story bible: character sheets, location rules, palette, prop inventory.
- Break each scene into shots with one-sentence shot lines and stated intent.
- Generate reference stills for every shot before animating any of them.
- Animate from stills where supported, using the four-part prompt to describe motion.
- Log continuity after each approved shot: wardrobe, blocking, light, props.
- Assemble a rough cut with the scratch track, then identify missing or broken shots.
- Regenerate selectively — only the shots that fail, with one variable changed at a time.
- Finish: color consistency pass, ambience layer, titles, export at target aspect ratios.
Step eight is where discipline pays off. Changing one variable per attempt turns troubleshooting into a controlled experiment instead of a guessing game.
Common Mistakes and How to Avoid Them
Prompting before planning. Generating a hero shot first feels productive and is usually wasted. Plan the sequence, then generate in order.
Paraphrasing character descriptions. Every synonym changes the face slightly. Copy and paste.
Overloading action. Two actions in one shot produce an averaged blur. Split them into two shots.
Ignoring aspect ratio until the end. Vertical and horizontal framing demand different compositions. Decide the delivery format before shot design, not after.
Generating too many variations. Ten options create ten decisions. Generate two or three, choose on stated criteria.
No continuity log. This is the single most expensive omission on multi-scene projects.
Chasing realism over readability. A slightly stylized look with consistent geography beats photoreal confusion every time.
Choosing Tools Without Locking Yourself In
Model capabilities shift quickly, so optimize for portability. Keep your prompts, references, and story bible in plain files and folders you control, not inside a single interface. When a new model handles hands or crowds better, you want to switch without rebuilding your project.
| Decision | Prefer | Avoid |
|---|---|---|
| Character fidelity | Tools with image referencing | Text-only models for leads |
| Long sequences | Consistent prompt grammar | Per-shot improvisation |
| Fast iteration | Low-resolution drafts first | Full-quality first passes |
| Portability | Plain-text prompts and local assets | Locked-in proprietary formats |
| Audio | Separate ambience and music layers | Baking sound into clips |
A useful test: can you hand your shot list and references to a collaborator and get recognizable results back? If yes, your system is portable. If no, your project lives in one tool's quirks.
Scaling to Episodic and Series Work
Once a single scene works, repetition becomes the challenge. Series work rewards templates: a fixed opening shot type, a consistent title treatment, a recurring establishing angle for each location. Viewers read those repetitions as style.
Keep a project-level style sheet — palette, grain, lens feel, pacing rule — and apply it to every episode. Build a shot library of reusable establishing shots and transitions, and refresh it gradually rather than per episode. And treat each episode's continuity log as a document you can search, because the question "which jacket was she wearing in the diner scene?" will come back.
FAQ
Do I need a storyboard artist?
No. Reference stills generated from your shot lines serve the same function. The discipline matters more than the drawing skill.
How many shots should a one-minute scene have?
Typically eight to fifteen depending on dialogue density and pacing. Count beats first, then assign shots to beats.
Why do my characters change between shots?
Almost always because the written descriptor drifted or the reference image changed. Freeze both and re-run.
Should I generate video directly or animate from a still?
Animate from a still when consistency matters. Direct generation is fine for establishing shots, textures, and montage material.
How do I fix a shot that is 90% right?
Change one variable: either the reference, the action wording, or the camera block. Re-render and compare against the log, not against memory.
What about vertical delivery?
Design vertically from the start. Close-ups and medium shots dominate, negative space goes above and below the subject, and wide establishing shots lose value fast.
How long should a project's story bible be?
One page per character and a single page for locations, palette, and props is usually enough for a short film or a small series.
Can I skip the scratch track?
You can, but expect longer shots than you need. Audio timing is the cheapest pacing tool available.
An AI director assistant does not replace taste. It replaces guesswork. Give it a script, a story bible, and a shot list, and it becomes the part of the process that holds everything together while you focus on the decisions only a director can make.
