Plan AI video like a film crew: shot sheets, prompt structure, camera language, native audio, continuity, iteration limits, and editing that hides the seams.
Most AI video projects fall apart before the first clip is generated, and rarely because the model is weak. They fall apart because nothing was planned: no shot list, no sense of what five seconds can hold, no idea how one clip connects to the next. A scene-based canvas — one workspace where you generate shots, chain them together, and reuse image and audio references — makes the plan visible. It does not write the plan for you.
This guide is about the plan. It covers how to define a deliverable, build a shot sheet a model can follow, write prompts that survive generation, direct camera and sound, hold continuity, budget your attempts, and cut the pieces into something that feels authored rather than sampled.
Start With the Deliverable, Not the Prompt
Before you open a generation canvas, write six constraints into a plain text file. Every later decision inherits from them.
- Runtime. Ten seconds, thirty seconds, ninety seconds. A short runtime forces discipline and exposes weak shots instantly.
- Aspect ratio. Vertical for social feeds, 16:9 for web and presentation, square for some placements. Decide now; reframing later is painful.
- Audio expectation. A mood piece with a music bed is a different project from a scene where someone speaks on camera.
- Visual register. Documentary realism, commercial polish, animation, archival grain. Pick one and stay inside it.
- Text policy. On-screen words belong in the edit. Generated typography is unreliable and dates quickly.
- The last shot. Decide the ending first. Everything before it is a runway.
Then apply the one-sentence test: if you cannot describe the video in a single sentence, the shot list will sprawl. "A runner tests a new shoe on wet city streets before dawn" is a deliverable. "Something inspiring about movement" is a wish.
The distinction matters because generation tools are compliant rather than opinionated. They will happily produce forty beautiful clips that share no argument. A deliverable is the argument.
Also write down who watches it and where. A fifteen-second vertical teaser for a product drop and a ninety-second widescreen brand film are not the same project at different lengths; they are different grammars. One rewards punch and pattern interrupts, the other rewards patience and a slow build. Finally, note what you will not attempt: no crowd scenes, no complex hand interactions, no readable signage. Naming your constraints up front turns them into style instead of failure.
Build a Shot Sheet the Model Can Follow
Professional production does not begin with a camera. It begins with a breakdown. Do the same, even for a fifteen-second clip.
The five-column sheet
| Shot | Story job | Framing and camera | Action | Audio |
|---|---|---|---|---|
| 1 | Establish place | Wide, static, 24mm | Empty street, rain, neon reflections | Rain, distant traffic |
| 2 | Introduce subject | Medium, slow push in | Runner ties a lace, exhales | Breath, fabric |
| 3 | Escalate | Low tracking, handheld | Runner accelerates past camera | Footsteps, splash |
| 4 | Detail | Extreme close-up, shallow focus | Sole grips wet asphalt | Compressed water hiss |
| 5 | Resolve | Wide, crane up | Runner disappears into fog | Music swell, fading steps |
Five shots at roughly five seconds each produce a coherent piece. Notice that each row has exactly one story job. That is the single most useful constraint in AI video, because models handle one clear action plus one clear camera behavior far better than a paragraph of simultaneous events.
The word "and" is a warning
If a shot description contains "and" twice, split it. "She opens the letter and gasps and the camera pushes in" is three decisions competing for the same five seconds. "She opens the letter. Camera pushes in slowly" is one beat the model can deliver.
Name the shot you are afraid of
Every project has one shot that decides whether the idea works: the macro pour, the profile turn, the hand-off between two characters. Find it in the sheet and generate it first. If the hardest shot fails after a fair attempt, the concept needs adjusting, and you have learned that in minutes instead of after a day of polishing establishing shots.
Prompt Anatomy for Clips That Hold Together
Prompts that work consistently follow the same internal order, because they mirror how a cinematographer thinks. Use this sequence as a checklist rather than a rigid formula:
Subject → action → environment → camera → light → texture → audio.
Here is one shot written in that order:
A woman in her thirties in an olive rain jacket stands at the edge of a rooftop, turning her head slowly toward the skyline. Overcast dawn, wet concrete, distant apartment windows with a few lights on. Camera: slow dolly forward at chest height, 50mm lens, shallow depth of field. Light: cool blue ambient with a warm sodium glow from below. Texture: fine film grain, muted contrast, documentary realism. Audio: wind over a microphone, faint city hum, her slow breathing.
Three things make this strong. It is specific about scale (50mm, chest height), it separates atmosphere from action, and its audio instruction is physically plausible.
Use reference images as anchors
When the tool accepts a reference image, stop describing what the image already shows. Describe only what changes: action, camera, and audio. Redundant description is a common cause of stiff output, because the model tries to reconcile text with pixels that already agree.
What to leave out
- Adjectives with no visual meaning. Gritty, epic, stunning. Replace them with what a camera would capture.
- Contradictions. Golden hour and noon sun in one scene produce mush.
- Long dialogue. Two short sentences per clip is a realistic ceiling; longer lines break rhythm and lip sync.
- Multiple camera moves. Pick one primary move per shot.
- Style names without specifics. "Cinematic" means nothing; "side-lit with soft shadow on the wall" means something.
Camera and Lens Language That Actually Works
Models respond best to a small, stable vocabulary. Learn these and stop guessing:
- Static or locked off. The safest choice and the most underused. Ideal for establishing shots and reaction beats.
- Slow push in. Adds tension without motion blur. Keep the modifier explicit: very slow.
- Dolly out. Good for reveals and endings.
- Tracking shot. Specify side tracking or following behind, plus the height: knee, chest, eye, drone.
- Handheld. Specify intensity, since handheld can mean a subtle breath or full documentary shake.
- Crane or tilt up. Strong for final beats, overused in openings.
- Rack focus. Powerful but fragile; name the foreground and background subjects explicitly.
- Orbit. Useful for product hero shots; keep the arc under 180 degrees for reliability.
Lens language is just as useful. Wide lenses around 18–24mm exaggerate space and speed. Normal lenses from 35–50mm read as neutral and human. Longer lenses from 85mm up compress space and flatter faces. Adding one focal length and one depth-of-field note does more for perceived production value than any stack of style adjectives.
The discipline that matters most: one move, one speed, one height per clip. If you want a push in and a tilt during the same five seconds, you are asking for a camera operator who does not exist in the model's world.
Treat Audio as a Second Script
Native audio generation is the feature people underuse most and misdirect most. Treat sound as a second script written alongside the picture.
Ambience carries realism. Rain on metal, room tone in a small kitchen, wind at a specific intensity. One ambience layer per shot is enough; stacking three makes the mix muddy and hard to cut.
Foley sells action. Footsteps on gravel, a zipper, a ceramic mug set down on stone. If the action matters visually, name the sound it makes.
Dialogue needs discipline. Keep lines under about ten words, avoid overlapping speakers, and describe voice quality: low and unhurried, close to the microphone, slight regional accent. If a line must be exact for brand reasons, generate the shot for timing and mouth movement, then record the voice separately. That hybrid almost always beats forcing a single generation to be perfect.
Music is usually better added later. Generated scores tend toward the generic, and you cannot cut them to a beat. Generate picture and sound design, then place a licensed or composed track in the edit where you control the timing.
A practical habit: fill in the audio column of the shot sheet before writing any prompts. It prevents the familiar disappointment of a visually strong clip carrying sound from a different film.
Continuity Across Shots
Continuity is where amateur AI videos visibly fall apart. A jacket shifts shade, a street changes city, a face changes age. Fix it with repetition rather than hope.
Lock a descriptor string
Write one exact phrase for every recurring element and paste it verbatim into each relevant prompt: a woman in her thirties with a short dark bob, olive rain jacket, silver hoop earrings. Small wording changes cause visible drift, so resist the urge to improve the sentence mid-project.
Build a character sheet
Prepare or generate a clean reference image: front view, neutral light, simple background. Most current tools accept it as a character anchor, and it stabilizes faces far more reliably than text alone. A reference image generator helps here because you can iterate on the face before any video attempt is spent.
Chain shots with end frames
Where the tool supports it, use the final frame of shot one as the opening frame of shot two. This is the closest thing AI video has to shooting coverage, and it removes the visual jump audiences notice immediately.
Keep locations boring
A location with one strong feature — a red door, a single window, a specific staircase — reproduces better than a rich, busy set. Complexity invites inconsistency, and inconsistency costs more time than a plainer frame costs beauty.
Worked Example: A Twenty-Second Product Teaser
Suppose you are launching a compact espresso machine: six shots, twenty seconds, vertical format, no dialogue.
- Establish (2s). Extreme close-up of steam against a dark tiled wall. Audio: soft hiss, kitchen room tone.
- Reveal (3s). Slow dolly back from the machine on a counter, morning light from frame left, 50mm. Audio: gentle clink of a cup.
- Process (4s). Macro shot of espresso filling a glass, crema swirling, static camera, shallow focus. Audio: liquid pour, low machine hum.
- Human beat (3s). A hand lifts the cup, thumb on the handle, sleeve visible. Audio: fabric, ceramic on saucer.
- Detail (3s). Rack focus from the metal badge to the cup in the background. Audio: near silence, letting the cut breathe.
- Resolve (5s). Wide of the kitchen, machine on the counter, faint steam, no people, static camera. Audio: room tone and a single music note.
Now the crucial part: generate shot three first. It is the shot most likely to fail, and if the macro pour looks wrong, the whole concept changes. Fail on the hardest shot, not the easiest.
This example also shows why the shot sheet is not bureaucracy. Six rows, six decisions, and the edit almost assembles itself. Compare that to the alternative: forty unplanned clips, three that look nice, and no idea what order they belong in.
Iteration Budgets and Cheap Failure
Generation is fast, so the temptation is to keep re-rolling until something miraculous appears. That is a slot machine, not a workflow. Use a loop with stop rules instead.
- Draft at lower fidelity first. Confirm composition and action before polishing. A cheap, ugly version that proves the blocking is worth more than a beautiful version of the wrong shot.
- Generate two variations per shot, not ten. Compare both against the shot sheet, not against each other.
- Change one variable at a time. Camera, action, or audio — never all three, or you will not know what fixed it.
- Set a three-attempt rule. If a shot has not landed after three honest attempts, the shot is wrong, not the prompt. Simplify the action, change the framing, or cut the shot.
- Log what worked. A short text file of winning prompts saves hours next project, because strengths carry across subjects.
- Plan volume before you start. Estimate two to four attempts per finished shot while learning, dropping to one or two once your descriptor strings are stable. Then add the edit, which is usually a third of the total effort.
If you are still learning how a model reads camera and light language, starting from a curated prompt library shortens the curve considerably, because you can see the structure behind prompts that already work.
Editing Discipline and Common Mistakes
A folder of good clips is not a video. Pacing, meaning, and polish happen in the edit.
Cut on action. End clips mid-movement and start the next mid-movement. This hides the seams that single-shot generation cannot avoid.
Vary shot length deliberately. Two-second cuts read as energy; five-second holds read as confidence. A teaser usually alternates between them.
Use sound bridges. Let the ambience of the next scene begin before the cut. It makes disconnected generations feel authored.
Grade for consistency. One subtle contrast and color adjustment applied across every clip does more for cohesion than any prompt tweak.
Add text and captions in post. Never rely on a model to render readable typography.
Do a music-first pass. Lay the track, then trim clips to its beats. It is faster than cutting picture and hoping music fits afterwards.
Now the mistakes worth naming out loud:
- Overloaded prompts. More than one primary action or camera move per clip invites failure.
- Ignoring audio direction. Silent prompts produce generic sound beds that clash later in the timeline.
- Chasing resolution first. Composition errors survive upscaling. Fix them early and cheaply.
- Describing mood instead of light. Direction, hardness, and color temperature are usable instructions; atmosphere words are not.
- No end frame. Starting each new shot from nothing almost guarantees a tonal jump.
- Skipping the shot sheet. Improvisation produces pretty clips with no argument.
- Aspect ratio mismatch. Decide vertical or widescreen before generating; crop reframing is miserable.
- Hero-shot syndrome. Spending all remaining effort on the opening and rushing the ending, which is the shot people remember.
When you compare tools, judge them on audio in the same pass, reference image support, usable clip length, camera control fidelity, predictability of cost, and how clean the output is on a timeline. It helps to test the same three-shot sequence everywhere instead of judging showreels; a shortlist of AI video generator alternatives can speed that up, but your own test clip is the real decision.
FAQ
How long should each AI-generated clip be?
Aim for three to five seconds per story beat. Shorter clips are easier to control and faster to cut; longer clips invite drift in faces and backgrounds. Save longer generations for static establishing shots where nothing needs to move except light.
Can I keep a character consistent across shots?
Yes, with discipline: one fixed descriptor string, a clean reference image, and end-frame chaining between clips. Expect to spend real time getting the first shot of a character right, then reuse that exact setup everywhere else. Drift usually comes from rewriting a description, not from the model.
Should I generate dialogue or record it separately?
Generate it for scratch timing and natural mouth movement, then replace important lines with recorded audio in the edit. Brand-critical copy should never depend on a single generation, and re-recording is faster than re-rolling a scene ten times.
Why does my camera movement come out chaotic?
You are probably asking for more than one move, or using abstract words like dynamic. Name one move, one speed, and one height, then delete everything else. Static shots are also a legitimate answer and often look more expensive.
Do I still need editing software?
Yes. Any editor works, including free ones. Generation replaces shooting, not cutting. Pacing, sound bridges, captions, and grade are what make the result feel professional rather than generated.
How many generations should a thirty-second video take?
Plan on three to five attempts per finished shot while you are learning, dropping to one or two once your prompt structure and descriptor strings are stable. Budget your time for the edit as well; it is usually a third of the project and the part that decides quality.
What if a shot keeps failing?
Change the shot, not the wording. Reduce the action to a single verb, switch to a static camera, or replace the shot entirely with a detail that carries the same story job. A shot sheet is a plan, not a contract.
Build your next scene with Orelon. A shot sheet, one clear idea per clip, camera language you can defend, and audio written before prompts — that is the whole method. Orelon is an AI video generator built for cinematic ideas in motion: use the video generator to turn a shot-by-shot plan into moving clips, the image tools for character and location references, and video templates when you want a starting structure instead of a blank canvas. Write your five-shot sheet first, then bring it to Orelon and watch the plan move.