Learn how to turn an idea into a shot-by-shot AI video prompt: camera language, lighting, motion, dialogue, iteration loops, and fixes for common failures.
Most disappointing AI video clips are not model failures. They are translation failures. A director holds a clear picture in their head, types three hurried sentences into a generator, and receives something technically moving but emotionally empty. The model did exactly what it was told; the problem is that what it was told was not what the director meant.
Prompt writing for modern video models — including Google's Veo family, Kling, Runway, Seedance, and Orelon's own generation stack — is closer to writing a shot list than to writing a search query. You are not describing a topic. You are specifying a moment: who is on screen, what changes between the first frame and the last, where the camera sits, how light behaves, and what the viewer should feel. This guide walks through a repeatable method for turning a rough idea into a prompt a cinematic model can actually follow.
A prompt is a script, not a search query
Search queries retrieve. Prompts direct. When you type "cyberpunk city at night" into a generator, you are asking it to retrieve a vibe. When you type a script, you are asking it to perform a scene.
The distinction matters because video models are temporal. They must decide what happens in second one, second three, and second seven. If your prompt only describes a static tableau, the model invents motion on its own — and it usually invents something generic: a slow push-in, drifting particles, a subject blinking at nothing.
A useful mental test: could a camera operator and an actor shoot your prompt without asking you a single question? If the answer is no, the prompt is still a mood board.
Three habits separate strong prompts from weak ones:
- Write in beats, not adjectives. "She turns from the window, sees the empty chair, and stops" gives the model a sequence. "Melancholic and beautiful" gives it nothing to animate.
- Commit to one idea per clip. Generators handle one clear action far better than three competing ones. Save the rest for the next shot.
- Describe what the camera sees, not what you know. The model cannot render backstory, only visible evidence of it.
The six layers of a cinematic prompt
Every strong prompt I have seen resolves into the same six layers. You do not need all six in every clip, but you should consciously decide which ones to include. Skipping a layer is a choice; forgetting one is an accident.
Subject and casting
Who or what is on screen, described in visual terms. Not "a businesswoman" but "a woman in her late forties, close-cropped grey hair, charcoal blazer, sleeves pushed to the elbow." Specificity in casting language reduces the model's tendency to default to a generic, symmetrical face.
For objects, describe material and wear: "a chipped enamel mug," "a rusted steel door with peeling green paint." Material cues are among the strongest levers for realism.
Action beats and timing
What changes during the clip. Use two or three beats maximum: an opening state, a turn, and an ending state. "He lifts the lid, steam escapes, his expression drops." That is a complete three-second story.
Beats also control pacing. "Slowly" and "in one continuous motion" tell the model to stretch the action across the full duration instead of finishing in the first second and idling.
Camera language
Lens, height, and movement. "35mm, eye level, slow dolly left" is more useful than "cinematic camera." Include the shot size — wide, medium, close-up, macro — because it determines how much the model must invent.
If you want stillness, say so explicitly. "Locked-off tripod shot, no camera movement" prevents the default drift that ruins dialogue scenes.
Light and color
Light is the fastest way to make a clip look intentional. Name the source and the direction: "single practical lamp from camera left, deep shadow on the right side of the face, warm 3200K." Then name the palette: "muted teal shadows, skin tones kept natural."
Avoid stacking contradictory lighting. "Harsh midday sun" and "soft candlelit glow" in the same prompt produce mud.
Style and texture
Film stock, grain, era, and rendering intent. "16mm grain, slight halation, documentary handheld feel" reads very differently from "clean digital, high dynamic range, commercial polish." Pick one register and stay in it for the whole sequence.
Continuity anchors
The details that must survive between shots: wardrobe, hair, props, time of day, weather, screen direction. Repeating these anchors verbatim across prompts is the cheapest way to make a multi-shot sequence feel like one film. More on this in the series section below.
A worked example: one sentence to a shot list
Start with a vague creative intention and refine it in passes. Here is the same idea at four levels of resolution.
| Level | Prompt | What the model will likely do |
|---|---|---|
| Raw | "A baker at dawn" | Generic sunrise, generic hands, drifting camera |
| Structured | "A baker in a small kitchen at dawn, kneading dough, warm window light, medium shot" | Recognizable scene, but flat motion and no arc |
| Cinematic | "Medium close-up, 50mm, a flour-dusted baker in her sixties kneads dough on a scarred wooden counter. Grey dawn light through a fogged window behind her, one warm bulb overhead. Slow handheld drift right, shallow depth of field, 16mm grain." | A coherent, textured shot with intentional framing |
| Scripted | "Medium close-up, 50mm, locked-off with a slow handheld drift right. A flour-dusted baker in her sixties kneads dough on a scarred wooden counter, pressing hard, then pauses and looks toward the fogged window. Grey dawn light from behind her, one warm bulb overhead, flour suspended in the air. Shallow depth of field, 16mm grain, muted palette, no dialogue." | A shot with a beginning, a turn, and an ending — the model has a beat to build toward |
The jump from level three to level four is where most creators stall. It is not more adjectives; it is a single, small piece of behaviour. The pause. The look toward the window. That is what makes a clip feel directed.
Dialogue, sound, and motion in the same prompt
Models that generate synchronized audio change the calculus, but they do not remove the need for discipline.
If a line of dialogue matters, quote it exactly and keep it short — under roughly twelve words per clip. State who speaks and how: "She says, quietly, 'You should have called.'" Then specify what the face does while speaking, because models often let performance collapse into a neutral mouth movement. "Her jaw tightens before she speaks" gives the generation something to work with.
For sound design, separate ambience from events. Ambience is continuous ("distant traffic, low room hum"). Events are punctual ("a door closes off-screen at the halfway point"). Mixing them into one phrase usually produces neither.
Motion deserves its own sentence. Distinguish between:
- Subject motion — what the person or object does.
- Camera motion — how the frame travels.
- Environmental motion — wind, water, steam, crowds, flickering light.
Specify at least subject and camera. Left unspecified, environmental motion tends to fill the vacuum with floating dust and slow-motion hair.
Iteration: change one variable at a time
Prompt crafting is an experimental discipline, and most people run it badly. They rewrite five things at once, get a better clip, and learn nothing they can reuse.
Adopt single-variable testing instead:
- Freeze a baseline. Write a prompt you consider 70% correct and generate it as-is.
- Change one layer. Only the lighting, or only the camera move, or only one action beat.
- Judge against a fixed rubric. Composition, motion realism, subject consistency, lighting intent, and whether the beat landed. Score each out of five.
- Keep a prompt log. One line per generation: what you changed and what happened. After twenty clips you will have a personal rulebook that no generic guide can give you.
Two practical supports make this faster. Seed or reference locking, where the tool exposes it, isolates style changes from composition changes. And a saved prompt library of proven blocks — your best lighting sentence, your best camera sentence — lets you assemble new prompts from tested parts rather than starting from a blank field.
Finally, resist the urge to fix a bad clip by adding clauses. Long prompts accumulate contradictions. When a generation fails twice, the fix is usually to delete a sentence, not add one.
Common failure modes and how to fix them
| Symptom | Likely cause | Fix |
|---|---|---|
| Clip looks flat and static | No action beat or camera instruction | Add one behaviour and one camera move |
| Subject morphs mid-clip | Too many competing actions | Reduce to two beats, shorten duration |
| Everything looks glossy and fake | No texture or light-source language | Name the light source, add grain or imperfection |
| Cuts feel unrelated across shots | Missing continuity anchors | Repeat wardrobe, palette, and time-of-day lines verbatim |
| Dialogue arrives out of sync | Line too long or performance unspecified | Trim to one short line, describe the face |
| Motion looks like slow motion | Vague pacing words | Use "in real time," "quick, decisive" |
| Composition drifts every take | Framing described in mood terms | Specify shot size, lens, and camera height |
One more failure deserves its own line: the prompt that describes the ending before the beginning. Models read prompts roughly in order and often front-load what they read first. Put your opening state first, the turn second, the resolution last. Sequence in the prompt becomes sequence on screen.
Adapting the same idea across different models
No two generators interpret identical text the same way, and pretending otherwise wastes time. Treat each model as a collaborator with habits.
Some models, Veo 3 among them, respond well to natural, director-style prose with embedded audio and dialogue instructions. Others, particularly those tuned on shorter captions, reward compressed, comma-separated descriptors and punish long sentences. Motion-heavy engines favour explicit physics — weight, inertia, contact — while stylised engines respond better to reference imagery and texture words.
A workable process: write your master prompt in full cinematic prose, then create two derivatives. A compressed variant (subject, action, camera, light — nothing else) for caption-style models, and a reference-led variant where you pair the prose with a still image or style frame. If you are comparing tools, side-by-side notes such as those in this Kling AI comparison help you judge which engine suits which shot type rather than which is "best" in the abstract.
Keep a small matrix: shot type against model, with a score for motion, realism, and prompt obedience. It takes an afternoon and saves months.
Building a reusable prompt system for a series
Single clips are a demo. Series are a craft. Once you plan more than three shots, move from prompts to a system.
Write a style bible first. Five lines: palette, light philosophy, lens set, grain and texture, and what the film never does. Paste the relevant lines into every prompt.
Lock a character block. A verbatim paragraph describing your subject — age, build, hair, wardrobe, distinguishing detail — repeated in every shot that features them. Change only the action and framing around it.
Number your shots and map screen direction. If a character exits frame right, they should enter frame left in the next shot. State this in the prompt; models do not infer geography.
Use templates for repeatable formats. Product reveals, talking-head explainers, and title sequences all have stable structures. Starting from a proven video template is faster than rebuilding structure each time, and it keeps your visual grammar consistent across a campaign.
Batch by location, not by story order. Generating all shots set in the kitchen together keeps lighting and wardrobe drift under control, because you are refining the same conditions rather than jumping between them.
FAQ
How long should an AI video prompt be? Long enough to specify six layers, short enough to avoid contradictions. Most well-crafted prompts land between 40 and 90 words. Beyond roughly 120 words, additions usually start fighting each other.
Should I describe the ending of the clip? Yes — describe the final state, because that is what gives the model a target to animate toward. Just keep the order chronological: opening state, turn, final state.
Do I need camera language if I want a simple shot? Yes, but simple camera language. "Locked-off medium shot" is a complete instruction and prevents unwanted drift.
How many generations should I expect per usable clip? Plan on three to six for a considered shot. If you are routinely needing fifteen, the prompt is doing too many jobs and should be split into two clips.
Can I reuse prompts across projects? Reuse blocks, not whole prompts. Lighting sentences and camera sentences travel well; character and wardrobe blocks rarely do.
What is the fastest way to improve? Keep a written log and score each generation against the same five criteria. Deliberate practice with feedback beats watching tutorials.
Should I include the word "cinematic"? It is a weak signal on its own. Replace it with the specifics that make something look cinematic — lens, light source, grain, camera height, and restraint in motion.
Turn your next prompt into a shot
Good prompting is a translation skill: you are converting intention into instructions a camera and an actor could follow. Nail the six layers, write in beats, test one variable at a time, and build a style bible the moment you move past a single clip. The results compound quickly.
Orelon is built for exactly this loop — cinematic ideas in motion, with room to test prompts, compare takes, and assemble sequences. Start in the AI video generator, generate a style frame in the image generator when you need a reference, and keep your best blocks in the Orelon blog workflow notes for the next project. The prompt is the script. Write it like one.

