Why the First Three Seconds Decide Everything
Short-form video is a brutal medium. A viewer decides whether to keep watching before they consciously register what they are looking at. That decision happens in roughly the time it takes to blink twice, and it is driven almost entirely by composition: where the subject sits in frame, how much motion is present, how legible the image is on a small screen, and whether something visually unresolved is happening.
Most AI-generated video fails not because the model is weak, but because nobody designed the shot. A prompt like "a futuristic city at night, cinematic" produces a pretty moving wallpaper. It does not produce a hook. The difference between the two is pre-production: a storyboard, a shot list, and a deliberate plan for what each shot is doing for the story and the retention curve.
This guide walks through a complete workflow for storyboarding and shot design in AI video production. It covers turning an idea into beats, converting beats into shot cards, writing prompts that models can actually follow, maintaining character and style consistency across shots, and finishing the result for the platform it will live on.
The Pre-Production Pipeline for AI Video
The temptation with generative tools is to skip straight to generation. Resist it. A thirty-minute planning pass saves hours of re-rolling, and it is the single highest-leverage habit in AI video production.
Step 1 — Lock the single-sentence promise
Before anything visual, write one sentence that states what the viewer gets. Not the plot — the payoff. "You will see a metal sculpture melt into a building and reassemble as a skyline." If you cannot write that sentence, you do not have a video yet; you have a mood.
Step 2 — Break the idea into beats
A short video usually runs on three to five beats. In a fifteen-second piece, that is roughly three seconds per beat. Write each beat as a plain-language action: reveal, escalate, turn, resolve. Beats are narrative units, not shots. One beat may need three shots; a different beat may need one long take.
Step 3 — Convert beats into a shot list
A shot list is a table. At minimum it should carry: shot number, duration, framing, subject action, camera behaviour, lighting note, and audio note. Keep it in a spreadsheet or a plain markdown file so it can be revised quickly.
Step 4 — Define the visual bible
Before generating anything, decide and write down the constants: colour palette, aspect ratio, lens character, grade, wardrobe, and the way light behaves in the world. These constants are what make a sequence of separately generated clips feel like one film. If they live only in your head, they will drift by shot four.
Translating Narrative Intent into a Structured Storyboard
A storyboard is not an art project. It is a compression format for directorial intent. For AI video work, you can build one with rough frames, with reference stills, or even with pure text — the format matters far less than the discipline of filling it in.
A practical storyboard grid for generative work looks like this:
- Shot ID and duration — so you can later match cuts to a music bed.
- Framing — wide, medium, close, extreme close, over-the-shoulder.
- Subject and action — one verb per shot, ideally.
- Camera behaviour — static, push in, pull out, pan, tilt, orbit, handheld.
- Lighting and time of day — the most commonly forgotten column.
- Continuity anchors — which props, garments, or set elements must match the previous shot.
- Prompt draft — the raw text you will feed the model.
Filling the continuity column is what separates a storyboard from a mood board. If shot two shows a character holding a red umbrella and shot three shows the same character in the rain with empty hands, the audience may not consciously notice — but they will feel the sequence as incoherent.
Storyboarding for retention, not for beauty
When you place frames side by side, ask a different question than usual: does each frame look meaningfully different from the last? Rapidly cutting between five visually similar medium shots produces a sequence that feels static even though it is technically edited. Vary scale aggressively — a wide, then a face, then a detail of a hand, then a wide again. Visual contrast is what makes an edit feel alive.
Shot Design Fundamentals That Still Apply to AI Footage
Generative models are excellent at texture and mediocre at intent. You supply the intent through classic shot design vocabulary, which translates surprisingly well into prompt language once you know what each term is doing.
Framing and negative space
Decide where the subject sits. Centred framing reads as graphic and poster-like; rule-of-thirds framing reads as observational. Negative space can carry meaning — a subject pushed to the far left with empty architecture to the right implies motion into that space. Say this explicitly in your prompt with phrases like "subject in the left third, large empty space to the right."
Camera movement
Movement is the strongest signal of production value and the most common source of artefacts. Slow, single-axis moves hold up best: a gentle dolly in, a slow orbit, a lateral truck. Complex combined moves tend to produce warping geometry and melting edges. If you need energy, get it from cutting fast between static-but-dynamic shots rather than from one chaotic camera path.
Depth of field and parallax
Depth is what separates amateur-looking AI footage from cinematic-looking AI footage. Specify a foreground element, a mid-ground subject, and a background layer, then ask for shallow depth of field. Foreground occlusion — a railing, a plant, a passing figure — instantly adds dimensionality and reads as intentional framing rather than a generated image with motion.
Lighting continuity
Pick a light direction and hold it across the sequence. If shot one has a hard key from the left, shot five should not have soft light from behind. Consistent light direction is one of the fastest ways to make separately generated clips feel like they were captured at the same time in the same place.
Writing Shot Prompts That a Video Model Can Follow
A reliable shot prompt has a fixed order. Once you internalise the order, you can write a shot in under a minute.
- Subject — who or what, described specifically.
- Action — one clear verb, present tense.
- Setting — environment plus time of day.
- Framing — shot size and subject placement in frame.
- Camera — movement type and speed.
- Optics — lens character, depth of field, foreground occlusion.
- Light — direction, quality, colour temperature.
- Style — grade, texture, film or digital feel.
- Format — aspect ratio and duration.
A weak prompt: "a robot in a city, cinematic, dramatic."
A strong prompt: "A rusted humanoid robot walks slowly toward the camera through a flooded street at dusk. Medium shot, subject centred, water reflecting orange sky. Camera dollies backward at walking pace. Shallow depth of field with a blurred foreground cable. Hard key light from the left, warm sodium street lamps, cool blue shadows. Muted teal-and-amber grade, fine grain. Vertical 9:16, four seconds."
The second prompt is not longer for the sake of length. Each clause removes a decision the model would otherwise make randomly.
One verb per shot
Generative video handles a single clear action far better than a chain of them. "She turns, picks up the cup, and walks out" will usually produce a muddled turn with a ghost cup. Split it into three shots. The edit will be more dynamic anyway.
Negative space in the prompt
Describe what you do not want, briefly. Warped hands, extra limbs, jittery geometry, text artefacts, and sudden lighting shifts are worth naming in a short negative list. Keep it short — long negative lists dilute the positive description.
Consistency: Locking Characters, Wardrobe, and Style
Consistency is the hardest problem in multi-shot AI video. It is also entirely solvable with discipline.
Character locking
Create a reference set for each character: one clean portrait, one full-body, and two or three expressions. Reuse the same descriptive language every single time the character appears, word for word. Changing "short black bob" to "short dark hair" in shot six is enough to produce a visible drift. Keep a text file of character descriptions and paste from it rather than retyping.
Wardrobe and prop locking
Name every visible element you care about: jacket colour, logo placement, bag, scar, vehicle. Then verify in the storyboard's continuity column that these elements appear in every shot where they logically should. Props are continuity anchors that viewers track unconsciously.
Style locking
Style drift usually comes from prompt drift. Fix it by building a reusable style block — grade, grain, lens character, contrast curve, colour palette — and appending it verbatim to every shot prompt. The style block should never change mid-sequence. If you want a shift, make it a deliberate act at a beat boundary.
When to use image-to-video instead of text-to-video
Text-to-video is fast and exploratory. Image-to-video gives you far greater control. Once you have approved a look through still generation, use those stills as the first frame of each shot. This converts consistency from a prompting problem into a selection problem, which is much easier to solve.
Worked Example: A Thirty-Second Teaser in Eight Shots
Here is how the pipeline comes together on a realistic brief: a thirty-second teaser for a fictional architecture studio, built around the idea of metal morphing into structure.
Shot 1 — Hook (2s). Extreme close-up of molten metal rippling. Static camera, hard rim light from the right, macro lens. Prompt emphasises the surface texture and slow motion.
Shot 2 — Context (3s). Wide shot of an empty industrial hall at dawn, light shafts from high windows. Slow dolly forward. Cool grey palette, one warm accent.
Shot 3 — Introduction of subject (4s). Medium shot of a figure in a dark coat walking through the hall, back to camera, left third of frame, large empty space ahead. Camera trucks laterally at walking pace.
Shot 4 — Detail (2s). Close-up of a gloved hand placing a metal shard on a workbench. Shallow depth of field, foreground blurred tool edge.
Shot 5 — Transformation (5s). The shard rises and unfolds into a structural lattice. Slow orbit, low angle, cool key light with a warm bounce from below.
Shot 6 — Escalation (4s). The lattice expands through the hall, filling the frame. Camera pulls back fast — the one high-energy move in the sequence.
Shot 7 — Reveal (5s). Wide exterior at sunrise: the finished building, the same warm accent from shot two now dominant. Static camera, subtle parallax from a passing foreground figure.
Shot 8 — Logo and loop (3s). Slow push in on a reflection in the building's glass, fading to a holding frame that mirrors the molten texture of shot one, so the video can loop cleanly.
Notice what the shot list bought you: a light direction that evolves with the story, a colour palette that shifts from cool to warm as the idea resolves, a single high-energy camera move placed at the emotional peak, and a final frame designed for the loop. None of that comes from a better model. It comes from a plan.
Editing, Pacing, and Platform Finishing
Generation is roughly half the work. Finishing decides whether the result feels professional.
Cut on movement
Place cuts where motion is already happening — mid-gesture, mid-step, mid-expansion. Cutting on a static frame draws attention to the seam. Cutting during movement hides it, because the eye is tracking the subject rather than the edit point.
Match duration to the platform
Vertical short-form rewards a hook in the first second and a payoff before the five-second mark. Horizontal formats tolerate a slower build. Whatever the platform, keep individual shots short — most AI shots reveal their weaknesses after four or five seconds, so cutting earlier is usually an upgrade, not a compromise.
Sound carries perceived quality
A clean ambient bed, one impact sound on the transformation beat, and a music drop on the reveal will do more for perceived production value than doubling your render time. Design sound to the shot list, not after the fact.
Captions and safe areas
Keep captions out of the bottom tenth of the frame and away from the right-hand rail where platform UI sits. If a key visual detail lives in that zone, it will be covered. Check a frame with the caption overlay enabled before you export.
Common Mistakes and How to Fix Them
Everything is a medium shot. Vary scale deliberately. If three consecutive shots share a framing, replace one with an extreme.
The style block changes mid-sequence. Freeze it. Variation belongs to subject and light, not to grade.
Too many actions per shot. Split them. One verb per shot is the rule that saves the most renders.
Camera moves that fight the subject. A fast orbit around a fast-moving subject produces the worst artefacts. Pick one source of energy per shot.
No continuity audit before assembly. Watch the sequence with the shot list open and tick off every continuity anchor. Fixing a mismatched jacket costs one regeneration; fixing it after publication costs a re-upload.
Generating before the storyboard is done. This is the mistake that makes AI video feel random. Planning is faster than re-rolling.
FAQ
Do I need drawing skills to storyboard AI video?
No. A shot list in a spreadsheet is a storyboard. Rough frames help, but the columns — framing, action, camera, light, continuity — are what actually direct the sequence.
How long should each AI-generated shot be?
Two to five seconds for short-form work. Generate slightly longer than you need so you have handles to trim, and cut before the model starts to drift.
Should I generate in vertical or horizontal first?
Generate in the aspect ratio you will publish. Cropping a horizontal render into vertical costs you composition and resolution, and vertical framing changes the shot design — you need more headroom and tighter subject placement.
How do I keep a character consistent across ten shots?
Create a reference image set, store a fixed written description, and paste that exact wording into every prompt. Where possible, drive shots from approved stills rather than from text alone.
What is the fastest way to improve my output quality?
Write one verb per shot, keep a frozen style block, and cut on movement. Those three changes alone will lift a sequence noticeably.
How many shots does a thirty-second video need?
Usually eight to fourteen. Fewer than eight means shots are too long; more than fourteen means you are cutting faster than the audience can read the images.
Putting the Plan First
AI video tools will keep improving. Composition, pacing, and continuity will keep mattering. Every hour spent on a shot list and a visual bible is repaid with fewer regenerations, cleaner continuity, and a sequence that holds attention past the first three seconds — which is the only metric that actually decides whether a video gets seen. Build the plan, then generate. The results will look deliberate, because they will be.
