Start With the Story, Not the Software
Most filmmakers who try AI for the first time make the same mistake: they open a video generator, type a dramatic prompt, and wait for something magical. What comes back usually looks impressive for three seconds and means nothing for ninety. The problem is not the model. The problem is that they skipped the part of filmmaking that AI is actually best at accelerating — the thinking stage.
Story development and shot design are the two places where artificial intelligence genuinely changes the economics of production. A treatment that used to take three weeks of note-passing can be drafted in an afternoon. A shot list that used to require a storyboard artist, a location scout, and a printed deck can be visualised before anyone books a camera. But none of that happens automatically. The quality of the output tracks the quality of the structure you feed in.
This guide walks through a complete pre-production workflow: building a story spine, mapping character arcs, generating a world, translating beats into shots, defining camera language, and holding visual continuity across dozens of frames. It assumes you are a filmmaker, a content creator, an animation director, or a solo video producer who wants a repeatable process rather than a pile of disconnected tools.
What AI Story Development Tools Actually Do Well
Before assigning work to any tool, it helps to separate what these systems are genuinely good at from what they merely appear to do.
Structural pattern recognition
AI models are excellent at recognising and reproducing narrative shape. Feed them a logline and a genre, and they can propose a three-act structure, a five-beat sequence, or a seven-point arc that follows established conventions. This is not originality — it is architecture. Treat the output as scaffolding, not as a finished script. The value is that you get a full skeleton in minutes, and you can immediately see which act sags.
Contradiction and continuity flagging
Ask a model to read your own draft and list every moment where a character behaves inconsistently, and it will often find problems you have read past twenty times. It does not understand your protagonist the way you do, but it can spot that a character who refuses to lie in act one is casually deceiving someone in act three without comment.
Volume generation for selection
Where AI shines is producing twenty options where you would previously have written three. Twenty possible opening images. Twenty ways a confrontation could end. Twenty names for a location. You are not outsourcing taste; you are outsourcing the blank page.
What it does poorly
AI is weak at subtext, at culturally specific humour, at silence, and at knowing when a scene should simply end. It also tends to flatten tone toward the generic middle unless you actively push it. Plan to spend as much time editing AI-generated material as you would spend editing a first draft from a human co-writer — sometimes more.
Worldbuilding and reference generation
Visual reference is one of the most useful and most overlooked applications. Instead of describing a rain-soaked coastal town in a paragraph, you can generate a dozen mood frames, pick three, and attach them to your shot list. Concept artists still earn their keep, but the conversation with them starts from an image instead of a mood board of borrowed stills.
Building a Story Spine Your Shot List Can Follow
A story spine is the smallest set of statements that fully describes your film. If your spine is vague, every downstream decision — lens choice, location, pacing, even colour — becomes a guess.
Write your spine in seven lines:
- A character who wants something.
- Is blocked by a specific obstacle.
- So they take an action that costs them something.
- Which produces an unexpected consequence.
- Forcing a decision they have avoided.
- Leading to a confrontation or revelation.
- Resolving in a way that changes them or their world.
That is roughly thirty seconds of writing and it will save you weeks. Once written, paste it into your AI assistant along with a genre, a target runtime, and a tone reference, and ask for a beat sheet with timestamps. Do not ask for a script. Ask for beats.
A practical tip: request beats in terms of change, not event. "Maya learns the truth" is an event. "Maya learns the truth and immediately lies about knowing it" is a change. Change is what generates shots, because a shot exists to capture a shift in information, emotion, or power.
Once you have your beats, annotate each one with three columns: what the audience knows, what the character knows, and what the gap between them is. That third column becomes your tension map and your editing rhythm. Scenes with a large gap between audience and character knowledge tend to be suspenseful; scenes with no gap tend to be flat regardless of how well they are shot.
Finally, set a hard runtime budget before you design anything. A three-minute short with fourteen beats will feel rushed. Four to six beats is usually the honest number. Knowing this early prevents the most common AI-assisted production disaster: a beautiful shot list for a film that cannot exist.
From Beats to Shots: Designing the Shot List
This is where the workflow becomes concrete. Take each beat and ask one question: what is the single most important piece of visual information in this moment?
One beat, one primary shot
For every beat, designate a primary shot that could carry the scene alone if everything else were cut. Then add supporting coverage: an establishing frame, a reaction, an insert, a transition. A useful ratio for short-form work is one primary shot plus two to four supports per beat. That keeps a six-beat short in the eighteen-to-thirty-shot range, which is realistic for a small team.
Write shot descriptions as behaviour, not adjectives
Weak: "beautiful wide shot of the city at dusk."
Strong: "Wide from rooftop level, camera static, a single lit window in the left third; a figure crosses it and stops."
The second version contains staging, camera behaviour, and a visual event. It is also directly usable by a storyboard artist, an AI image generator, or a cinematographer. Behavioural descriptions survive translation between tools; adjective piles do not.
Group shots into sequences, not lists
Once you have thirty shots, group them into sequences of four to seven shots that share a location, time of day, or emotional register. Sequences are the unit you will actually shoot, so designing them as units prevents the classic problem of a shot list that is technically complete but practically unshootable because it bounces between six locations in ninety seconds.
Mark the shots you can actually get
Add a column for feasibility: practical, generated, or hybrid. Practical means you can shoot it with a phone or a camera you own. Generated means it will come from an AI video tool or animation pipeline. Hybrid means a real plate composited with generated elements. Labelling these early stops you from discovering in post that a third of your film requires an effect you have no way to produce.
Camera Language: Lenses, Movement, and Framing
AI video tools respond well to cinematographic language once you learn which parts of it they understand. Vague terms produce generic results; specific terms produce recognisable intent.
Focal length and perspective
State focal length in millimetres. "35mm, eye level, medium shot" produces a different result than "85mm, chest-up, compressed background." Wide lenses exaggerate space and distance; long lenses compress and isolate. If you want a character to feel trapped, a long lens with a busy background does more than any dialogue line.
Camera movement vocabulary
Use established terms: static, slow push in, pull back, dolly left, handheld follow, crane up, whip pan, orbit. Add speed qualifiers — slow, deliberate, snap. Most generators handle one movement per shot well and two movements poorly. If a shot needs both a push and a tilt, consider whether it should be two shots.
Framing and headroom
Specify framing explicitly: extreme wide, wide, medium wide, medium, medium close-up, close-up, extreme close-up. Also specify subject placement — left third, centre, low in frame — because generators tend to default to centred subjects unless told otherwise, and centred framing everywhere reads as flat and amateurish over a full sequence.
Lighting and time of day
Lighting is where specificity pays off fastest. "Golden hour backlight with lens flare" and "overcast noon, flat and cold" produce genuinely different films. Build a small lighting palette for your project — three or four setups — and reuse it. Consistency of light reads as authorship.
Build a reusable prompt block
Rather than writing camera language from scratch for every shot, maintain a reusable block that you paste and then modify:
[focal length], [framing], [camera movement], [lighting], [time of day], [colour palette], [film stock or texture reference], [subject action]
Keep the first seven elements nearly identical across a sequence and change only the last. This single habit does more for visual consistency than any post-production filter.
Keeping Visual Continuity Across Dozens of Shots
Continuity is the hardest problem in AI-assisted filmmaking, and it is solved with documentation rather than with better models.
Create a continuity bible
One document containing: character appearance descriptions written in consistent language, wardrobe for each act, location layouts with compass directions, key props and their positions, and the lighting palette. Every prompt you write should be traceable to an entry in this document.
Use reference images, not adjectives
"Weathered fisherman in his sixties" will drift shot to shot. A specific reference image or a locked character description with fixed physical details — scar on left cheek, grey beard cropped short, olive jacket with a torn right cuff — will hold much better. Three identifying details per character is usually the sweet spot: enough to anchor, few enough to survive generation.
Track screen direction
If a character exits frame right, they should enter the next shot from the left for the geography to read correctly. This is basic film grammar and it is the single most common continuity error in AI-generated sequences, because each shot is produced independently. Keep an eye-line and direction column in your shot list and check it before generating anything.
Generate in blocks, not one at a time
When possible, generate an entire sequence in one session with identical style parameters. Tools drift slightly between sessions and across prompt edits; keeping a sequence together reduces the chance that shot four looks like it belongs to a different film.
Accept a controlled amount of variance
Perfect continuity is not the goal — legibility is. Minor variations in background detail read as texture to an audience. Variations in a character's face read as a mistake. Prioritise consistency where the audience is looking: faces, hands, wardrobe silhouette, and the direction of movement.
A Full Walkthrough: 90-Second Short Film
Here is the workflow end to end for a hypothetical short called The Last Delivery.
Day one — spine. Seven lines: a courier must deliver a package after the city has shut down; the recipient's address no longer exists; she opens the package to find a key; the key opens a door she has been avoiding; she goes in and finds her brother's belongings; she decides to leave them and keep the key; she walks out into daylight. Runtime budget: ninety seconds, six beats.
Day two — beats and tension map. Six beats, each annotated with audience knowledge, character knowledge, and the gap. The largest gap sits in beat four, where the audience suspects the address before she does. That beat gets the most coverage.
Day three — shot list. Twenty-two shots across six sequences. Primary shots: six. Supports: sixteen. Feasibility: fourteen practical interior and street shots, six generated establishing and transitional frames, two hybrid.
Day four — camera language. Two lighting palettes: sodium streetlight with wet reflections for exteriors, single-window daylight for interiors. Lenses: 24mm for exteriors, 50mm for interiors, 85mm for two close-ups in beat four. Camera primarily handheld with two deliberate static frames to punctuate the ending.
Day five — references. Twelve generated mood frames, three kept. One character reference sheet with three identifying details.
Day six — shot generation. Six generated frames produced in one session with locked parameters. Two rejected for screen-direction errors. Replaced and re-generated in the same session.
Day seven — assembly and review. A rough edit using still frames on a timeline with temp audio. This reveals pacing problems before any footage exists — which is the entire point of the exercise. One beat is cut. The film is now eighty-two seconds and better.
Total pre-production: seven days, one person, no budget spent on a camera crew.
Choosing the Right Tool for Each Stage
Different stages need different capabilities, and the most common mistake is asking one tool to do everything.
For story structure and beats: a general-purpose language assistant with strong long-context handling. You need something that can hold an entire treatment in mind and answer questions about it.
For world and mood reference: an image generator with consistent style conditioning and reference-image support. Speed matters more than fidelity here — you will discard most of what you make.
For shot visualisation: a storyboard-oriented tool that lets you attach camera metadata to each frame, or a plain document plus an image generator if you prefer control.
For moving footage: a video generator that accepts image-to-video input, so you can drive motion from your selected reference frames rather than from text alone. Text-to-video is for exploration; image-to-video is for production.
For continuity tracking: a spreadsheet. Genuinely. A structured sheet with columns for shot number, sequence, framing, lens, movement, lighting, screen direction, and status will outperform any specialised app you can find, because you will actually maintain it.
Decision criteria, in order of importance: does it accept reference images, does it let you lock style parameters, does it output at a resolution and aspect ratio you can edit with, and does it cost you less than the time it saves. If a tool fails the last test, drop it — novelty is not a workflow.
Mistakes That Undermine AI-Assisted Pre-Production
Generating before structuring. The most expensive error. Every hour spent clarifying the spine saves roughly a day of regeneration.
Treating a shot list as a wish list. If it cannot be shot or generated with the resources you have, it is a fantasy. Mark feasibility early and cut ruthlessly.
Over-prompting. Long prompts with fifteen aesthetic adjectives produce mush. Six to ten specific elements is the practical ceiling; beyond that, elements start cancelling each other out.
Chasing perfect faces across shots. Spend your effort on staging and light instead. Audiences forgive slight facial drift in a fast cut; they do not forgive a sequence that makes no spatial sense.
Ignoring sound during design. Design shots with sound in mind. If a sequence is carried by one continuous ambient bed, that affects how many cuts you can make. Pre-production that ignores audio produces assemblies that fight themselves.
Skipping the animatic. Always cut your frames together with temp audio before committing to production. It is the cheapest possible way to find out that your structure does not work.
Letting the tool set the tone. Generators gravitate toward a slick, high-contrast, generic look. If your film needs flat light, awkward framing, or ugliness, you must ask for it explicitly and check every frame against your intent.
FAQ
Do I still need a storyboard artist if I use AI? For most independent work, no — you can produce a functional animatic yourself. For client work, commercials, or anything with heavy action choreography, a human storyboard artist working from your AI-generated reference frames is faster and more accurate than prompting alone.
How many shots should a one-minute film have? Between twelve and twenty-five, depending on pacing. Fast, music-driven pieces sit at the high end; dialogue or mood pieces sit lower. More shots is not more professional — it is more work.
Can AI write a full screenplay? It can produce a full-length draft that reads competently and means very little. Use it for structure, options, and contradiction-checking. Write the dialogue yourself, or at least rewrite every line that carries emotional weight.
What is the biggest continuity problem to watch for? Screen direction and eye-line. They are easy to record in a spreadsheet and nearly impossible to fix after generation, because the fix requires re-generating both shots on either side of the cut.
Do I need expensive video tools to start? No. A language assistant, an image generator, a spreadsheet, and a free editing application will get you through the entire workflow described here. Upgrade only when a specific bottleneck is costing you measurable time.
How do I stop the output looking like everyone else's? Constrain your palette. Pick three colours, two lighting setups, and two lenses for the whole project, and refuse to deviate. Restriction reads as style; variety reads as incoherence.
Should I generate video for every shot? Rarely. Most shorts work best with a small number of generated showcase shots embedded in practically shot material, or in an animatic-driven style where stills carry the weight. Generate motion where motion adds information, not decoration.
How do I present this to a client or producer? Deliver a one-page spine, a beat sheet with tension annotations, a twenty-shot list with feasibility marks, three mood frames, and a thirty-second animatic. That package communicates intent more clearly than a finished-looking generated clip that hides its own weaknesses.
The through-line in all of this is unglamorous: write the spine, annotate the beats, describe behaviour instead of beauty, lock your camera language, document your continuity, and cut an animatic before you spend a minute of production time. AI compresses the distance between an idea and a testable version of it. It does not remove the need for decisions — it just makes bad decisions visible sooner, which is exactly what a well-run pre-production is supposed to do.
