Why AI Assistance Changed the Shape of Video Pre-Production
For most of the history of film and video, the expensive part of production was capture. Cameras, crews, locations, lighting, permits, reshoots — every decision had to be defensible because changing it later cost real money. Generative video flipped that. Capture is now the cheapest it has ever been, and the bottleneck has moved upstream to the one thing that cannot be generated on demand: a clear decision about what the video is supposed to be.
That shift has a practical consequence. Vague ideas are no longer free. A half-formed concept used to be harmless because a director, a cinematographer, and an editor would resolve it on set. In an AI-assisted pipeline, an unresolved idea turns into forty generated clips that do not cut together. The hours you invest writing a script and designing shots are now the highest-leverage hours in the entire project.
The second reason scriptwriting and shot design moved to the center is that generative visual tools reward specificity. "A woman walks into a cafe" is not a shot; it is a premise with thousands of valid interpretations, and the model will pick one at random. "A woman in a rust-colored coat pushes through a glass door into a narrow cafe, morning light raking across the tile floor, handheld medium shot at chest height" describes one shot with one intention. When you write that sentence you are not writing a prompt — you are writing a shot list entry. The two activities have effectively merged.
Third, iteration economics favor text. Rewriting a scene costs minutes. Generating a scene also costs minutes, but multiplied across every shot, voice pass, and revision. Every problem you solve in the script is a problem you do not have to solve across a hundred render attempts. The workflow below is built on that principle: decide in text, generate once.
The Six-Stage Workflow, From Rough Idea to Locked Shot List
A repeatable pipeline matters more than any single tool. Tools change monthly; the sequence does not.
Stage 1: Lock the brief before you open a generator
Write a one-page constraint sheet and keep it open in a second window for the rest of the project:
- Logline in one sentence: subject plus tension
- Target runtime, in exact seconds
- Aspect ratio and platform, plus safe areas for captions
- Audience and the one feeling they should leave with
- Tone references: three existing films or ads, not adjectives
- Must-include elements: product, logo, claim, brand color
- Must-avoid elements: competitor lookalikes, cliches, prohibited claims
This page is what you hand an AI writing assistant as context. Without it, the assistant will produce something competent and generic, and you will burn an hour discovering that "modern and cinematic" is not a brief.
Stage 2: Build a beat sheet
A beat is a change in the story, not a scene. Under ninety seconds, five to nine beats is usually right. Around three minutes, twelve to twenty. Ask an AI assistant for three structurally different beat sheets from the same brief — one linear, one starting in the middle, one built around a single continuous moment — then merge the strongest beats into one outline.
Assign a rough duration to every beat as you go. This is the single most useful habit in short-form video: a beat sheet with numbers attached is a budget you can actually defend.
Stage 3: Break beats into scenes
Each scene gets an ID, a location, the characters present, time of day, dramatic function, and estimated duration. A compact table works well. When the durations stop adding up to the target runtime, you have a decision to make — cut a scene or shorten three — and it is much cheaper to make it here than in the edit.
Stage 4: Translate scenes into a shot list
This is the document you will actually work from. Use a consistent naming convention from the start, such as S02_SH03_TAKE01, and never rename mid-project. Every shot entry carries:
- Shot size and camera angle
- Camera movement, or an explicit note that the shot is static
- Subject action in one active verb
- Lighting note
- Estimated duration in seconds
- Audio note: dialogue, ambience, or music cue
If a shot entry takes more than two lines to describe, it is usually two shots.
Stage 5: Generate keyframes, then motion
Generate a still keyframe for every shot first. Stills are fast and cheap relative to video clips, they expose composition problems immediately, and they give you a locked visual target. Only when the whole board of stills feels coherent should you animate them. Teams that skip this step spend their budget discovering composition errors in motion.
Stage 6: Assemble with temporary audio
Cut a rough version with placeholder narration and a scratch music bed. Pacing problems are almost always visible at this stage, before the visuals are polished, and fixing them here costs nothing but a re-edit.
Working With an AI Writing Assistant Without Losing Your Voice
Use it as a structural editor, not a ghostwriter
The most valuable use of an AI script assistant is not first-draft generation — that output tends to be smooth, generic, and oddly paced. It is critique. Feed it your draft and ask targeted questions: Where does tension drop? Which scene repeats information already delivered? What is the strongest possible first line? Give me three alternative endings in under forty words each. Then write the revision yourself. You keep ownership of the voice, and you gain an outside read in seconds.
Track characters and continuity deliberately
Keep a character sheet per project: appearance, wardrobe, vocal register, mannerisms, relationships, and one secret that informs behavior. Update it after every scene where something changes. For series work, add a continuity log — what changed, in which episode, and what it affects downstream. This document is what keeps an AI image or video generator from redesigning your lead character in the third episode.
Run dialogue and pacing passes
Read every line aloud. Anything you stumble over, a voice model will stumble over too. Cut fifteen to twenty percent of the words in a typical draft. Keep voice-over sentences under about fourteen words, spell out ambiguous numbers and abbreviations, and avoid homographs that text-to-speech models mispronounce. Small formatting discipline here saves an entire round of audio regeneration.
Designing Shots: Camera Language That Survives Generation
Learn the shot-size vocabulary and use it consistently
Establishing wide shots orient the viewer. Full shots show body language. Medium shots carry dialogue. Medium close-ups carry emotion. Close-ups carry intensity. Inserts carry information — a hand, a screen, a detail. One idea per shot. A useful discipline is to alternate wide and tight deliberately rather than defaulting to medium for everything, which is what happens when a shot list has no size column.
Move the camera only with motivation
Static shots cut together more reliably than moving shots, in AI pipelines and in traditional editing alike. Use movement when it has a job: a slow push to reveal something, a follow to create momentum, a pan to connect two subjects. One moving shot per beat is a reasonable ceiling. If a shot needs three simultaneous movements to feel dynamic, the problem is the shot, not the movement.
Build a look bible
Create a single document containing three to five exact phrases that define your visual identity, and paste them verbatim into every visual prompt:
- Palette: three named colors plus one accent
- Lighting: direction, quality, and time-of-day rules
- Lens character: focal-length feel, depth of field, distortion
- Texture: grain, halation, contrast curve
- Format: aspect ratio, frame-rate feel
Because the phrases are identical across shots, the output feels like one film instead of a mood board. The moment you start paraphrasing your own look bible, drift begins.
Check continuity between scenes
Screen direction, eyelines, wardrobe, prop placement, time of day, weather, and background extras. AI generation will happily flip a character's orientation or change the light between two shots of the same conversation. A short continuity checklist applied to the keyframe board catches most of it before animation.
Building a Reproducible Prompt System
Treat prompts as production assets, not one-off messages. A reliable prompt has a fixed anatomy:
[shot size and angle] + [subject and wardrobe] + [single action] + [environment] + [lighting] + [lens and format] + [style reference] + [exclusions]
For example: "Handheld medium shot at chest height, woman in a rust-colored wool coat, pushing open a glass door, narrow tiled cafe interior with a chrome counter, morning light raking across the floor from the left, 40mm lens, shallow depth of field, muted cinematic color grade, no text overlays, no extra people."
Then keep a shot file for each entry containing the prompt, the seed or reference images used, the style tokens, and the take numbers you generated. When a client asks for a change in week three, you can reproduce the original look exactly instead of guessing. Version your prompts the way you version code.
Choosing Tools for Each Stage
Script and planning
Look for outline templates, beat-sheet views, and the ability to keep character sheets attached to a project. Plain documents work fine; the structure matters more than the software.
Storyboards and keyframes
Image generation with reference-image support is the key feature. Character consistency across many shots depends far more on how well the tool accepts references than on raw image quality.
Video generation
Prioritize control over spectacle: camera-motion controls, start and end frame support, consistent aspect ratios, and clip lengths that match your shot durations. Test three tools on the same ten shots; the one that holds your character's face still across them is usually the right answer.
Audio
Text-to-speech with pronunciation control, a small sound-effects library, and a music source with clear commercial licensing. Audio design is what makes AI video read as professional rather than experimental.
Decision criteria that actually matter
Control and consistency first. Then clip length, resolution, aspect-ratio flexibility, licensing terms for commercial use, generation speed, cost predictability at your real volume, and how cleanly exports move into your editor. Demo reels are not criteria; your own test shots are.
A Worked Example: Sixty-Second Product Film
| Shot | Description | Duration |
|---|---|---|
| 1 | Establishing wide, city street at dawn, empty | 4s |
| 2 | Medium, subject walking with product in hand | 5s |
| 3 | Insert, hands opening the product | 3s |
| 4 | Close-up, face reacting | 4s |
| 5 | Medium close, product in use at a desk | 7s |
| 6 | Insert, detail of a feature | 3s |
| 7 | Wide, subject leaving the room, product on desk | 5s |
| 8 | Product beauty shot, static, clean background | 6s |
| 9 | Voice-over and end card space | 8s |
That is forty-five seconds of picture plus an eight-second end card, with opening and outro music covering the rest. Note how the durations are uneven — real films breathe — and how only two shots involve camera movement.
Common Mistakes That Break an AI Video Workflow
- Writing for humans instead of generators. Vague verbs like "experiences" or "reflects on" give a model nothing to render. Use physical, observable actions.
- Changing the look mid-project. A new style phrase on shot thirty forces you to regenerate shots one through twenty-nine.
- Overloading every shot with movement. It reads as chaos and it breaks continuity between cuts.
- Ignoring audio until the end. Weak sound is what makes generated footage feel cheap.
- Skipping the keyframe stage. Composition problems are far cheaper to fix in stills.
- No naming convention. Untracked files turn a two-hour revision into two days.
- Generating before the script is locked. Every script change invalidates generated footage.
- Forgetting captions and safe areas. Vertical crops and subtitle zones should be planned in the shot list, not patched afterward.
A Quality-Control Checklist Before Final Render
- Every shot in the list has an approved keyframe
- Character appearance is consistent across all scenes
- Screen direction and eyelines are coherent
- One idea per shot; no shot exceeds its planned duration
- Audio levels normalized, no clipping, ambience continuous under cuts
- Captions legible against every background
- Aspect-ratio safe areas respected for every target platform
- Color and contrast consistent across shots
- Licensing verified for every voice, music track, and reference asset
- Export settings match the destination platform
Scaling: Series, Campaigns, and Localization
Once the workflow is stable, it becomes a template. Keep a project skeleton with the brief, beat sheet, scene table, shot list, look bible, and prompt file in place, and duplicate it for each new episode. Reusable asset libraries — character sheets, location keyframes, transition shots, music beds — shrink production time dramatically on the second and third videos in a series.
Localization is where this approach pays off most. Because narration and on-screen text are separate layers, one master edit can produce multiple language versions with re-recorded voice-over and swapped text, provided you reserved space in the shot list for longer translated lines. Keep two extra seconds per text-heavy shot as a general rule.
Frequently Asked Questions
Do I still need a human writer if I have an AI assistant?
Yes, and more than before. The assistant accelerates structure and critique; taste, judgment, and voice remain human responsibilities. The output is only as distinctive as the person directing it.
How long should a shot be in an AI-generated video?
Match the shot duration to what the moment requires, then check it against your generator's maximum clip length. Four to seven seconds per shot is a practical default for short-form work, with inserts as short as two seconds.
What is the single biggest cause of inconsistent visuals?
Paraphrasing your own style description. Copy the exact phrases from your look bible into every prompt and keep reference images fixed.
Can I use AI-generated footage commercially?
That depends on the terms of the specific tools you use and the input assets you supply. Check licensing for every generator, voice model, and music source in your pipeline before delivery.
How do I handle dialogue-heavy scenes?
Separate performance from picture. Generate the visual with clear mouth movement, or shoot it from angles where the mouth is not the focus, then layer clean recorded or synthesized audio on top.
What should I do when a shot keeps failing?
Simplify. Remove movement, reduce the number of subjects, shorten the action to one verb, and re-check that the shot duration is within the model's clip limit. Most failures are complexity problems, not model problems.
