AI Video Prompt Writing: A Practical Cinematic Guide

A prompt is not a wish. It is a compressed shot list written for a system that has never seen your storyboard and will never ask a clarifying question. When an AI video model returns something generic, it is usually doing exactly what it was told. The gap sits between the shot you pictured and the sentence you typed.

Text-to-video systems make hundreds of small production decisions for every second of footage: where the camera stands, what the light does, how fabric moves, whether the air has any weight. Every one of those decisions is inferred from your words. Leave a slot empty and the model fills it with the statistical average of everything it has learned, which is why so much generated footage settles into the same soft-focus, mid-distance, gentle push-in look.

This guide closes that gap. It covers how a video model reads a prompt, how to structure one that survives iteration, which camera and lighting vocabulary actually changes the output, how to keep people and products consistent across a sequence, and how to run a generation loop that ends with a cut instead of a folder of near-misses.

Start with intent, then turn it into a shot spec

Good prompts are written twice. First in plain language, with no jargon at all, so you know what the shot is for. Then in structured language, so the model knows what to render.

The first draft might be: a cyclist waits at a rain-soaked crossing, then decides not to cross and pushes off again. That sentence already contains the two things weak prompts lack, a subject with a situation and a change of state. It is worth writing down before you touch a single camera term, because once you start typing lens numbers it is easy to forget what the shot was about.

The second draft turns that intent into a specification:

Medium shot of a cyclist in a soaked yellow rain jacket paused at a crossing, hand on the handlebar, then pushing off and turning away, rain-soaked street at dusk, single cold streetlight overhead with warm shop-window spill on the left, 35mm lens at chest height, static camera with slight handheld sway, rain visible in the light beam, muted blue and sodium palette, fine film grain, no on-screen text.

Same idea, now renderable. Nothing in the second version is poetic; everything is a decision.

Give every shot a word budget

Forty to eighty words per shot is the sweet spot. Under thirty words, too many slots stay empty and the model improvises. Over about 120, the middle of your prompt starts getting dropped, because models weight the beginning heavily and the end moderately, and the sentence you were proudest of is usually the one that disappears.

If your idea genuinely needs more than 120 words of description, it needs two shots. A five-second clip comfortably holds one or two beats. Three actions in three seconds smears into mush.

Replace mood words with the thing that creates the mood

What you mean	What to write instead
Make it cinematic	35mm lens, one slow push-in, warm practical behind the subject
Make it epic	Low camera near the ground, 18mm lens, subject small against a cliff
Make it moody	Single hard source from camera left, deep shadow falloff, cool grade
Make it premium	85mm close-up, soft window light from the right, condensation detail, no hands, no text
Make it dynamic	One tracking move at walking pace, subject crossing frame left to right

Every row on the right is checkable. You can watch the output and say whether it happened. That is the test for any prompt phrase: if you cannot verify it by watching the clip, it is decoration.

The five slots every video prompt fills

A text-to-video model converts your sentence into a short list of production decisions. Those are slots. Your prompt either fills them or leaves them to chance.

1. Subject and condition

A man gives the model an average man. A weathered fisherman in an oil-stained yellow slicker gives it a specific one. Wardrobe, material, age, and wear are cheap words that buy enormous specificity. For products, replace adjectives with physical facts: brushed aluminium, condensation beading on the lid, a chipped enamel rim, a hairline scratch across the screen.

2. Action with a visible change

A clip needs a change of state, not a pose. A woman stands in a kitchen produces a still image with drift. A woman sets down a heavy pan, exhales, and wipes her hands on her apron produces three beats the model can distribute across the duration. Count your beats and match them to your clip length.

3. Place, time, and weather

Location sets the physics of the shot: what reflects, what moves, what color the shadows are. Dusk on a wet street gives you sodium light and reflections. A dust-lit interior gives you beams and floating particles. Name the time of day and the weather even when it seems obvious, because the model will otherwise pick a neutral, well-lit afternoon.

4. Camera placement

Where the camera is, at what height, with what lens, and how it moves. This is the slot most writers under-specify and the one audiences notice most. A close-up at ground level and a wide shot from a crane can hold the same action and feel like different films.

5. Style, grade, and constraints

Style tags work best as a short cluster at the end: documentary realism, 16mm grain, muted grade. Keep it to three or four. Ten adjectives do not add up to a stronger look; they average out into a weaker one.

Constraints are the most underused part of a prompt. No on-screen text, no logos, single subject, no camera shake. Each one removes an entire failure mode before it happens. If a previous generation produced a specific artifact, name it in the negative and move on.

A fill-in template

[subject + wardrobe or condition]
[action with a visible change]
[place, time of day, weather]
[light source, direction, quality]
[shot size + lens]
[one camera movement]
[atmosphere: particles, haze, reflections]
[style and grade]
[constraints]

Fill this in and you will rarely write a bad first draft. The order matters: subject and action early because they carry the most weight, style and constraints late because they are modifiers.

Camera and lens language that actually changes output

Most prompt vocabulary is decorative. Camera terms are not, because they map to real geometry the model has learned from footage.

Term	What it produces
Dolly in / push-in	Steady forward travel, background compresses
Tracking shot	Camera moves with a moving subject
Orbit / arc	Camera circles the subject, background parallax
Crane up	Vertical rise that reveals scale
Whip pan	Fast rotation with motion blur
Static tripod	No movement, stable frame
Handheld drift	Small organic sway, documentary feel
Rack focus	Attention shifts between two planes
Macro	Extreme close-up, tiny depth of field
Anamorphic	Wider frame, oval bokeh, horizontal flares

Pair every movement with a speed and a reason. Slow push-in at walking pace toward the subject's hands is usable. Dynamic camera is not a direction; it is a shrug.

Lens language is just as literal. Wide lenses around 18 to 24mm exaggerate space and can distort faces near the edges. A 35mm reads as observational. A 50mm is neutral. An 85mm flatters faces and compresses backgrounds. Naming a focal length is often faster and more reliable than describing depth of field in words, and it removes a whole category of ambiguity from the prompt.

One movement per shot. Two movements, a push-in plus an orbit for example, usually resolve into an unintended wobble, because the model averages the two trajectories instead of sequencing them.

Lighting, palette, and atmosphere

If you improve only one part of your prompts, improve this one. Lighting tells the audience what kind of scene they are watching before anything moves.

• Hard single source: drama, tension, visible shadow edges.
• Overcast ambient: documentary, melancholy, even skin tones.
• Practical sources in frame: lamps, screens, neon, headlights. Adds realism and color contrast.
• Backlight with haze: silhouettes, atmosphere, a sense of air.
• Bounced daylight with a window pattern: interiors that feel lived in.

Write light as source plus direction plus quality. A single hard key from camera left. A window behind the subject blowing out softly. Neon spill from a bar sign across wet pavement. An overcast sky that flattens shadows. That is a complete instruction; good lighting is not.

Color follows light. Name a palette rather than a mood: warm sodium against cold blue shadow, bleach-bypass desaturation, sun-bleached pastels, teal and amber. Two colors in tension read as more cinematic than one overall tint, because tension is what the eye reads as intention.

Atmosphere is the cheapest realism you can buy. Sea spray hanging in a light beam, dust motes drifting through a shaft of sun, steam curling off a cup, rain streaking past a streetlight. One atmosphere element per shot is usually enough; three turn the frame into fog.

Consistency across shots: people, products, places

Consistency is the hardest part of AI video, and it is a workflow problem more than a prompt problem. Four habits solve most of it.

Generate the anchor frame first. Create a still of your character or product with an AI image generator, iterate until it is exactly right, then use that frame as the visual reference for every shot in the sequence. A reference image is worth more than a paragraph of description, because the model no longer has to guess.

Freeze the descriptor block. Copy the same subject, wardrobe, and palette sentence word for word into every prompt in the sequence. Only the action, shot size, and camera movement should change between shots. The moment you paraphrase, the jacket becomes a coat and the coat becomes outerwear, identity starts to drift.

Change one variable per shot. If shot two needs a new angle, do not also change the lighting and the wardrobe. When something breaks, you need to know which change broke it.

Keep a sequence bible. A plain text file with the locked descriptor block, the palette, the lens set, and the approved reference frames. Ten minutes of bookkeeping saves hours of regeneration, and it means a collaborator can pick up the sequence without interviewing you first.

Locations need the same treatment. Write one location block with architecture, surface materials, light direction, and time of day, then reuse it. Audiences forgive a lot, but a street that changes color between cuts reads as a mistake rather than a style choice.

Image-to-video: describe motion, not appearance

Starting from a reference image changes the job of the prompt. The model already knows what the subject looks like, so your words should describe motion, camera, and change, not appearance.

Re-describing the subject in different terms is the most common way people accidentally morph a face. If the reference shows a red coat, do not write a burgundy jacket; write that the coat shifts as she turns. If the reference shows a bearded man in his forties, do not add adjectives about age or features. Every new descriptor is a small instruction to redraw.

A clean image-to-video prompt looks like this:

Camera slowly arcs right; her hair lifts in the wind; the coat fabric ripples; the light stays consistent; subject identity unchanged; no camera shake; no text.

Short, motion-first, and explicit about what must be preserved. Add one preservation clause such as identity unchanged, lighting consistent, or same framing, and the model has a constraint to respect rather than a suggestion to consider.

The iteration loop that ends in an edit

Generating video is not a single act. Treat it as a loop with a fixed number of steps so you do not drift into endless rerolling.

1. Write the shot in plain language. One sentence, no jargon. That is your intent.
2. Convert it to the template. Now it is a specification.
3. Generate three or four variants at the fastest setting that still shows real motion, so you judge the idea rather than the polish.
4. Name the single biggest flaw in the best variant. Not every flaw, the biggest one. The camera drifts upward unexpectedly.
5. Change exactly one thing. Add a camera constraint, swap one lighting word, shorten the action by one beat.
6. Lock and log. When a prompt works, save it with the model name, settings, and reference frame. A personal prompt library turns lucky accidents into repeatable craft.
7. Shoot coverage. For each beat, generate a wide, a medium, and a detail. Cutting between them hides imperfections that a single long clip would expose.

Most finished shots take between four and twelve generations. If you are past twenty on the same sentence, the sentence is wrong, not your luck.

Deciding when to rewrite instead of reroll

Use these three tests before you spend another generation.

• Can you point at the failure? If you cannot name what is wrong in one short phrase, the prompt is too vague to fix by tweaking.
• Did two attempts produce the same flaw? That is a contradiction in the prompt or a missing constraint, not a bad seed.
• Is the problem visible at all? Precise hand action, rendered lettering, and complex crowd choreography remain unreliable. If the shot depends on them, plan to solve it in the edit instead of the prompt.

Four prompt recipes worth adapting

Product spot

Close-up of a ceramic coffee cup on a dark oak table, steam rising, camera pushes in slowly, single warm window light from the right with deep shadow falloff, 85mm lens, shallow focus, dust motes in the beam, clean premium finish, no text, no hands.

Adapt it by swapping the atmosphere element, not the structure. Change dust motes for condensation and you have a cold-drink spot with the same bones.

Narrative scene

Medium shot of two people at a bus stop in the rain at night, one turns away as the bus arrives, cold streetlight from above with warm interior glow through the bus windows, 35mm lens, static camera with slight handheld sway, rain visible in the light, muted blue and sodium palette, film grain, no on-screen text.

The change of state is the whole shot. Remove it and you have a photograph.

Abstract or music video

Macro shot of ink dispersing through water, camera orbits slowly, hard backlight turning the ink into silhouette, 100mm macro lens, very shallow depth of field, high-contrast monochrome with a single crimson streak, no subject, no text.

Abstract work rewards one strong light decision more than a stack of style words.

Social ad

Wide shot of a runner on an empty coastal road at sunrise, camera tracks alongside at running pace, low warm sun flaring into the lens, 24mm lens, deep focus, dust and sea mist in the air, energetic grade with warm highlights, negative space on the right for copy, no text.

Notice that the recipe specifies the composition need, meaning copy space, single subject, and no text, before the model has a chance to invent something in the wrong place. That is the difference between a clip you can post and a clip you have to reframe.

Mistakes, criteria, and a review checklist

These are the errors that quietly ruin otherwise good ideas.

• Adjective stacking. Stunning, epic, hyper-realistic, high resolution, masterpiece. No visual information, and it dilutes the words that do carry meaning.
• Contradictions. Handheld plus locked-off. Golden hour plus harsh noon shadows. Shallow focus plus everything sharp. The model resolves these by ignoring one half, usually the half you cared about.
• Describing a scene instead of a shot. A prompt is one camera and one moment. A three-scene story needs three prompts.
• Crowding the duration. Six actions in five seconds produce mush.
• No motion verb. Without a change of state you get a still image with drift.
• Re-describing a reference. The fastest way to lose a face.
• Forgetting constraints. Text, logos, extra hands, and watermarks all have suppressors. Use them.

Five checks before you approve a clip

1. Does the subject match the reference frame?
2. Is the motion physically plausible, or does something slide?
3. Did the camera do the one thing you asked for?
4. Does the lighting stay consistent from the first frame to the last?
5. Would this cut against the shot before and after it?

A beautiful clip that cannot be edited into the sequence is a dead end. When you browse video templates or study examples, watch for editability rather than spectacle. That is the skill that compounds.

FAQ

How long should a video prompt be?

Forty to eighty words for a single shot. Under thirty leaves key decisions to the model; over about 120 starts losing instructions in the middle. If the idea needs more room, split it into two shots.

Does adding the word cinematic help?

Barely. It is a summary word, not an instruction. Replace it with the specific thing you mean: a lens, a light source, a palette, or a camera movement.

How do I stop faces from morphing?

Anchor the face with a reference image, describe motion instead of appearance, keep the descriptor block identical across shots, and remove any conflicting adjectives about age or features.

Can prompts include dialogue or sound cues?

Some models accept short spoken lines and will shape mouth movement or atmosphere accordingly. Keep lines brief and put them in quotes so they are not read as description. Always plan to add real sound in the edit; generated audio is a starting point, not a finished mix.

Why does on-screen text look wrong in generated video?

Rendered lettering is still unreliable at small sizes and in motion. Design around it: specify no on-screen text, then add typography in post where you control the font, kerning, and timing.

How many generations should a finished shot take?

Four to twelve is normal for a shot with camera and performance requirements. Past twenty, rewrite the prompt rather than reroll it.

Should I write differently for image-to-video?

Yes. Appearance is already solved by the reference, so spend your words on movement, camera, and what must stay unchanged.

What if I only have one still image and no time to build a sequence?

Start with the still. Generate a couple of variants with different single camera moves, then cut them together with a title card and a music bed. A three-shot sequence built around one consistent anchor frame usually beats a single long, drifting generation.

Pick one shot you have been struggling with and rebuild it with the template: subject, action, light, lens, one movement, style, constraints. Generate three variants, change exactly one variable, and log what worked. That single loop teaches more than another hour of prompt lists.

When you are ready, generate the shot in Orelon by writing the prompt, locking your reference frame, and iterating toward the cut instead of collecting clips. For more breakdowns of camera language, workflow, and craft, the Orelon blog is the next stop. Your best footage is usually one well-specified sentence away.