Why Story Design Still Decides Whether an AI Video Works
Generative video tools are now good enough that a single shot can look genuinely cinematic. What they still cannot do is decide what the story needs next. Choosing which shot, from which angle, at which emotional moment, is the work of story design and shot listing, and it remains the highest-leverage activity in any AI-assisted production pipeline.
Teams that skip this stage usually end up with footage that looks impressive in isolation and incoherent in sequence. A character changes jackets between two shots of the same conversation. A room rearranges itself mid-scene. The emotional arc flattens because every shot carries the same intensity, so nothing lands.
Treating pre-production as a real discipline fixes most of this. This guide walks through a repeatable workflow: break a script into visual beats, convert those beats into camera parameters, assemble a shot board, and protect continuity while you generate. The specific tools matter less than the order in which you use them.
The Pre-Production Chain: From Idea to Shootable Shot List
A shot list is not a document you write once. It is the middle link in a chain of artifacts, each one narrower than the last. When a project goes wrong, it is almost always because one link was skipped.
The four artifacts every project needs
1. The beat sheet. One page. Eight to twenty lines, each describing a change in the story: a decision, a reveal, a reversal. No camera language yet. If a line does not change anything about the situation, it does not belong on the beat sheet.
2. The shot list. A numbered table where each row is one generated clip. It records what the audience sees, not what the writer imagined: subject, action, shot size, camera movement, location, time of day, and duration.
3. The shot board. The same rows as the shot list, plus a visual reference for each row. The reference can be a sketch, a still from a reference film, a generated keyframe, or a simple blockout. Its purpose is to make continuity failures visible before they cost generation time.
4. The continuity bible. A living document holding character descriptions, wardrobe, props, colour palette, lighting direction, and the reference images that define each. Every shot board row points back to entries here.
Where AI genuinely helps, and where it does not
AI is excellent at the mechanical parts: parsing a script for named entities, summarising scenes, proposing shot coverage, generating keyframe thumbnails, and reformatting a spreadsheet into consistent prompt syntax. It is unreliable at taste. A model will happily suggest twelve shots where four would be stronger, because it optimises for plausibility rather than restraint.
The practical division of labour is: humans decide the beat sheet and the emotional intent, AI drafts the shot list, humans prune it, AI generates the visuals, humans judge whether the sequence works.
Breaking a Script Into Visual Beats
Extracting characters, locations, and props
Feed the script to any capable language model and ask for three lists: named characters, distinct locations, and objects that are touched, moved, or shown in close-up. Then ask a second question: which of these appear in more than one scene? Those recurring elements are your continuity risks, and they deserve reference images before anything else is generated.
Push further with a structured output. A useful schema is one row per scene with columns for characters present, location, time of day, emotional tone, and the single most important visual detail. That last column is where directors earn their money: the detail that tells the audience where to look.
Mapping the emotional curve
Shot lists fail most often because they are emotionally flat. Fix this by assigning each beat a value from minus five to plus five and plotting it. A scene that sits at plus four for its entire length has nowhere to go; a scene that travels from minus two to plus four has shape.
Once the curve exists, it constrains the shot list in useful ways. High-intensity beats tolerate handheld motion, tight framing, and short clips. Low-intensity beats want stillness, wider framing, and longer holds. The contrast between them is what makes the high points read as high points.
Turning Narrative Beats Into Camera Parameters
Shot size, angle, movement, lens
Every shot in your list should specify four things: shot size (wide, medium, close, extreme close), camera angle (eye level, low, high, overhead), camera movement (static, pan, dolly in, handheld), and implied lens (wide-angle for spatial context, long lens for compression and isolation). These four parameters do more for perceived production value than any amount of post-processing.
Be deliberate about repetition. Reusing an identical framing for two different scenes creates an unintended visual rhyme; audiences read it as meaningful even when it is accidental. If you want a callback, plan it. If you do not, vary at least one parameter.
Writing prompts a generator can actually obey
The most common prompt error is overload. A prompt that stacks six stylistic adjectives, three lighting conditions, and two camera moves gives the model contradictory instructions, and the output resolves the conflict randomly. Write prompts that read like a shot description on a call sheet:
- Subject and action, stated plainly.
- Shot size and angle.
- Lighting direction and quality.
- One, or at most two, style anchors.
- Continuity references, listed separately from the creative description.
Keep a house style block, the paragraph of consistent lighting, palette, and texture language you paste into every prompt, and keep it short enough that you can maintain it by hand. Consistency across a project comes from repeating the same style block, not from adding more words to each prompt.
Translating intent into parameters
A shot list row is a translation. 'She realises she has been lied to' is intent. 'Close-up, eye level, slow dolly in, soft side light, two seconds' is a parameter set. Practise the translation explicitly for the first ten shots of a project; after that the pattern becomes fast, and you will start writing shot lists that a generator can execute without guesswork.
Building a Shot Board That Survives Production
The columns that matter
A shot board with too many columns becomes a document nobody updates. Start with these: shot number, scene, description, shot size, angle, movement, duration, character references, location reference, style block, status. Then add columns only when a specific failure proves you need them.
The status column is the one people forget and the one that saves the most time in a long project. Use five states: drafted, prompted, generated, approved, revised. Any shot that has been generated but not approved is a liability, and seeing twenty of them at once is a reliable signal that the sequence needs rethinking rather than more generation.
Thumbnail discipline and versioning
Attach a thumbnail to every row as soon as a keyframe exists. A shot board with images reveals problems a text list hides: three consecutive shots at the same framing, two characters with nearly identical silhouettes, a location that looks like a different location each time it appears.
Version numbering should be boring: shot 14 gets versions 14a, 14b, 14c, with the approved version marked in the status column. Overwriting a file named final-final is how working sequences get destroyed.
Continuity: Keeping Characters and Spaces Consistent
Reference sheets, seeds, and style anchors
Consistency is a pre-production problem disguised as a generation problem. Before generating scene one, build a character sheet: one clean portrait, one three-quarter view, one full-body shot with wardrobe, and a written description of roughly forty words covering age, build, hair, wardrobe, and one distinguishing feature.
Do the same for every recurring location: a wide establishing image, a detail image, and a note about lighting direction. Where are the windows, and which side of the room does the key light come from? When a location reappears at a different time of day, change the light, not the architecture.
Where your tools support a seed or reference image parameter, reuse the same seed family for a character across shots, and vary the prompt around it. Where they support reference images, attach two: one for the subject and one for the environment. Two well-chosen references beat six competing ones.
Start-and-end frame conditioning for transitions
Many generators now let you supply a first frame, a last frame, or both. This is the cleanest way to hide a cut: end clip A on a frame that matches the first frame of clip B, and the join becomes invisible. It also gives you a way to control where a movement lands, which matters for match cuts and for shots that need to hand off to the next scene.
The workflow cost is real, so use it selectively. Reserve start-and-end frame conditioning for transitions that the story actually leans on: the reveal, the match cut, the moment where a character's decision changes the direction of the scene.
A Practical Workflow: The 90-Minute Shot List Sprint
This is the process that fits an afternoon and produces a shot list you can actually generate from.
Minute 0-10: Read the script out loud. Mark every line where something changes. Those marks become the beat sheet.
Minute 10-25: Build the entity lists. Characters, locations, recurring props. Flag anything appearing in two or more scenes as a continuity item.
Minute 25-40: Draft coverage. For each beat, write two to five shots. Do not polish. Aim for coverage, then edit.
Minute 40-55: Prune. For every shot, ask what it does that the previous shot does not. Delete anything that only restates. A ten-shot scene that becomes six is a stronger scene.
Minute 55-70: Assign parameters. Shot size, angle, movement, duration, and style block for each surviving row.
Minute 70-80: Generate thumbnails. Rough keyframes for the first shot of every scene, plus any shot with complex blocking. Sketches work fine here.
Minute 80-88: Check continuity. Walk the board left to right and name the subject, wardrobe, location, and light for each row out loud. Anything you cannot state confidently is not ready.
Minute 88-90: Freeze the list. Export it, mark it as the working version, and note the date. Changes after this point are revisions, not drafts.
The point of the sprint format is not speed for its own sake. It is that a shot list written in one sitting tends to be coherent, while a shot list assembled over a week of scattered decisions tends to be a collection of unrelated good ideas.
Choosing Tools: Decision Criteria
Script and beat analysis
Look for structured output. A tool that returns a clean table you can export is worth more than one that returns beautiful prose you have to retype. Check whether it can hold a project-level memory of characters and locations, and whether you can correct it when it misreads a scene.
Generation and shot control
Prioritise start-and-end frame control, reference image inputs, and aspect ratio variety over raw resolution. Resolution is easy to buy later; control is what determines whether a sequence works. Also check clip length limits honestly, because a five-second ceiling shapes the shot list you are able to write.
Review and asset management
The unglamorous criterion that matters most. Can you attach a thumbnail to a row? Can you mark a version as approved? Can you find shot 27 from three weeks ago in under ten seconds? If not, your review process is the bottleneck, not your generation quality.
Common Mistakes and How to Fix Them
Writing shots as moods. A line like 'melancholy city shot' gives a generator nothing to hold. Fix: name the subject, the framing, and the light.
Repeating the same framing. Three medium shots in a row read as an accident. Fix: alternate wide, medium, and close, and vary the angle when you return to a size.
Ignoring duration at the planning stage. A shot list without durations produces a sequence you cannot time to music or narration. Fix: assign seconds per shot and add them up per scene.
Overloading the style block. Six style adjectives cancel each other out. Fix: two anchors, applied everywhere.
Generating before continuity references exist. This is the expensive mistake. Fix: build the character and location sheets first, even if they feel slow.
Never deleting shots. Every unused shot in the board is a decision deferred. Fix: prune at every review, not once at the end.
Changing two variables at once. If you regenerate a shot with a new angle and a new wardrobe at the same time, you cannot tell which change fixed it. Fix: one variable per iteration, then lock it.
FAQ
How many shots should a one-minute video have?
Roughly fifteen to twenty-five for a narrative piece with dialogue, fewer for a mood-driven piece. If your count is much higher, you are probably cutting faster than the story justifies. Let the beat sheet set the count, not a target number.
Can I skip the shot board and prompt directly from the script?
You can, and it shows. The board is what makes continuity visible. Even a minimal version, a spreadsheet with a thumbnail column and a status column, is enough to prevent most of the failures that ruin long projects.
What is the best way to keep a character consistent?
Three things, in order: a written description you never change, one to three reference images used every time, and the same style block in every prompt that features them. Consistency comes from repetition and restraint, not from more descriptive words.
How do I decide when a shot is approved?
Ask one question: if every other shot in the scene matched this one in quality and intent, would the scene work? If yes, approve it. If you are approving shots because regenerating them feels tedious, you are about to build a sequence you cannot fix in editing.
Do I need storyboards, or is a shot list enough?
A shot list is enough for simple sequences with one subject and stable framing. Add a board when you have more than one character in frame, when blocking matters, or when a location reappears across scenes.
How should I handle scenes that change time of day?
Change lighting direction and colour temperature, not set dressing. If the architecture moves, the audience reads it as a different place, and the scene loses its anchor.
What should I do when a generated shot is almost right?
Write down exactly which single parameter is wrong before touching the prompt. Almost-right shots usually fail on one axis: framing, movement, lighting, or wardrobe. Change that axis, keep everything else identical, and regenerate once. Repeating the same prompt rarely improves the result; changing the right variable almost always does.
