AI video tools have made it easy to produce a beautiful shot. They have not made it easy to produce a story. Most disappointing AI shorts do not fail because the render quality is poor. They fail because the script drifts, the character's jacket changes color between scenes, and the camera keeps resetting to the same medium shot.
Pre-production discipline is what separates a clip from a film. This guide walks through a complete, tool-agnostic workflow for AI-assisted scriptwriting and shot design: how to structure a story, how to translate it into a shot list, how to keep characters and places consistent across dozens of generations, and how to assemble everything into something that holds attention. It is written for solo creators, small studios, and anyone who wants a repeatable pipeline instead of improvising one video at a time.
Why pre-production is the real bottleneck in AI video
Generative models are probabilistic. Give the same prompt twice and you get two different results. That randomness is a feature when you are exploring a look, and a liability when scene 12 has to match scene 3. Every decision you leave undefined becomes a variable the model will resolve differently on each pass.
Pre-production for AI video works differently than on a traditional set. You are not scheduling crew and locations. You are defining constraints. A locked shot list, a fixed character description, a defined palette, and a consistent lens language are the guardrails that make generation predictable. The more precisely you specify, the less you regenerate, and regeneration is where most project time disappears.
There is also a narrative cost to skipping this phase. Without a beat sheet, AI-assisted scripts tend to produce three variations of the same scene: a character explains the problem, a character reacts to the problem, a character resolves the problem. The result reads like a summary rather than a story. Structure creates escalation, and escalation is what makes an audience keep watching.
The practical rule: spend roughly one fifth of your project timeline in pre-production. It will save you more than that in generation and editing, and it will improve the finished piece at the same time.
Stage 1: Turn a logline into a beat sheet
A logline is one sentence. A beat sheet is the skeleton that tells you what actually happens, in what order, and why each moment matters. This is the stage where an AI assistant is genuinely valuable, because it can generate and compare structural options faster than you can outline by hand.
The three questions every beat sheet must answer
Before you accept any structure, check it against three questions:
- What does the protagonist want, specifically enough that you could film it?
- What stands in the way, and how does that obstacle change over time rather than simply repeating?
- What is measurably different at the end of the story?
If the obstacle never changes, you have a loop, not a plot. If nothing is different at the end, you have a mood piece. That is fine if it is intentional, but you should know which one you are making.
Working with an assistant as a structural partner
Ask for three competing structures rather than one. For example: a linear escalation, a structure that opens on the ending and works backwards, and a structure built around a single location where pressure accumulates. Comparing three outlines side by side reveals what your premise is actually about.
Once you pick a direction, push on it. Ask where the tension sags. Ask which beat is doing no work and could be cut. Ask what the second act would look like if the protagonist made the opposite choice at the midpoint. These are cheap questions at the outline stage and expensive ones after you have generated forty shots.
Keep the beat sheet to one line per beat. It should fit on a single page. If a beat needs a paragraph to explain, it probably contains two beats.
Validate before writing dialogue
Read the beat sheet out loud as if narrating it to someone. Where you stumble or add explanation, that is where the story is under-built. Fix it here rather than in the script, because every structural problem becomes three times harder to solve once dialogue is attached to it.
Stage 2: Drafting the script with an AI assistant
With a locked beat sheet, the script becomes an exercise in expansion rather than invention. You already know what happens; now you are choosing how it is revealed.
Structure analysis: arcs, themes, and scene density
A useful technique is to have an assistant read your draft and tag it in layers. Map the emotional curve scene by scene, note which theme each scene touches, and estimate how much narrative weight each scene carries. When several consecutive scenes carry the same weight and hit the same emotional note, the middle will feel flat. You do not need an elaborate report, just a clear view of where the shape sags.
The same pass can flag scenes that exist only to deliver information. In short-form AI video, where every second competes for attention, a pure exposition scene is a candidate for compression into a single line of voiceover or a visual detail.
Dialogue that survives text-to-video
AI-generated visuals struggle with subtle performance. Long speeches from a static character look unnatural and waste the medium. Practical adjustments:
- Keep lines short enough to fit a single shot, ideally under fifteen words.
- Give characters something physical to do while they speak, so the frame has motion.
- Prefer conflict in dialogue over explanation, and let visuals carry context.
- Reserve one or two longer monologues for moments that genuinely earn them.
Formatting for handoff
Write the script in a consistent format with clear scene headings, action lines, and dialogue blocks. Even if you are the only person reading it, structure helps you search, cut, and reorder later. Include a brief slug line with time of day and location, because those details become prompt components downstream.
Stage 3: Designing shots before you generate anything
This is the step most creators skip, and it is the single largest quality lever. A shot list is not bureaucracy; it is the plan that prevents you from generating attractive but unusable footage.
Anatomy of a usable shot list
Build a simple table with one row per shot:
- Shot ID, consistent with your script scene numbers
- Scene and beat reference
- One-sentence description of what the audience must understand
- Shot size: wide, medium, close-up, insert
- Camera behavior: static, slow push, pan, handheld drift, orbit
- Approximate duration in seconds
- Audio note: dialogue, ambience, music cue
- Generation note: reference image, style anchor, or model choice
The description column is the important one. If you cannot state what a shot communicates in one sentence, it is decorative, and decorative shots are the first thing to cut in the edit.
Shot sizes and moves that models handle well
Generative video is strongest with clear, single-variable motion. A slow push in on a face, a lateral pan across a landscape, and a static wide with moving elements in frame all translate reliably. Rapid, compound camera movement, complex hand interactions, and crowds passing in front of the lens tend to produce artifacts. Design around the strengths rather than fighting them.
Also think in coverage. Generate two or three options for key beats: a wide establishing version, a medium performance version, and a detail insert. Having choices in the edit is what lets you control rhythm, and options are cheap compared to reshooting later.
Build edit points into the plan
Before generation, mark where you intend to cut. A scene that is meant to be cut in three places should be planned as three shots, not one long take trimmed afterwards. Decision points written down in advance keep the pacing intentional instead of accidental.
Stage 4: Locking character and location consistency
Continuity is the hardest problem in AI video and the most visible failure. The solution is documentation, not luck.
Character sheets
For every recurring character, write a fixed description that you reuse verbatim in prompts: approximate age, build, hair, wardrobe, distinguishing features, and overall style register. Then generate a small set of reference stills and keep the best one as the canonical anchor. Image-to-video and reference-conditioned workflows will hold identity far better than a text description alone.
Never paraphrase a character description between shots. The moment you write the same person two different ways, you introduce variation the model will happily exploit.
Location bibles and lighting continuity
Do the same for places. A location bible contains a reference frame, the materials and colors that define the space, and a note on time of day and light direction. If a scene takes place at dusk, keep the light warm and low across every shot in that scene. Changing the lighting between shots of the same conversation reads to an audience as a mistake, even if they cannot articulate why.
Keep a continuity log as you generate. One line per accepted shot noting wardrobe, props, light direction, and which reference frame you used. It is a two-minute habit that prevents hours of rework.
Stage 5: The generation pipeline
With a script, shot list, and reference library ready, generation becomes mechanical, which is exactly what you want.
Batching and prioritization
Group shots by character, location, and style so that once you have tuned a prompt, you can reuse it almost unchanged. Prioritize hero shots first: the opening, the climax, and any shot the story depends on. If a hero shot cannot be made to work, you need to know that before you have generated thirty supporting shots around it.
Prompt templates that keep style stable
Create a reusable prompt skeleton with slots for subject, action, shot size, camera move, lighting, and style. Write the style clause once and paste it into every prompt in the project. Consistency in the prompt is consistency on screen.
Keep a changelog when you adjust the template mid-project, and note which shots used the older version. If the look shifts halfway through, you will want to know why before you start regenerating.
When to stop iterating
Set a hard iteration limit per shot, typically three attempts. If a shot is not working by the third, the problem is usually conceptual rather than technical. Change the shot design rather than the prompt.
Stage 6: Assembly, rhythm, and sound
Editing AI footage is different from cutting live-action, because the raw material rarely has performance continuity. You are building rhythm from scratch.
Cut on motion and on contrast. If two consecutive shots are both slow and wide, the sequence will feel static regardless of content. Alternate scale and pace deliberately. Use inserts and detail shots as connective tissue between larger beats, and let sound carry transitions where visuals would need to match too precisely.
Sound design does more work here than in traditional video. A consistent ambience bed under a scene makes separately generated shots feel like one place. Music cues can bridge cuts that would otherwise feel abrupt. If you use synthetic voice, keep one voice per character across the whole project and match pacing to the edit rather than the other way around.
Finally, watch the assembled piece with the sound off, then with the picture off. If the story is unclear in either pass, the issue is in structure, not in polish.
A quality control checklist before you export
Run through this list once, at the end, with fresh eyes:
- Does the first five seconds establish a question the viewer wants answered?
- Is every character visually identical to their reference across all appearances?
- Do lighting and time of day stay consistent within each scene?
- Does the cut rhythm vary rather than settling into one tempo?
- Is any shot present that the story would not miss?
- Do dialogue lines fit the shots they are attached to?
- Is the audio balanced, with no jarring level jumps between scenes?
- Does the ending resolve or deliberately refuse to, rather than simply stopping?
Anything you hesitate on is usually a cut, not a keep.
Common mistakes and how to avoid them
Writing the script and shot list at the same time. They inform each other, but doing both at once produces shots that serve sentences instead of scenes. Finish the script first.
Changing character descriptions mid-project. Even a small variation in wording creates visible drift. Copy and paste, never retype.
Over-relying on a single long take. Models lose coherence over extended durations. Several shorter shots cut together almost always look better and are easier to control.
Generating before designing. Generating fifty clips and then looking for a story among them is the most common way to waste a week.
Ignoring sound until the end. Sound is a structural tool, not a finishing layer. Plan ambience and music alongside the shot list.
FAQ
How much of the script should AI write?
Use it for structure, options, and rewriting. Keep final dialogue decisions human, because voice and specificity are what make a script feel authored rather than assembled.
Do I need a shot list for a short video?
Especially for a short video. Short pieces have no room for filler, and a shot list is the fastest way to see which shots are unnecessary before you spend time generating them.
How do I keep a character consistent across many shots?
Lock a written description, generate reference stills, choose one canonical frame, and use image-conditioned generation wherever possible. Reuse the exact same wording in every prompt.
What if a shot simply will not generate correctly?
Change the shot. Simplify the camera move, reduce the number of subjects, shorten the duration, or communicate the same information from a different angle. Chasing a difficult shot rarely pays off.
Can this workflow scale to a team?
The structure is the point. A shared beat sheet, shot list, and continuity log let several people generate footage that cuts together, because everyone is working from the same constraints rather than from individual interpretation.
Where should a beginner start?
Start with a one-page beat sheet and a six-shot list for a thirty-second piece. Completing one small project end to end teaches more than planning a large one indefinitely.
The tools will keep changing. The discipline of defining structure, shots, and continuity before generation does not. Build that habit once and every new model becomes easier to use well.
