Orelon logoOrelon
Tarifs

Prompts for AI Video Generators: A Practical Guide

11 nov. 2026 · Par Orelon Team

Explorez les modèles vidéo IA

Parcourez quelques créations de la communauté pour trouver l’inspiration, puis ouvrez n’importe quel modèle pour continuer à créer dans Orelon.

Learn how to write prompts for AI video generators: structure, camera language, lighting, iteration loops, and fixes for the most common consistency problems.

A video model has no idea what you meant. It only knows what you wrote. Two people can ask for "a cinematic shot of a city at night" and get wildly different results, not because the model changed, but because one prompt gave it something to direct and the other gave it a mood. The prompt is the entire interface between the sentence in your head and the pixels on the timeline, and the gap between a frustrating generation and a usable shot is usually three or four very specific words.

This guide is about closing that gap. You will get the structural pieces that actually change output, a formula you can reuse across genres, worked examples, an iteration loop that converges instead of wandering, and the reference-frame techniques that make multi-shot sequences hold together. If you want to test ideas as you read, open the AI video generator in one tab and keep a scratch file of prompts in another.

Why prompt quality decides output quality

Every generation is a compression problem. Your text gets compressed into a latent representation, then decompressed into motion. Anything you leave vague is filled in by the model's average of everything it has ever seen, which is exactly why vague prompts look generic. "Beautiful woman walking" does not produce a bad image; it produces an average one, assembled from thousands of unrelated sources.

Specificity works because it narrows the distribution. Saying "a woman in a rust-colored wool coat walking through wet slate streets, seen from behind at waist height" removes the model's freedom to choose wardrobe, weather, framing, and camera position. It still has freedom over how she moves and how the light falls, which is the part you actually want the model to improvise on.

There is a second reason prompts matter more in video than in stills: time. A still image only has to be coherent for one frame. A clip has to stay coherent across dozens or hundreds of frames, and every ambiguous word becomes an opportunity for the model to change its mind halfway through. Ambiguity in a prompt does not just soften the look, it creates drift.

The anatomy of a strong video prompt

Most strong prompts contain the same six or seven ingredients in a different order. You do not need all of them every time, but knowing what each one does lets you add or remove them deliberately.

Subject and action

Name who or what is on screen and what they are doing, using one clear action. "A mechanic wipes grease from her hands" beats "a mechanic doing mechanic things." If the action is complex, split it into two shots instead of asking a single clip to perform a sequence.

Setting and time of day

Location and hour determine the light more than any other single detail. "An empty laundromat at 2 a.m." carries a complete lighting plan inside it. If the setting matters for continuity, describe the same three landmark objects every time so the model has anchors.

Camera position, lens, and movement

This is the most underused part of most prompts. Say where the camera is (low angle, shoulder height, overhead), what it is doing (locked off, slow push in, handheld follow, crane up), and what kind of lens it implies (wide 24mm, normal 50mm, long 85mm with compressed background). A locked-off wide shot and a handheld close-up of the same scene read as two different films.

Light and color

Describe the source, the quality, and the contrast. "Single practical lamp, hard shadow, warm falloff into blue ambient" gives the model a lighting diagram. "Cinematic lighting" gives it a catalog of clichés.

Motion and pacing

Video prompts need a tempo. Words like drifting, snapping, settling, gusting, or gliding tell the model how fast things should move. Without them, the model defaults to a medium, slightly floaty speed that reads as artificial.

Style and texture

Film stock, grain, rendering style, and reference era belong here. Keep it to two or three descriptors. Stacking eight style words creates internal contradictions, and contradictions resolve into mush.

Negative instructions

State what must not appear: no text overlays, no lens flare, no extra fingers, no crowd. Keep the list short and concrete. Long lists of prohibitions tend to pull attention toward the very thing you are excluding.

A reusable prompt formula

Here is a skeleton that works across most models. Fill the brackets, then delete whatever the shot does not need.

[subject + single action], [setting + time of day],
[camera position] + [lens feel] + [camera movement],
[light source + quality + color],
[motion tempo of the subject or environment],
[style: 2-3 texture or era descriptors],
[aspect ratio / duration note]

A filled version:

A ceramicist presses her thumb into a spinning bowl, clay slurry on her wrists,
bare studio at dawn, camera at table height on a 50mm, slow dolly right,
low window light from the left, soft with a hard rim on the wheel,
the wheel spins fast, her hands move slowly and deliberately,
muted earth tones, fine 16mm grain, no text overlays

Notice what is absent: no adjectives about beauty, no vague praise, no request for a story. The prompt describes a physical situation and lets the emotional read emerge from the light and the pace.

Worked examples across genres

Different genres lean on different parts of the formula. These four examples show where the weight goes.

Product film

Weigh the prompt toward surface, light, and camera precision. "Macro shot of a matte black watch case rotating on a black glass surface, camera locked off at 30 degrees, single softbox from behind creating a thin edge highlight, dust motes drifting in the beam, no reflections of the studio, 3 seconds." Product prompts fail when they ask for a rotation and a reveal and a logo in the same clip. Pick one movement.

Narrative scene

Weigh it toward performance and blocking. "Two brothers stand on either side of a kitchen doorway at night, neither speaking, camera at chest height slightly behind the older one, shallow focus on the younger one's face, overhead bulb flickering once, no camera movement, both hold still except for a small shift in weight." Restrained motion is a feature here, so say so explicitly.

Landscape or establishing shot

Weigh it toward atmosphere and scale. "Aerial push over a flooded river delta at golden hour, camera 40 meters up moving forward at walking pace, water channels catching a warm specular sheen against dark green reed beds, mist sitting in the low ground, long lens compression, no birds, no boats." Scale comes from camera height plus a slow speed; fast drone moves make big landscapes look like drone stock.

Abstract or loop

Weigh it toward motion physics and texture. "Ink plumes blooming in clear water, filmed at 120 frames per second against a white background, single diffused top light, slow symmetrical expansion, shallow depth of field, high shutter so edges stay crisp, seamless loop." For loops, describe the start and end states as identical, otherwise the cut will show.

Iteration: how to refine without wandering

Prompt work is a search, and searches go badly when you change several variables at once. Treat each generation as a test with one variable.

Change one thing per pass

If the framing is right and the light is wrong, edit only the lighting clause. If you rewrite the whole prompt between attempts, you lose the ability to know which phrase did the work. Keep a fixed core and rotate a single slot.

Freeze a winning fragment

When a phrase reliably produces the look you want, stop rewriting it. Build a small library of proven fragments: a lighting clause, a camera clause, a texture clause. Mixing known-good fragments is faster than writing new prose every time.

Read the output as feedback about the prompt

The model is not failing randomly. If every generation drifts into soft focus, you probably implied shallow depth of field. If every character turns to face the camera, your action clause was probably too passive. Diagnose the prompt instead of adding more adjectives as compensation.

Cap your attempts per shot

Set a ceiling, often five to eight generations, then either accept the best result or change approach entirely: a different camera angle, a still image start frame, or a simpler action. Endless micro-edits on a fundamentally wrong approach wastes more time than starting over.

Consistency across shots: reference frames and anchors

Single clips are easy compared to sequences, where the same character, wardrobe, and location must survive multiple generations. Text alone rarely holds this together.

Use an image start frame

Generate or supply a still frame that establishes the scene, then animate from it. The model inherits composition, wardrobe, and light from the image, which removes most of the drift. The AI image generator is useful for building those frames before you spend time on motion.

Build a character sheet

Create three or four canonical images of a character: front, three-quarter, and a full-body shot in the actual wardrobe. Reuse the same description text alongside those images in every shot prompt, down to the same adjectives. Consistency comes from repetition, not from a single perfectly written paragraph.

Write a continuity block

For each project, keep a short block of text that never changes: location landmarks, wardrobe hex-adjacent color words, time of day, lens family, grain level. Paste it into every prompt. It is boring, and it is the single biggest difference between sequences that feel like one film and sequences that feel like a mood board.

Match motion, not just appearance

Two shots can have identical characters and still feel disconnected if one moves fast and one drifts. Give the whole sequence a motion rule, something like "all camera moves slower than a walking pace," and apply it consistently.

Shot lists before prompts

A prompt is a shot, so decide the shots first. Write a plain-text shot list with one line per clip: shot number, framing, action, duration, and the one thing the audience must notice. Only then write prompts.

This order matters because prompt writing tempts you into writing five beautiful unrelated clips. A shot list forces you to answer questions the model cannot: which shot establishes, which one reverses, which one is the reaction. A ten-shot sequence with a coherent list beats thirty gorgeous orphan clips every time.

A practical template for each line:

  • Shot 04 — medium, over the shoulder, she reads the letter, 3s, the audience must notice her hands shaking.

Then the prompt writes itself outward from that intention: framing, action, light, movement, style, in that order.

Common mistakes that flatten results

Stacking style words. Five references to different directors produce an average of five incompatible looks. Pick two.

Contradictory light. "Golden hour" and "harsh overhead noon sun" cannot coexist. Contradictions get silently dropped, and you never learn which one went missing.

No camera instruction. If you do not specify position or movement, you accept the model's default, which is usually a slow, centered, slightly floating medium shot.

Asking for a story. One clip cannot contain a beginning, middle, and end. Ask for a moment; arrange the moments in the edit.

Ignoring aspect ratio and duration. A vertical 5-second prompt behaves differently from a widescreen 10-second one. State the frame and the length when the model supports it.

Legible text requests. Logos, signage, and on-screen type remain unreliable. Generate the plate and add type in post.

Building a prompt library that compounds

Keep a single document with four columns: the prompt, the model and settings used, a one-line verdict, and the reusable fragment you extracted. After a few weeks you will have a personal vocabulary of clauses that reliably do what you want, which is worth more than any generic prompt list. Public collections such as the prompt library are useful for range, but your own log is what makes you fast.

If you want a head start on structure, video templates give you a working shape to modify rather than a blank box. Treat them as scaffolding: swap the subject and the light, keep the camera grammar.

FAQ

How long should a video prompt be? Long enough to remove ambiguity, short enough to stay internally consistent. Most reliable prompts land between 30 and 70 words. If you pass 100 words, check for contradictions before adding anything else.

Does punctuation or formatting matter? Commas and line breaks mainly help you stay organized. Some models weight early tokens more heavily, so put the subject and camera first if the output feels unfocused.

Why does my character's face change between shots? Text descriptions rarely pin facial features precisely. Use a reference image per shot, keep the wardrobe description word-for-word identical, and avoid describing the face differently each time.

Can I reuse the same prompt across different models? Partly. The subject, light, and action clauses transfer well; camera and style phrasing often do not, because each model has its own vocabulary for motion and lens behavior. Expect to re-tune the camera clause.

Do negative prompts actually work? They help with a small number of concrete exclusions, like text overlays or extra limbs. Long lists of exclusions are unreliable, and sometimes mentioning an object pulls it into frame.

What is the fastest way to improve? Rewrite one prompt three ways, changing only the camera clause, and compare. Camera language is the highest-leverage variable and the one most people skip.

Direct your next clip with a better prompt

Prompts are not incantations, they are briefs. Describe the situation, place the camera, name the light, set the tempo, and let the model handle the rest. Then iterate one variable at a time until the shot behaves.

Orelon is built for exactly this loop: cinematic ideas in motion, with room to test a camera angle, lock a reference frame, and build a sequence shot by shot. Start with the AI video generator, pull a starting structure from the prompt library, and keep a log of what works. Your tenth prompt will be better than your first, and your prompt file will be the most valuable asset in the project.