Every AI video project that falls apart in the edit usually failed long before the first frame was generated. The model did exactly what it was asked to do; the problem is that nobody had defined what to ask for. Creators working with generative video quickly discover that the bottleneck is not rendering power or resolution — it is preparation.
This guide walks through a practical pre-production workflow for AI-driven video: using a script-analysis assistant to turn a rough idea into structured scenes, converting those scenes into a shot list, generating a storyboard you can actually approve, and carrying all of it into a video generator without losing the thread. It is written for solo creators, small marketing teams, and indie filmmakers who need consistent output without a full crew.
Why Pre-Production Decides Whether an AI Video Succeeds
Two things changed at roughly the same time. First, text-to-video and image-to-video models became good enough that a single prompt can produce a usable shot. Second, the cost of a bad shot dropped to nearly zero, which sounds like good news until you realize it also means you can generate three hundred bad shots in an afternoon and still have nothing to cut together.
Volume is not the problem. Direction is.
A generative model has no memory of your intent. It does not know that the character in scene four is the same person who appeared in scene one, that the phone on the desk matters later, or that the emotional turn happens on the third beat and not the second. Every one of those decisions has to be encoded somewhere — in a document, in a reference image, in a prompt structure — and encoding them is exactly what traditional pre-production does.
The practical consequence is a shift in where your time goes. In a crew-based shoot, most of your hours are spent on set. In AI video, most of your hours are spent upstream: writing, structuring, describing, and reviewing. Teams that accept this and invest in the front end consistently ship better videos with fewer revision cycles than teams that treat prompting as the whole job.
There is also a quality-control dimension. When a shot is wrong, you need to know what specifically was wrong. "It looks off" is not actionable. "The framing is a wide when the beat needs a close-up, and the light is warm when the scene is established as cold morning" is actionable, and you can only make that judgment if you wrote it down first.
What AI Script Analysis Actually Does
An AI script assistant is not a replacement for a writer. It is a structural analyst. Feed it a treatment, a rough draft, a transcript, or even a bullet list of ideas, and it will surface the shape of the story: who wants what, where the tension sits, which scenes are load-bearing and which are decorative.
The useful output is not polished prose. It is a set of decisions you can act on.
From Logline to Beat Sheet
Start with one sentence: who is the character, what do they want, what stops them, what changes. That is your logline, and it should fit in under thirty words. Everything downstream is checked against it.
From there, expand into a beat sheet. A 60-second social video typically needs four to six beats. A three-minute brand film usually needs eight to twelve. Each beat should be describable in a single line that names a change — something is different after this beat than before it. If a beat cannot be described as a change, it is probably a mood, and moods belong inside other beats rather than as separate ones.
An AI assistant is genuinely good at this compression step. Give it your messy draft and ask for the beat structure, then check whether the beats actually escalate. Weak structure is easy to spot once it is laid out in a list.
Scene Cards and the Structural Outline
Once the beats hold up, convert each one into a scene card. A good scene card has six fields: location, time of day, characters present, what happens, what the audience must learn, and how the scene ends. Keep each field to one line.
The scene card is the bridge between writing and production. It is short enough that a generator prompt can be derived from it, and structured enough that you can hand it to an editor or a stakeholder for approval before anything expensive happens.
This is also where continuity lives. If a scene card lists a red jacket, that jacket must appear in every scene where the character wears it. Keeping continuity in the scene card rather than in your head is the single highest-leverage habit in AI video production.
Dialogue, Voiceover, and On-Screen Text
Not every video needs spoken words, but every video needs to know whether it has them. Decide early:
- Voiceover carries exposition and works well over montage or B-roll.
- On-camera dialogue is expensive in AI video because lip sync and performance consistency are still the hardest problems to solve.
- On-screen text is cheap, reliable, and readable on mute — which matters because a large share of viewers watch without sound.
Write the voiceover separately from the visuals. Record or generate it as an audio file early, because having a locked audio track changes how you cut. If you storyboard first and write narration later, you will spend most of your edit forcing visuals to fit words that arrived too late.
Building a Shot List That Survives Generation
A shot list is not a storyboard. The shot list is the technical plan; the storyboard is the visual plan. Do the shot list first, because it forces you to answer questions that are easier to answer in text.
For each scene, list every shot you intend to generate. One scene of ten seconds usually means two to four shots, not one. Coverage gives you options in the edit, and options are what save a video when one generated clip comes back unusable.
Shot Size, Lens, and Camera Movement Vocabulary
Use standard terminology so your prompts stay consistent across a project:
- Shot size: extreme wide, wide, medium wide, medium, medium close-up, close-up, extreme close-up, insert.
- Camera movement: static, slow push in, pull out, pan left or right, tilt, handheld drift, crane, orbit, dolly alongside.
- Lens feel: wide-angle distortion, normal, long lens compression, macro detail.
A shot list entry might read: "MS, static, long lens, subject at desk, warm practical light from left, 4 seconds." That single line contains everything a generator prompt needs and everything an editor needs to know where the clip goes.
Resist the urge to make every shot dynamic. Movement is a punctuation mark, not a default. A video made entirely of orbiting camera moves reads as noise.
Continuity Rules for Generated Frames
Write down the recurring elements and never paraphrase them. A character described as "woman, late twenties, silver-streaked dark hair, olive field jacket" should be copied word for word into every prompt where she appears. Changing one adjective between shots is the fastest way to produce a video where the lead appears to be three different people.
Keep a continuity sheet with four categories: characters, wardrobe, locations, and hero props. Each entry gets a locked description and, where possible, a reference image. Then treat that sheet as a constraint rather than inspiration.
Duration Budgeting
Generators often produce clips in fixed lengths. Plan around that instead of fighting it. If your model reliably returns four- to five-second clips, write your shot list in units of four seconds and build beats from multiples of that. If you need an eight-second shot, either generate two clips and join them on a matched frame, or design the shot so a cut in the middle feels intentional.
Total runtime discipline matters more than most creators expect. Cut your target length by roughly ten percent before you start, because AI-generated footage tends to run slightly long once you add transitions and audio padding.
Storyboarding with Generated Visuals
A storyboard used to mean sketches. Now it can mean low-resolution generated frames, thumbnails pulled from existing footage, or still images composed from reference material. The point is not beauty; the point is approving composition, blocking, and tone before you spend time on motion.
Composition, Light, and Color as Prompt Fields
Treat the frame as a set of independent variables you can adjust one at a time:
- Composition: rule of thirds, centered symmetry, negative space on the right, foreground obstruction, low angle looking up.
- Light: direction (key from left, backlit, top light), quality (soft, hard, diffused), time of day, practical sources.
- Color: palette temperature, dominant and accent hues, saturation level, contrast curve.
When a board frame looks wrong, change one variable and regenerate. Adjusting three at once teaches you nothing and usually produces three new problems.
Sound and Pacing Notes on the Board
Under each board frame, add a one-line audio note: ambient bed, music cue, sound effect, silence. Silence is a legitimate and often underused note — dropping the music for two seconds before a reveal does more work than any camera move.
Pacing notes matter as much. Mark the frames where the edit should breathe and the frames where the cut should be hard and fast. A storyboard that shows only images tells you what the video looks like; a storyboard with timing notes tells you how it feels.
Mood Boards vs. Locked Boards
Distinguish between exploration and lock. A mood board is a collection of references that communicate a feeling; it is deliberately loose. A locked board is a sequence of specific frames with specific compositions that you have approved and will not change without a reason.
Most AI video projects get into trouble because they never leave the mood-board stage. The creator has fifty gorgeous references and no commitment to any of them, so each generated clip drifts in a different direction. Set a lock point and hold it.
A Repeatable End-to-End Workflow
Here is the sequence that works reliably, whether the final video is fifteen seconds or fifteen minutes.
Step 1: Define the deliverable. Aspect ratio, runtime, platform, audio-on or audio-off, and the single action you want the viewer to take. Write it at the top of the project document so it stays visible.
Step 2: Draft the logline and beat sheet. One sentence, then four to twelve beats. Run it through a script-analysis assistant and check that each beat represents a change.
Step 3: Build scene cards. Location, time, characters, action, information, exit. Approve these before moving on.
Step 4: Write the shot list. Two to four shots per scene, with shot size, movement, subject, light, and duration on every line.
Step 5: Create the continuity sheet. Locked descriptions for characters, wardrobe, locations, and props. Attach reference images wherever you have them.
Step 6: Generate board frames. Use still-image generation first. Iterate composition and light at low cost, then lock the board and stop exploring.
Step 7: Produce audio. Voiceover, music bed, and key sound effects. Lock the timing of the voiceover to the second, because it determines your edit points.
Step 8: Generate motion. Work shot by shot, in storyboard order, keeping prompts structurally identical to the locked shot list. Generate two or three variants of every shot that carries emotional weight.
Step 9: Assemble and review on mute. Cut the picture first with no audio. If the sequence does not read silently, the structure is the problem, not the sound design.
Step 10: Finish sound and color. Add music, effects, and a light grade to unify the generated clips. Minor color matching does more for perceived quality than another round of generation.
Common Mistakes and How to Fix Them
Writing prompts instead of plans. If your prompt is a paragraph and your shot list does not exist, you are improvising at scale. Fix: write the shot list line first, then derive the prompt from it.
Changing the character description between shots. This is the most common cause of unusable sequences. Fix: copy-paste the locked description every time. Never retype it.
Generating full scenes in one clip. Models rarely sustain a coherent action across ten seconds. Fix: break the scene into coverage and cut on motion.
Approving frames you do not actually like. A board frame that is "good enough" becomes a finished shot that you resent. Fix: regenerate at the board stage, where it is cheap.
Skipping audio until the end. Audio determines timing. Fix: lock voiceover before you cut picture.
Making every shot move. Constant motion flattens emphasis. Fix: hold the camera still on the shot that matters most.
Ignoring the first three seconds. Most viewers decide whether to keep watching almost immediately. Fix: open on your strongest frame, not your establishing shot.
Choosing Tools Without Locking Yourself In
Pre-production documents should be portable. If your beat sheet, scene cards, and shot list live in plain text or a standard screenplay format such as Fountain, you can move between writing tools, storyboard generators, and video models without redoing the work.
A practical split looks like this:
- Structure and beats: any script-analysis assistant, plus a plain-text or screenplay-format document you own.
- Board frames: any still-image generator, saved as a numbered sequence that matches your shot list.
- Motion: a video generator that accepts both text and image input, so your locked board frames can drive the animation.
- Audio: a separate voice tool and a separate music source, exported as stem files.
- Edit and finish: a standard nonlinear editor, so you are never dependent on a single generator's timeline.
The principle is simple: keep the source of truth in your own files, and treat every generation tool as a renderer, not a home.
Quality Control Checklist Before You Generate
Run this list at the point where you are about to spend real time on motion:
- Does every shot on the list trace back to a beat?
- Is every recurring element described identically in every prompt?
- Does the shot list include enough coverage to survive two failed clips?
- Is the audio locked, with timings written into the sequence?
- Does the assembled board read as a story on mute?
- Is the opening frame the strongest frame you have?
- Have you cut ten percent of the runtime?
- Is the deliverable spec — ratio, length, captions — confirmed at the top of the document?
If any answer is no, the fix is upstream and cheap. Fix it there.
FAQ
Do I need a screenplay to make an AI video? No. But you do need a structure. A beat sheet and scene cards are enough for most short-form projects; a full script only becomes necessary when dialogue, performance, or complex continuity is involved.
How many shots should a one-minute video have? Anywhere from ten to twenty-five, depending on pacing. Fast social edits sit at the high end; narrative pieces sit lower. Coverage matters more than the exact count.
Can I skip the storyboard if I already have a shot list? You can, but you lose your cheapest iteration step. Board frames let you catch composition and lighting problems before you pay for motion generation.
What is the biggest cause of inconsistent characters? Inconsistent wording. Write the character description once, store it in a continuity sheet, and paste it into every prompt without editing.
Should I generate audio first or video first? Lock voiceover first, then picture, then music and effects. Dialogue and narration define the timing that everything else follows.
How do I handle a scene that keeps failing to generate? Change one variable at a time: framing first, then light, then action complexity. If a shot still fails after three attempts, rewrite the shot list entry — the problem is usually that the shot is doing too much.
Final Thoughts
The gap between creators who ship polished AI video consistently and those who keep restarting is rarely technical skill. It is documentation discipline. A beat sheet, scene cards, a shot list, a continuity sheet, and a locked board take an afternoon to produce and save days of regeneration.
Start with the smallest version of this workflow: one logline, four beats, twelve shot lines, and a locked description for your main character. Generate from that. You will notice immediately that your clips match each other, your edit comes together faster, and the question you ask yourself in the edit bay shifts from "what is this supposed to be?" to "which of these two good takes works better?" That shift is what directing actually is.
