Why Pre-Production Decides Whether an AI Video Works
Generative video tools have become extraordinarily capable in a very short time, and that has created a strange new problem: the visuals are rarely the bottleneck anymore. Anyone can generate a beautiful four-second clip. Far fewer people can produce a coherent ninety-second story in which the lighting matches, the character keeps the same jacket, the voice sounds like it belongs to the face on screen, and the cut lands on the beat.
The difference between those two outcomes is almost never the model. It is the plan that sits in front of the model.
Traditional production works the same way. A director does not walk onto a set with a camera and hope for the best. They arrive with a script, a shot list, a storyboard, a casting decision, a lighting plan, and a rough sense of how the edit will feel. Each of those artifacts exists to remove uncertainty before expensive resources get spent. In AI video, the expensive resources are time, iteration count, and attention.
Without pre-production, AI video projects fail in predictable ways. The story drifts because nobody wrote down what the second act was supposed to accomplish. Characters morph between shots because no reference set was ever locked. The edit feels sluggish because the script was written for reading rather than for watching. And the whole project quietly dies inside a folder of sixty near-identical clips.
This guide lays out a director-style workflow you can run with almost any current generation toolset. It is deliberately tool-agnostic, because model names change faster than process does. What follows is a practical system: how to decompose a brief, structure a script, build shot lists and storyboards, lock visual consistency, handle sound, manage iteration, and quality-check the result before publishing.
The Six-Stage Pipeline, From Brief to Final Cut
Before diving into each stage, it helps to see the whole shape of the process. A healthy AI video project moves through six stages, each producing one concrete artifact that the next stage depends on.
- Brief decomposition. Input: a one-line idea or client request. Output: a logline, a target audience, a runtime, and a narrative spine.
- Script architecture. Input: the logline. Output: a beat sheet and a formatted script with timings.
- Shot list and storyboard. Input: the script. Output: a numbered list of shots with camera, subject, action, and duration, plus reference frames.
- Consistency lock. Input: the shot list. Output: character reference sets, style anchors, and a prompt vocabulary sheet.
- Generation and assembly. Input: everything above. Output: raw clips, voice tracks, music, and a rough cut.
- Polish and QA. Input: the rough cut. Output: the final export, plus captions and thumbnail assets.
The critical rule is that no stage begins until the previous artifact exists as a document. Not a thought, not a mental model, a document. This is what keeps a project from collapsing into improvisation once generation starts and the clock is running.
Where most people skip ahead
Almost everyone skips stage one and two and jumps directly to prompting. That feels efficient for a single clip. It becomes catastrophic at thirty seconds and above, because the model has no way of knowing that a detail shown in shot three needs to pay off in shot eleven. Only you know that, and only if you wrote it down.
Stage 1: Brief Decomposition, From One Line to a Narrative Spine
A brief arrives as a sentence: a product launch teaser, a founder story, a comedy sketch about a malfunctioning smart fridge. Your job is to convert that sentence into something a director could shoot.
The four questions
Ask these before writing a single prompt:
- Who is the viewer, and what do they already believe? A viewer who knows the product needs a different opening than a cold audience.
- What is the single change in the viewer's head? Curiosity into interest. Skepticism into trust. Boredom into amusement.
- What is the runtime? Fifteen seconds, thirty seconds, and ninety seconds are three different art forms, not three lengths of the same thing.
- What is the last image? Deciding the final frame early gives the whole story a destination.
Writing the logline
A working logline fits in one sentence and contains a protagonist, a pressure, and a turn. For example: A night-shift barista keeps hearing a customer order the same drink she dreamed about the night before, until she realizes the customer is her own future self. That sentence already implies genre, tone, location, and a shot list shape. A vague brief like something cool with coffee and a twist implies nothing and will produce nothing coherent.
Setting the constraint budget
Constraint is a creative tool. Before generation, decide how many locations, how many characters, and how many distinct lighting setups the video will contain. A useful default for a first AI video: two locations, one or two characters, one lighting mood per location. Every additional variable multiplies the number of things that can drift out of consistency.
Stage 2: Script Architecture, Beats, Timing, and Dialogue Economy
A script for AI video is not a screenplay. It is a timing document with emotional checkpoints.
Build a beat sheet first
List five to eight beats, each one sentence, each one a change in the situation. For a thirty-second piece: hook, context, complication, escalation, turn, resolution, button. If a beat does not change anything, cut it. Flabby beat sheets produce flabby videos, and flabbiness is far more visible on screen than on the page.
Do the timing math honestly
Speaking pace for natural narration runs roughly 140 to 160 words per minute, so a thirty-second voiceover holds about 70 to 80 words. That is shorter than most people expect. Write the voiceover, read it aloud with a stopwatch, and cut until it fits. Then cut 10 percent more, because generated voice tracks often need breathing room that text does not show.
Write shot-level stage directions
Instead of prose description, annotate the script with intent. Each line should carry a light note about how it looks: wide, cold morning light, subject small in frame or tight, warm practical lamp, handheld feel. These notes become the seed of your prompts and prevent the visual language from lurching between scenes.
Dialogue that survives text-to-speech
Generated voices stumble on stacked clauses, dense proper nouns, and unmarked sarcasm. Keep lines short. Break long sentences into two. Write numbers the way you want them spoken. If a joke depends entirely on timing, plan to place it in the edit rather than trusting the model to land it.
Stage 3: Shot Lists and Storyboards That Actually Guide Generation
A storyboard for AI video is not an art project. It is a specification. Sloppy storyboards produce sloppy videos because the generation step has nothing precise to aim at.
The anatomy of a usable shot record
Every shot in your list should contain these fields, even if you fill them in shorthand:
- Shot number — sequential, never reused after a revision.
- Duration — in seconds, targeting three to six seconds for most generated clips.
- Subject and action — who does what, in one active verb.
- Camera — framing (wide, medium, close), angle, and movement (static, push in, pan).
- Setting and light — location plus one dominant lighting quality.
- Audio — dialogue line, ambient bed, or sound effect.
- Purpose — the reason this shot exists in the story.
That last field is the one people forget and the one that saves the most time later. If a shot's purpose is establish that the office is empty, you know immediately during review whether it succeeded, regardless of how pretty it looks.
Turning shot records into prompts
A good prompt is a compressed version of the shot record, ordered from largest to smallest decisions: subject, action, setting, lighting, lens and framing, movement, then style. Keep it under roughly sixty words. Longer prompts do not produce more control; they produce competing instructions, and the model quietly picks one.
When to skip the storyboard
You can skip drawn frames when the video is a montage of unrelated visuals, when it is purely abstract, or when it is a single continuous shot. You should never skip the written shot list. Drawing is optional; specification is not.
Storyboard frames as reference images
If you do generate reference frames, treat them as continuity documents rather than final artwork. Export them at consistent aspect ratio, label them with the shot number, and keep them in a single folder that you consult while prompting. This one habit prevents the most common failure in generated video: a character whose face changes identity every time the camera moves.
Stage 4: Consistency, Characters, Props, and Visual Style
Consistency is the hardest problem in AI video and the one most improved by process rather than by tools.
Build a character reference set
Collect or generate four to eight images of each character: front, three-quarter, profile, and a full-body shot. Ideally vary the lighting across them so the model learns the face, not the lamp. Name the set clearly — mara_reference_v2 — and never mix versions inside one project.
Write a style anchor paragraph
Draft one short paragraph describing the look of the entire video: color palette, contrast, film grain or cleanliness, lens character, and mood. Paste a trimmed version of it into every prompt. This is the single cheapest consistency technique available, and it costs nothing but discipline.
Keep a prompt vocabulary sheet
Words drift. If you describe a jacket as mustard in one prompt and ochre in another, you will get two jackets. Maintain a running glossary of the exact terms you use for each recurring element — wardrobe, props, locations, hairstyles — and copy them verbatim.
Continuity pitfalls to watch for
- Lighting direction flips between shots in the same scene, breaking the illusion of a shared space.
- Wardrobe changes mid-scene because a prompt omitted the jacket detail.
- Prop disappearance, especially small items like glasses or phones that models drop when the framing tightens.
- Scale drift, where a character appears taller or shorter relative to the set.
- Style whiplash, when one shot looks photographic and the next looks illustrated.
Review each scene as a group, not shot by shot. Problems that are invisible in isolation become obvious in sequence.
Stage 5: Sound Design, Voice, and Pacing
Viewers forgive imperfect visuals. They rarely forgive bad audio.
Casting the voice
Choose the voice before generating the visuals if dialogue matters. The timbre of a voice constrains the face, the age, and even the wardrobe. Generate a few test lines, listen at normal speed rather than scrutinizing them, and pick the one that carries emotion without pushing.
Layering the audio bed
A finished track usually has three layers: dialogue, an ambient or room tone layer, and music. Ambient tone is the most frequently skipped and the most transformative — a faint hum, distant traffic, or a room's natural reverb instantly makes a generated scene feel like a real location.
Cutting to rhythm
Place your music first, mark the beats, then trim shots to those marks. Cuts that land on the beat read as intentional; cuts that land a quarter second late read as amateur. For short-form vertical video, keep the opening image under two seconds before the first cut or camera move.
Lip sync and mouth coverage
If characters speak on camera, favor medium and wide shots where mouth detail is less scrutinized, and reserve tight close-ups for moments without dialogue. This is a classic film trick and it works even better with generated footage.
Stage 6: Iteration, Version Control, and Render Discipline
Generation invites infinite tweaking. Structure prevents it from consuming the project.
The three-pass rule
Allow three generations per shot. Pass one is exploration, pass two incorporates the best elements of pass one, pass three is the final attempt. If none of the three works, the problem is usually upstream — the shot description is ambiguous or the shot does not belong in the edit. Fix the plan, not the prompt.
Naming conventions that save you later
Use a strict pattern: project_scene_shot_version. For example, coffee_s02_sh07_v03. Add a one-word status tag when a clip is approved: coffee_s02_sh07_v03_ok. This makes assembly mechanical instead of archaeological.
Batch by similarity
Group generations by location, lighting, and character so you can reuse the same reference images and prompt scaffolding across a run. Batching reduces the mental overhead of switching contexts and produces more consistent results within a scene.
Keep a decision log
A simple running note of what you tried and why you rejected it pays off enormously on longer projects. Two weeks later, when a client asks why a shot looks the way it does, the answer is already written down.
A Ten-Minute Quality-Control Checklist
Run this before exporting. It catches the majority of issues that make AI video look unfinished.
- Story check: Does each shot have a purpose? Could any shot be removed without loss?
- Continuity check: Wardrobe, props, lighting direction, and location consistent within each scene?
- Motion check: Are there fake-looking micro-movements, warping edges, or morphing faces?
- Text check: If signs or screens appear, is the lettering legible or deliberately abstract?
- Audio check: Dialogue audible, music not masking speech, no abrupt level jumps between cuts?
- Pacing check: Does the first two seconds earn attention? Does the final frame land?
- Format check: Correct aspect ratio, safe margins for captions, and a thumbnail frame that reads at small size.
- Accessibility check: Burned-in or uploaded captions, and a description that makes sense without sound.
Common Mistakes That Sink AI Video Projects
Chasing model novelty instead of finishing. A new tool appears every week. Switching mid-project breaks consistency and rarely improves the outcome. Finish with what you started, then experiment on the next project.
Writing a script that only works on paper. If the script reads well but has no visual rhythm, the edit will feel static. Read it aloud while imagining cuts.
Overloading prompts. Ten descriptive clauses fight each other. Six clear decisions beat ten vague ones.
Ignoring scale and geography. Audiences track spatial logic unconsciously. If a character walks left in one shot and enters from the right in the next, something feels off even if nobody can name it.
Publishing without sound design. Music alone is not sound design. Add room tone and at least one diegetic effect per location.
Treating the first generation as the final. Generating is cheap; assembling is expensive. Budget your time accordingly.
FAQ
How long should an AI-generated shot be?
Three to six seconds is the sweet spot for most current models. Shorter shots give you more editorial control and hide imperfection. Longer shots look impressive in a demo but are harder to cut and more likely to drift.
Do I need a full storyboard for a thirty-second video?
No drawings, but you need a complete written shot list. A storyboard becomes genuinely valuable once a project involves recurring characters, complex blocking, or a client approval step.
What is the fastest way to improve character consistency?
Lock a reference set of four to eight images per character, write one style paragraph that you paste into every prompt, and maintain a vocabulary sheet for wardrobe and props. Process beats parameter tweaking.
Should I write the script or the voiceover first?
Write the script, then immediately record a rough read to check timing. If the read runs long at first attempt, the script is too dense. Cut before you generate anything.
How do I handle dialogue scenes?
Keep lines short, favor medium shots while speaking, and place close-ups on reactions rather than speech. If sync still looks off, convert the moment into a voiceover with a visual of the listener.
What is the right order of operations when a client wants revisions?
Go back to the artifact, not the render. If the story changed, update the beat sheet. If the look changed, update the style anchor. Regenerating clips from an outdated plan simply produces new versions of the same problem.
How many shots should I plan per minute?
Roughly fifteen to twenty-five for fast-paced short-form, and eight to twelve for calmer narrative work. Plan the count before generating so you can budget time honestly.
Is it worth doing this much planning for a single clip?
No. The pipeline pays off from about thirty seconds upward, or any time a client, a brand, or a recurring character is involved. For a one-off experiment, prompt and move on.
