Why Pre-Production Decides the Quality of AI Video
Most disappointing AI video results are not caused by weak models. They are caused by decisions made before a single frame is generated. When a clip feels random, disconnected, or emotionally flat, the reason is usually one of three things: there is no beat structure to give the sequence shape, there is no shot logic explaining why the camera is where it is, or there is no continuity plan holding visual details steady from one shot to the next.
Generative tools are genuinely excellent at one specific job: producing variations quickly. They can draft ten different beat structures before you finish your coffee, describe a location in five visual registers, and assemble a shot list you would never have the patience to type by hand. What they cannot do is decide what your story is about, which emotional turn matters most, or whether a close-up belongs on this line of dialogue or the next one. Direction remains human work.
The practical consequence is simple. Treat AI as a collaborator inside a pipeline you own, rather than as a machine that replaces the pipeline. The rest of this guide walks through that pipeline stage by stage: structure, script, shot list, visual anchors, continuity, animatic, review. Along the way you will find prompts, worked examples, decision criteria, and the mistakes that derail most first attempts.
A working pre-production kit looks like this:
- A one-sentence logline you wrote yourself
- A beat sheet with rough durations attached to every beat
- Scene cards with an explicit change field
- A script that has survived at least three passes
- A shot list built from subject, action, camera, and light
- One paragraph of style description reused in every prompt
- A continuity table filled in for every shot
- An animatic you have watched end to end at real speed
If any of those artifacts is missing, the gap will eventually show up in the render. The model will not fill it in for you, because the missing piece is intent, and intent is not a parameter.
The Pipeline: Five Stages From Idea to Animatic
Stage outputs matter more than stage names
Every stage should produce a document you can hand to the next stage. A stage with no artifact gets skipped the moment a deadline appears, and the sequence always shows it. Name your files consistently, for example scene-shot-version, so that revisions never overwrite each other. Unnamed files are the most common cause of lost work in fast generative workflows.
Time-box each stage on purpose
For a one-to-three minute piece, a realistic schedule looks something like this: structure and logline, thirty minutes; script, sixty to ninety minutes; shot list, forty-five minutes; visual anchors, twenty minutes; boards, sixty to ninety minutes; animatic assembly and review, forty-five minutes. The purpose of time-boxing is not speed for its own sake. It is to stop you from polishing a script for six hours while the shot list stays empty, which is the most common form of procrastination in AI-assisted production.
Where AI helps, and where it hurts
| Stage | AI is strong at | Keep this human |
|---|---|---|
| Structure | Generating three alternative beat shapes | Choosing which shape serves the story |
| Script | Producing volume, alternates, compression passes | Voice, subtext, final line decisions |
| Shot list | Expanding described action into coverage | Deciding what a shot is for |
| Visuals | Style consistency, palette exploration, keyframes | Motivating any stylistic break |
| Review | Flagging repeated framing, timing drift | Judging whether the sequence moves anyone |
One rule that prevents compounding errors
Never let the model own two adjacent stages. If AI writes both your beat sheet and your shot list, its interpretation of the story becomes the shot list, and any structural mistake gets baked in twice. Own the structure, delegate the expansion.
Structure First: Loglines, Beats, and Scene Cards
A logline that actually constrains
A logline is one sentence containing a character, a want, an obstacle, and a stake. Write it yourself, by hand, before opening any tool. It takes ninety seconds and it anchors every downstream decision. If your logline does not name an obstacle, you do not have a story yet, you have a subject.
Beat sheets in three flavors
Ask the model for options rather than a finished outline:
Here is my logline and my tone reference. Give me three beat structures for a ninety-second piece. For each, list the beats with estimated durations in seconds and name the emotional turn inside each beat. Do not write dialogue yet.
The three structures you get back usually fall into recognizable shapes: escalation, where a normal state worsens through repeated attempts until collapse forces a resolution; reveal, where a contradiction reframes what the audience thought they understood; and loop, where the same situation repeats with one variable changing each pass. Escalation suits action and comedy. Reveal suits drama and documentary-style pieces. Loop suits short-form and social formats where rhythm matters more than arc.
Pick one, then cut it hard. A ninety-second piece usually supports four to six beats, not twelve. Attach durations to every beat, because a beat with no duration is a wish rather than a plan. If your total exceeds the target length, remove a beat rather than compressing all of them. Compressed beats produce rushed pacing, and no amount of visual polish repairs rushed pacing.
Scene cards and the change field
A scene card is a compact block: location, time of day, characters present, one sentence of action, and the single change that occurs. Convert each beat into a card, then rewrite the cards yourself. Prompt template:
Convert each beat into a scene card in this format: slugline, characters, one-sentence action, the change, and a suggested visual emphasis. Keep every card under sixty words.
The change field is the one people skip and the one that matters most. If a scene ends in the same emotional and informational state where it began, it is filler. Cut it or merge it into a neighbor. When you have your cards, sort them physically into a strip board, because reordering cards reveals structural problems that a linear document hides: a sagging middle, two consecutive beats of identical intensity, or a reveal placed so late it never pays off.
Drafting Scripts With AI While Keeping Your Voice
The three-pass drafting method
The fastest way to make an AI-drafted script sound generic is to ask for a script. Separate drafting from voice instead.
Pass one, content with no style: ask for the scene in plain minimal prose. Who wants what, what blocks them, how it ends. Ignore how it reads.
Pass two, voice: supply two or three samples of your own writing, or describe the register precisely, such as clipped and professional, warm and digressive, or formal and cold. Ask for a rewrite that preserves the action of pass one.
Pass three, compression: ask for a twenty percent cut with no loss of information. This pass removes the throat-clearing that AI drafts reliably produce, including redundant adverbs, restated emotions, and dialogue that explains what the image already shows.
Dialogue diagnostics worth running
- Subtext check: flag every line where a character states exactly what they mean, then rewrite those lines by hand.
- Voice differentiation: strip the character names and ask the model to identify who is speaking from word choice alone. Any line that could belong to anyone needs work.
- Read-aloud pass: have the script read aloud and mark where your attention drops. Those marks are your edit list.
- Contradiction pass: ask where a character behaves inconsistently with their stated want. Sometimes that inconsistency is the best scene in the piece, but only when it is intentional.
A worked example
Weak draft: I am really frustrated because you never listen to me, and it makes me feel invisible in this partnership.
Compressed rewrite: You had already decided. Why did you ask?
The second line carries the same information with subtext, and it gives a performer something to play rather than something to announce. Notice that the compression pass is not about brevity for its own sake. It is about moving information from dialogue into behavior, which is where it belongs in a visual medium.
Shot Lists That Generative Models Can Follow
The four-part shot line
Every shot line should contain four elements: subject, action, camera, and light. That is the minimum information a generative model needs to produce something usable, and it is also the minimum a human crew needs to shoot it later.
Shot 12: Woman opens the door and freezes, eye level, medium shot with a slight push in, warm practical light spilling from the hallway against a cool interior.
Build vocabulary banks and reuse them
Four categories are worth maintaining as your own reusable lists:
- Framing: extreme wide, wide, medium wide, medium, medium close, close-up, extreme close-up, over-the-shoulder, insert.
- Movement: static, pan, tilt, dolly in, dolly out, tracking, crane, handheld, gimbal, whip pan.
- Lens feel: wide-angle distortion, normal, short telephoto compression, shallow depth of field, deep focus.
- Light: key direction, contrast ratio, practical sources, motivated light, time-of-day palette, atmosphere or haze.
Consistency comes from repeating the same vocabulary across shots. If shot three is a medium close-up with shallow depth of field and warm practical light, then shot nine in the same scene should not suddenly become a wide deep-focus frame unless that change is deliberate and motivated by a beat. Random variation reads as carelessness even when every individual frame is attractive.
From shot line to prompt
Keep the shot line as the source of truth and generate prompts from it. A prompt is a rendering instruction, not a creative document, so it should be replaceable. When a frame comes back wrong, you regenerate from the same shot line rather than improvising a new prompt, because that keeps the parameters documented and the sequence coherent.
Style Anchors, Palettes, and Visual Consistency
Write one style paragraph and paste it everywhere
A style anchor is a single paragraph describing palette, contrast, texture, and mood reference. It goes at the beginning of every prompt in a sequence.
Muted teal and amber palette, soft contrast, subtle film grain, overcast daylight through large windows, restrained camera movement, mid-century interior.
Repeating that block is what holds a sequence together when you generate shots across multiple sessions, and it makes drift obvious. When one frame suddenly looks like it belongs to a different film, the anchor is your reference for what changed.
Use reference images for characters and props
Written descriptions of a person are approximations. A reference image is a constraint. Whenever a tool supports reference input, use it for recurring characters, signature props, and hero locations. Combine the reference with the style anchor, and you get both identity and mood held steady.
Break the style only on purpose
Every palette or lighting shift should be motivated by a story beat. The moment a character enters a new emotional state, a shift in color temperature can do work that dialogue cannot. But if you shift style because you got bored during generation, the audience reads it as an error, not a choice.
Continuity Tracking Across Shots and Scenes
Continuity is where AI sequences break most visibly. A jacket changes color, a window moves, the light flips direction, a character's hair length drifts. None of these problems are model failures. They are tracking failures.
Build a continuity table and fill it in for every shot:
| Field | Why it matters |
|---|---|
| Character look | Clothing, hair, props carried |
| Screen direction | Which way the subject faces and moves |
| Light direction | Where the key light comes from |
| Time of day | Determines palette and shadow length |
| Location details | Furniture, signage, recurring background elements |
| Emotional state | Prevents tonal jumps between adjacent shots |
Two rules prevent most failures. First, keep the number of distinct locations small in short pieces, because every new environment is a new set of variables to control. Second, when a generated frame is close but not correct, regenerate from the same prompt rather than hand-editing, so the underlying parameters stay documented and reproducible.
Screen direction deserves special attention. If a character exits frame right in shot four and enters frame left in shot five, the audience reads a continuous journey. If they enter frame right instead, the geography collapses and the cut feels wrong even to viewers who cannot explain why. Draw a simple overhead map of your scene, mark which side of the frame each character occupies, and keep it consistent until an intentional crossing marks a turn in the story.
Reviewing the Animatic and Iterating Fast
Assemble your still frames into a timeline before generating any motion. Add scratch dialogue, a temporary music bed, or even a metronome click. Then watch the sequence end to end at real speed, without pausing.
This single step catches problems that are invisible in a frame gallery: shots that run three seconds too long, two consecutive shots with identical framing that cut like a jump, a reveal that lands after the audience has already guessed it, and pacing that collapses in the middle third.
Keep a revision log with the date, the shot number, what changed, and why. When someone asks for a change weeks later, the log tells you which earlier decision you are reversing and what else depended on it.
Once the animatic holds up, generate motion only where movement carries meaning. Hybrid sequences, where a handful of animated frames are cut against stills with subtle camera drift, often read as more deliberate than a fully animated reel, and they are far cheaper to revise. Static frames with slow drift are also more forgiving than motion generation, which is worth remembering when a deadline is close.
Choosing Tools Stage by Stage
Match the tool to the stage instead of searching for one platform that does everything.
- Structure and beats: a general-purpose text model with a long context window works well, because you are asking for organization rather than prose.
- Script drafting: whichever model best matches your register, ideally after you have supplied style samples.
- Frames and stills: an image generator with strong style consistency and reference-image support.
- Motion: a video generator that accepts a starting frame, so composition carries forward from your board.
- Animatic assembly: any timeline editor, including free ones. The board is the deliverable, not the render.
Decision criteria that matter more than feature lists: can it accept a reference image, can it hold a style across prompts, does it document the parameters you used, how fast is one revision cycle, and how easily can you export a frame at full resolution. A tool that renders beautifully but gives you no route back to a specific frame will cost more time than it saves.
It also helps to keep a small personal library: your style anchor, your vocabulary banks, a continuity template, a scene card format, and your revision log. That library, not any single model, is what makes AI-assisted pre-production reliable. Tools will keep changing; the pipeline and the library stay yours.
Common Mistakes, Troubleshooting, and FAQ
Mistakes that appear again and again
- Starting with visuals. Boards built before the script produce beautiful nonsense. Write the script first, even a rough one.
- Too many shots. Beginners board forty shots for a sixty-second piece and get a trailer-like blur. Aim for roughly two to four seconds of screen time per shot on average, then cut the remainder.
- Vague camera language. Dynamic shot means nothing. Say what the camera does.
- Unmotivated style changes. Every palette shift should map to a beat.
- Ignoring audio. Dialogue length determines shot length, so time your lines before you time your shots.
- Treating the first output as final. The advantage of generative tools is iteration volume, so budget at least three passes on any shot carrying emotional weight.
- No naming convention. Use scene-shot-version and keep versions side by side.
Troubleshooting quick fixes
If the sequence feels flat, check whether two adjacent shots share the same framing and energy; change one. If a character drifts, add a reference image and re-anchor the style paragraph. If pacing drags, look at the middle third first, since that is where padding accumulates. If a shot never looks right after five attempts, the problem is usually the shot itself, not the prompt, so rewrite the shot line or cut the shot.
FAQ
Do I still need a script if AI can generate one?
Yes. You need a script you have edited, because the script is where dramatic decisions live. AI is efficient at producing options and poor at knowing which option serves your story. An unedited generated script is a starting point, never a shooting document.
How many storyboard frames does a short video need?
Roughly one frame per shot, plus an extra frame for any shot with a significant internal change. A ninety-second piece with twenty-five shots typically needs twenty-five to thirty frames.
Can AI keep a character consistent across many shots?
Increasingly yes, especially with a reference image and a fixed style block repeated in every prompt. It still requires a continuity document and regular manual checks, because consistency is a process rather than a setting.
What is the fastest way to improve output quality?
Slow down at the beat sheet stage. Most visual problems trace back to a structural problem that no model can repair downstream. Fixing the structure takes twenty minutes; fixing it after boarding takes a day.
Should I generate motion for every shot?
No. Generate motion where movement carries meaning. Static frames with subtle camera drift are often stronger and much easier to revise.
How do I present an AI storyboard to a client or collaborator?
Present the animatic rather than individual frames. People respond to pacing and sequence, and a single frame invites arguments about details that do not matter yet.
How long should a first project be?
Sixty to ninety seconds. That length is short enough to finish and long enough to teach you every stage of the workflow, including continuity tracking and the review loop. Scale up only after one complete cycle feels routine.
