AI Scriptwriting and Storyboarding: A Video Pre-Production Guide

Every disappointing AI video project tends to fail in the same place. Watched shot by shot, the generated clips look impressive. Watched end to end, they fall apart: faces drift between cuts, pacing sags through the middle, and the original idea dissolves somewhere between prompt three and prompt thirty.

The generators are rarely the weakest link in that chain. Planning is. When a shot takes ninety seconds to render instead of a full day on set, teams quietly stop writing things down. They iterate by feel, regenerate constantly, and finish with a folder of beautiful orphan clips that refuse to behave like a film.

This guide lays out a neutral, tool-agnostic pre-production workflow for AI-assisted video. It covers how to build a story spine before touching a generator, turn a finished script into a shot list a machine and an editor can both read, produce storyboard frames that act as control signals rather than decoration, hold characters and locations steady across dozens of generations, plan camera movement and sound deliberately, choose tools by shot type instead of by hype, and run a review loop that does not collapse under version sprawl.

Why planning decides whether an AI video works

Classical pre-production existed for a reason that generative tools have not removed: the cost of a bad decision rises sharply the later you make it. In traditional production the ladder looks like this. A story flaw caught while outlining costs a rewrite of a paragraph. The same flaw caught in a table read costs an afternoon of discussion. Caught on the shoot day, it costs a location, a crew, and overtime. Caught in the edit, it costs a reshoot you cannot afford.

AI production flattens that ladder but does not remove it. A story flaw discovered before prompting costs a document edit. The same flaw discovered after two hundred generations costs a week of rework, a blown deadline, and a client who has lost confidence in the process. The middle rungs are cheaper than they are on a real set, which is exactly what makes teams skip them — and skipping the cheap rungs is how you end up paying for the expensive one.

There is a second, subtler reason planning matters more with generative tools rather than less. Video models are not neutral executors of intent. They interpolate, reinterpret, and invent. Where your brief is vague, the model fills the gap with its own default taste. The result is a shot that looks polished and is polished about the wrong thing — a confident close-up of a character whose emotional beat the script never asked for, or a wide shot that solves a geography problem you did not know you had.

So the useful mental model is this: pre-production is where you shrink the model's freedom to exactly the dimensions that matter, and leave it free everywhere else. Writing, structure, and visual references are the three levers — the same three a human crew needs before the first camera rolls.

One practical consequence: treat the planning documents as the real deliverable of the first phase. If you can hand a shot list, a character sheet, and a style block to a collaborator and they can produce a sequence that matches yours, the planning worked. If they cannot, the plan is still living in your head, where it will not survive a deadline.

What AI planning help does well — and where it misleads

Being honest about the division of labour saves a lot of wasted regeneration later.

Where AI planning assistance earns its place

• Structural drafting. Models are strong at proposing beat structures, three-act shapes, hook variations, and cold-open options. Getting twenty ways to open a ninety-second video in under a minute is genuinely useful for ideation, even if you keep only one.
• Condensation. Turning a long article, transcript, webinar, or product document into a script outline is fast and competent work, provided you state the target runtime, audience, and platform.
• Shot expansion. Given a script, a model can propose breakdowns, suggest coverage, and flag beats that will be hard to depict without human performance — a useful early warning system.
• Visual reference generation. Concept frames, mood boards, palette studies, and character sheets are quick to produce and quick to iterate.
• Continuity bookkeeping. A running list of wardrobe, props, locations, time of day, and screen direction is tedious for humans and trivial for a well-prompted assistant.

Where it consistently misleads

• Dramatic judgement. Models default to explaining instead of showing, and they resolve tension far too early. A script can read cleanly and still be dramatically inert.
• Runtime intuition. Ask for a sixty-second script and you may receive 140 words of dialogue that cannot possibly be delivered alongside visuals. Comfortable narration runs at roughly 2.5 words per second, so a minute of voice-over is about 150 words before you subtract pauses, music, and any breathing room.
• Physical plausibility. Blocking suggestions frequently ignore geography. Characters walk through walls, or a two-person conversation is staged across an impossible distance, or a hand reaches for an object that was never established in frame.
• Taste convergence. Left unsupervised, models drift toward a glossy, generic aesthetic. Your differentiation comes from the constraints you impose on purpose.

The working rule: use AI for volume and structure, and reserve human attention for dramatic logic, physical continuity, and style decisions. Those are the three places where a bad call is expensive and a machine has no stake in the outcome.

Choosing a tool stack: decision criteria that actually matter

Most AI video projects do not need one perfect tool. They need a small stack where each layer does one job and hands off cleanly. A typical stack has five layers: a writing and structuring layer, an image layer for concept frames and reference portraits, a video generation layer, an audio layer for voice and music, and an editing layer where everything is graded and assembled.

When comparing options inside a layer, judge them against the work you actually do, not against feature lists. The criteria that change outcomes most often:

• Identity retention. Does the model keep the same face, hair, and build across different angles and lighting conditions? This single criterion decides whether your film has a protagonist or a rotating cast of lookalikes.
• Reference image support. Can you feed an image and have it respected? Text-only consistency is fragile; image conditioning is the strongest control signal available.
• Camera control. Can you specify a move, a height, and a direction, or are you writing poetry and hoping?
• Clip length and pacing. Short clips are easier to control; longer clips preserve motion but drift on identity. Know which trade-off your shot type needs.
• Motion character. Some tools produce smooth, almost weightless movement; others render weight and friction better. Neither is universally correct, but mixing them without a grade in between creates visible seams.
• Text and graphic rendering. On-screen text, logos, and infographics are still far better handled in conventional editing or motion design tools than generated. Plan to composite them.
• Handoff quality. Frame rates, aspect ratios, colour space, and export options determine how much cleanup happens in the edit.
• Predictable cost structure. Not the price itself, but how the work is metered. A layer that punishes experimentation will quietly kill your willingness to iterate.

Run a style-matching test before committing. Generate the same shot — same framing, same subject, same lighting note — in three candidate tools, then compare them side by side for grain, colour response, skin tones, and the feel of motion. If one tool's default palette is noticeably warmer or more saturated, you will be fighting it in every shot unless you either standardise the grade in the edit or leave that tool out of the project. The test costs an hour and prevents weeks of correction work.

Step 1: Build the story spine and beat sheet before prompting

A story spine is the shortest possible description of what changes over the course of the video. Not what happens — what changes. A product explainer in which a frazzled freelancer becomes a calm freelancer has a spine. A montage of features does not, which is why feature montages feel like slideshows no matter how good the footage is.

Write the spine in three sentences: the starting state, the disruption, and the resulting state. Then add one sentence describing what the viewer should feel by the final frame. Keep that document open for the whole project. Every later decision — shot choice, music, pacing, colour — gets tested against it.

Once the spine exists, expand it into beats. For a video under ninety seconds, five to seven beats is usually right. Longer pieces can carry twelve to fifteen before the structure starts to feel padded. Each beat gets a one-line description and an emotional temperature. This beat sheet, not the finished script, is the first artefact worth feeding to an assistant for expansion.

A prompt pattern that works well here is constraint-first: state the runtime, the audience, the platform aspect ratio, the tone, and the number of beats, then ask for three alternative expansions of each beat rather than one finished script. You are shopping for options, not accepting a draft. Keep the beat descriptions short enough that a shot list can be derived from them line by line, because that derivation is the next step and it is much easier from ten clean lines than from two pages of prose.

Step 2: Turn the script into a machine-readable shot list

The bridge between a finished script and generated footage is the shot list, and this is the artefact most AI workflows get wrong. People write prose shot descriptions that read pleasantly and cannot be automated, batched, or audited.

A production-ready shot list is tabular and disciplined. Each row describes exactly one generation. Each column answers exactly one question that a model, an editor, or a continuity reviewer will ask:

• Shot ID — a stable identifier such as S01, S02, S02b. Never renumber. Append suffixes instead, so notes written last week still resolve.
• Duration — target seconds, plus a note on whether it is flexible.
• Shot size — wide, medium, close, extreme close, insert.
• Subject and action — one primary action per row. Two actions means two shots.
• Setting — where, what time of day, what weather, what is in the background.
• Camera — angle, height, movement, lens feel.
• Lighting and palette — the mood described in visual terms, not adjectives like epic or cinematic.
• Continuity notes — wardrobe, props, which hand holds an object, which direction a character faces.
• Audio — dialogue line, voice-over line, ambience, or music cue.
• Reference asset — the storyboard frame or character sheet this shot must match.

The discipline pays off the moment something goes wrong. When a shot fails, you know whether the failure was the prompt, the reference image, or the concept itself, because each concern lives in its own field. The table also makes batching realistic: rows that share a camera setup, lighting note, and style block can be generated with identical parameters, which is where consistency is cheapest to achieve.

A quick sanity check before generating anything: read only the action column from top to bottom. If it reads like a sequence of causally linked events, your shot list is a film. If it reads like a list of unrelated images, you have a mood board and you should go back to the beat sheet.

Step 3: Storyboard frames and reference assets that hold continuity

Storyboards for AI production serve a different purpose than traditional ones. A human storyboard artist communicates an idea to a crew that will interpret it. An AI storyboard communicates a target to a model and a reference set to a continuity reviewer. That means the frames should be more literal, not more expressive.

Start with rough blocking on paper or in any sketching app. You are solving geography: where people stand, which way they face, what sits in the foreground, where the camera is. Twenty seconds per frame is plenty. The goal is to discover that a two-person dialogue scene does not work in one wide shot before you generate six versions of it.

Only then generate polished concept frames. Two techniques work consistently well:

1. Frame one and multiply. Generate a single strong, fully specified image of your protagonist in the correct wardrobe, location, and light. Use it as an image reference for every subsequent shot, changing only pose, shot size, and angle.
2. Block the location as a grid. Generate the same place from four angles — wide, medium, reverse, and detail. This mini location bible stops the background from redesigning itself in every cut, which is one of the most common and most distracting failures in generated sequences.

Name every frame with its shot ID. The single most frequent cause of continuity chaos is a folder of images named render_final_v3.png that nobody can map back to a row in the table. If a filename cannot be traced to a shot in three seconds, the naming convention has already failed.

Step 4: Prompt architecture for characters, locations, and style

Consistency does not come from better adjectives. It comes from architecture: a reusable vocabulary that appears, unchanged, in every prompt that shares a subject.

Build three reference documents and treat them as locked assets.

A character sheet. For each recurring person, define age range, build, hair, wardrobe with specific colours and materials, distinguishing features, and default expression. Write it as a compact block of keywords, not flowing sentences. Then generate a reference portrait and keep it as an image input for every shot featuring that character. When a model drifts, the fix is almost always a stronger reference image rather than a longer description.

A location bible. The same treatment for places: architecture, materials, signage, vegetation, the direction the light comes from, time of day, and a fixed palette. Generate two or three angles per location and reuse them across the sequence.

A style block. This is the global suffix appended to every prompt: rendering style or film stock, lens character, colour grade, contrast, grain, and aspect ratio. Keeping it identical across the project is what makes separately generated shots feel like one film rather than a sampler reel.

Here is the practical test for a working prompt architecture. If you deleted the source script, could a stranger regenerate your video from the shot list, character sheet, location bible, and style block alone? If not, your documentation is still carrying tacit knowledge that will not survive a busy week.

Step 5: Plan camera movement, pacing, and sound

Camera movement is where AI video is most seductive and most expensive. A slow push looks cinematic in a demo, but twenty pushes in a row produce nausea and a flat rhythm. The audience stops noticing movement that never stops.

Plan motion at three levels. At the sequence level, decide where movement is allowed at all — static shots give moving shots their impact. At the scene level, choose a dominant grammar: locked-off interview framing, handheld documentary energy, or smooth dolly moves. At the shot level, specify one movement per shot, with a direction and a speed.

Guidance that holds across most current models:

• Simple moves render reliably. Pushes, pulls, pans, tilts, and slow orbits are safe. Complex combinations — a push that becomes a crane that becomes a whip pan — usually produce artefacts.
• Movement is easier to add in post than to control in generation. Cutting a static shot slightly shorter with a digital push is often cheaper and cleaner than generating motion.
• Motion should follow motivation. Move the camera because the subject moves, or because the audience needs to discover something in the frame.
• Keep a motion ledger in the shot list. If three consecutive shots all orbit, you will spot the repetition on paper and fix it for free.

Sound deserves the same planning. Generating voice separately often gives better control and easier pickups than generating it inside a video tool, at the cost of alignment work in the edit — usually a minor trade for the flexibility gained. Even a rough scratch track changes your pacing decisions, because silence hides dead stretches that narration exposes immediately.

Step 6: Review passes, versioning, and asset hygiene

Quality collapses at scale without naming discipline. Adopt a convention on day one: project_shotID_variant_notes. Keep originals immutable and derive new versions rather than overwriting, so a rejected direction can be revisited without regenerating from scratch.

Run reviews in passes instead of reacting to each generation as it arrives:

1. Continuity pass. Watch the sequence with the sound off. Does the world stay consistent? Do characters keep their wardrobe, props, and screen direction?
2. Story pass. Watch with audio only, or read the script against the timeline. Does the spine still land? Do the beats escalate?
3. Rhythm pass. Watch at speed with half your attention. Are there dead stretches? A sequence that only works under full concentration is usually too slow.
4. Craft pass. Only now assess individual shots for artefacts, softness, and finish.

Keep a decision log. When you reject a shot, write one line explaining why. After ten rejections, patterns appear — and those patterns are almost always fixable upstream in the shot list rather than in a prompt.

A worked example: a forty-five-second product explainer

Suppose you are making a short explainer for a scheduling app. The spine is simple: a freelancer starts the day overwhelmed by scattered client messages, adopts one scheduling habit, ends the day in control. The viewer should feel relief.

The beat sheet has six beats: morning chaos, the specific failure point, discovery of the tool, first successful use, a montage of small wins, and a calm close. The shot list runs about fourteen shots. Shot sizes alternate wide, medium, and insert so no two consecutive shots feel the same. Two shots are static; the rest carry a single simple move. Three shots are screen-recording style and are produced in the editor, not generated.

Reference assets are one character sheet with fixed wardrobe, one location bible for the home-office desk, and one style block specifying soft window light, a warm neutral palette, and shallow depth of field. Model assignment follows shot type: identity-heavy shots go to the tool with the strongest character consistency, the laptop and coffee inserts go to a texture-strong option, and the establishing city shot goes to whatever handles wide scenes best. The grade is unified in the edit so the seams disappear.

In review, the continuity pass catches that the protagonist's coffee cup switches hands between shots. The story pass reveals that the discovery beat arrives too late, so one shot is deleted and the timing shifts by two seconds. Neither problem required regenerating hero footage. That is the return on planning: most fixes cost editing time rather than generation time.

Common mistakes to avoid

• Writing prompts before writing the spine. Prompts cannot fix an absent story.
• One prompt, one shot, no system. If every shot is described from scratch, consistency is luck.
• Over-specifying. Long, poetic prompts often perform worse than disciplined keyword blocks because they contain contradictory instructions. Cut every adjective that does not change a pixel.
• Ignoring runtime arithmetic. Count the words in your narration and divide by 2.5. Scripts that skip this always run long.
• Skipping reference images. Text-only consistency is fragile, and the failure appears late, when swapping in images is most disruptive.
• Generating before blocking. Geography errors are cheap on paper and expensive on screen.
• Chasing single-shot perfection. A good sequence with one soft shot beats a perfect shot inside a broken sequence.
• Treating the paperwork as optional. The shot list, character sheet, location bible, motion ledger, and decision log are the actual product of pre-production.

Frequently asked questions

How much of a video script should be written with AI assistance?

Use it for structure, options, condensation, and first drafts. Rewrite dialogue and the opening thirty seconds yourself, because those carry the most judgement per word and are the parts a viewer will judge most harshly.

Do I need storyboards if the final video is generated?

Yes, but lighter ones. You need blocking and reference frames, not detailed illustration. Even rough boxes on paper eliminate the geography and continuity errors that are hardest to fix later.

How do I keep a character consistent across many generations?

Lock a written character sheet, generate a reference portrait, and pass that image into every shot featuring the character. Change only pose, camera, and framing between prompts, and keep the style block identical across the whole project.

Should video and audio come from the same tool?

Not necessarily. Generating voice separately often improves control and makes pickups easier. The cost is alignment work in the editor, which is usually minor compared with the flexibility you gain.

How many attempts should one shot get before I move on?

Set a budget by shot type — often three to six attempts for hero shots and two for inserts. If a shot exceeds its budget, the problem is usually in the shot list, not in the wording of the prompt.

What is the single highest-leverage habit in this workflow?

Writing the story spine and the continuity notes before generating anything. Nearly every downstream frustration traces back to a decision that was never written down.

When should I stop planning and start generating?

When the spine fits in three sentences, the beats are numbered, the shot list has stable IDs, every recurring character has a reference portrait, every location has at least two angles, and the style block is frozen. That is usually a few hours of work for a short video — a fraction of the time you would spend regenerating inconsistent shots.

A checklist to carry into the next project

Before opening a generator, confirm that you have: a three-sentence spine and a one-sentence emotional target; a beat sheet with five to fifteen beats; a script paced to the target runtime; a tabular shot list with stable IDs and one action per row; rough blocking for every scene with two or more characters; a character sheet and reference portrait per recurring person; a location bible with multiple angles; a locked style block; a per-shot tool assignment; a motion ledger; a naming convention and versioning plan; and a four-pass review routine.

That list looks unglamorous next to a shiny render. It is also the difference between a folder of impressive clips and a video that holds a viewer from the first frame to the last. Generators will keep improving, prompts will keep getting shorter and smarter, and the tools will keep changing names. The part that stays hard — knowing what you are making and why — is still solved with a document, a table, and the habit of writing things down before you start generating.