Orelon logoOrelon
Tarifs

How to Write Prompts for AI Video Generators: A Practical Guide

12 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: subject, camera, light, motion, texture, iteration loops, and a checklist for consistent cinematic clips.

A strong prompt for an AI video generator is not a poem and not a spec sheet. It is a short, ordered brief that answers the questions a camera operator, a gaffer, and an editor would ask before a take: who is in frame, where they are, what they do, how the camera behaves, where the light comes from, and what the finished image should feel like.

Everything below assumes one thing: the model will always fill silence with an average. If you do not choose the lens, it chooses a safe middle lens. If you do not choose the light, it chooses flat light. Prompting is the craft of removing the gaps that would otherwise be filled with someone else's taste.

This guide walks through the layers of a complete prompt, how much detail each shot really needs, recipes for common deliverables, a diagnostic loop for weak takes, and the habits that keep a series visually consistent. You can follow along in the AI video generator while reading.

Start With the Shot, Not the Sentence

Most people write prompts the way they write captions: left to right, one idea flowing into the next. That produces paragraphs the model has to summarize. Shot thinking is different — you decide what the frame contains before you decide how to say it.

Four decisions come first, and all four fit on a sticky note:

  • Who is on screen, described specifically enough to be cast (age band, wardrobe, posture, energy).
  • Where they are and what time of day it is, because time drives every lighting choice downstream.
  • The single event of the clip — one action with a beginning, a middle, and an end.
  • The emotional temperature: tense, warm, clinical, nostalgic.

If you cannot state those four in one sentence, the prompt is not ready. Two clips, each carrying a clean idea, will almost always beat one clip carrying two.

A quick test: read your planned prompt out loud and imagine the first and last frame. If either frame is fuzzy in your head, the model's version will be fuzzier.

The Six Layers of a Prompt

Prompts fail more often from missing layers than from wrong words. Run these six in this order every time, and skip a layer only when you have consciously decided it does not matter.

Layer 1 — Subject and action

Describe the subject with nouns that carry visual information: "a woman in her late thirties, damp trench coat, hair pinned back" rather than "a person". Then give one continuous action: "she steps off the curb and glances left". Continuous verbs hold motion together. Stacked beats — walking, turning, laughing, opening an umbrella — collapse into a blurry middle, because the model has to average four physical states into a few short seconds.

If the shot needs a second beat, make it a second clip. Editing two clean takes is faster than rescuing one crowded take.

Layer 2 — World and time

Location plus time of day plus weather: "a narrow side street at dusk, wet asphalt, low mist". Time of day is the highest-leverage token in that sentence, because it determines the light, the palette, and the contrast range at once.

Add one environmental detail that reacts to the subject — steam from a vent, dust drifting through a shaft of light, rain striking a puddle. Reactive environments read as footage; static backdrops read as a render.

Layer 3 — Optics and framing

Say where the camera stands and what it looks through: "low-angle medium shot, 35mm, shallow depth of field" or "overhead wide, 24mm, deep focus". Shot size and angle do most of the storytelling work; lens choice does the rest, since wide lenses exaggerate space and motion while long lenses compress backgrounds and isolate faces.

When a take looks generic, the optics line is usually the missing piece. "Medium shot" is a category; "85mm medium close-up with a compressed background" is a decision.

Layer 4 — Light and color

Describe sources and direction, not moods. "Warm sodium streetlights from camera left, cool ambient fill, teal shadows, one practical lamp in the background" gives the model a lighting plan. "Cinematic lighting" gives it a coin flip.

Limit the palette to two or three colors plus a neutral. Beyond that, the frame stops reading as a designed image and starts reading as a color test.

Layer 5 — Motion and tempo

State whether the camera moves and how fast, and state subject speed separately. "Locked-off static camera" is different from "slow push in, roughly a foot over four seconds", which is different again from "handheld follow with slight sway". Unspecified camera motion is the most common reason a short clip drifts, because the model has no reason to hold the frame still.

Pair camera tempo with subject tempo. A slow dolly plus a hurried walk creates tension; both fast creates noise.

Layer 6 — Texture and finish

Grain, halation, anamorphic bokeh, atmospheric haze, slight lens flare, documentary softness. One texture token often separates footage from a synthetic render, and one is usually enough. Two texture references can coexist; five average into mud.

How Much Detail Is Enough?

Length is the wrong metric; density is the right one. A busy prompt on a simple shot adds noise, and a thin prompt on a complex shot invites improvisation.

A workable rule set:

  • 15–30 words for beauty shots, abstract textures, loops, and backgrounds where only one element needs control.
  • 40–80 words for most narrative, product, and social work — enough for subject, setting, camera, and light without stacking style references.
  • 100+ words only when blocking genuinely requires it: multiple subjects, a specific interaction, a precise camera move, and a defined end state.

Whatever the length, keep the order constant. Consistent ordering makes two prompts comparable, and comparable prompts make iteration a measurement instead of a guess. When your tenth prompt is structured like your first, you can see exactly which change moved the image and which change did nothing.

Style Directives That Actually Change the Render

Style words work when they map to visible decisions. Everything else is decoration.

Instead of Write
cinematic anamorphic framing, warm practicals, cool shadows
realistic available light, natural skin texture, light grain, handheld
epic wide establishing shot, low horizon, slow crane up, storm light
premium single large softbox upper right, matte surfaces, crisp specular highlight
retro warm stock emulation, soft contrast, slight halation, muted greens

The pattern: replace judgment with observation. Do not tell the model the result is impressive — describe what makes it look that way.

Film and era references are efficient shortcuts when they name a trait. "Warm stock emulation with lifted blacks" is actionable; "vintage" is not. If you admire a specific look, extract the trait — the grain, the palette, the contrast curve — rather than naming the source and hoping the association lands.

Keep a reusable style block: lens family, lighting logic, palette, texture, aspect ratio. Paste it into every prompt in a sequence and your clips start matching without editing tricks. The prompt library has ready-made blocks if you would rather start from a known-good baseline.

Prompt Recipes for Five Deliverables

Product hero

Lead with shape and material, because edges and reflections are where generation most often breaks. "Matte ceramic bottle centered on a seamless gradient backdrop, slow 90-degree orbit, single large softbox upper right, crisp specular highlight along the left edge, 50mm, deep focus." One light, one move, restrained palette. Commercial realism comes from restraint, not from adding more surface detail.

Vertical social hook

Vertical changes composition, so prompt for it explicitly. "9:16, close on hands opening a cardboard box, warm window light from behind, shallow focus, locked-off camera, sharp silhouette readable in the first frame." Faces need headroom, hands need space to move, and the action must read with sound off.

Narrative beat

Use optics and blocking to imply genre. A thriller beat: "static wide, subject small in frame, hard side light, empty foreground." Intimacy: "85mm, shallow focus, soft window light, slow drift, quiet background." Add one imperfection such as grain or a slight sway so the frame does not look plastic.

Explainer b-roll

Clarity beats drama. "Medium shot of a technician at a workbench, neutral palette, steady slow push in, right third of frame empty with soft background blur." The empty area is deliberate, reserved for captions. Without that instruction, text lands on a face.

Loops and backgrounds

"Seamless loop, slow drifting clouds over a muted gradient sky, minimal contrast change, no camera movement, no subject." Loops need stable composition and no hard shifts in light, so specify stillness instead of hoping for it.

The Iteration Loop: Diagnosing a Weak Take

Treat generations as footage. Keep the takes with good bones, name the specific failure, and fix that failure with one change. Rewriting the entire prompt after every attempt destroys the information you just gained.

What you see Likely cause Fix
Frame drifts or wanders camera motion unspecified add "locked-off static camera" or an explicit move with speed
Subject changes shape mid-clip stacked actions or vague nouns one continuous action, more specific nouns
Flat, gray image no light source described name the source, direction, and quality
Plastic, synthetic look no texture token add one grain or halation reference
Composition ignores the description too many competing elements cut to one subject, one foreground, one background
Unexpected text or people appear negation used describe the desired state instead of what to exclude

Fix the broadest problem first: composition, then light, then texture, then micro-detail. A perfectly textured frame with broken motion is unusable, while a well-composed frame with flat light can often be finished in post.

Also separate exploration from production. Explore with short, low-commitment generations until a look appears; then rebuild that look as a longer, fully specified prompt for the final take. Mixing the two modes is how an afternoon quietly disappears.

Keeping Characters and Locations Consistent

Consistency across clips comes from repeated language, not luck. Three habits do most of the work:

  1. Freeze the subject description. Copy the exact same words for age, wardrobe, hair, and posture in every prompt for that character. Paraphrasing invites a new face.
  2. Anchor with a still. Generate a reference image first, then start motion from it. The AI image generator is useful for locking a look before committing to moving footage.
  3. Lock the environment. Reuse the same location sentence and the same palette across a sequence, and vary only camera position and action.

If a character still shifts, the description is probably doing two jobs at once — describing appearance and mood together. Split it: appearance belongs in the subject line, mood belongs in the light and tempo lines.

Prompt Habits That Hurt

  • Contradictions: "static camera, fast dolly in" or "bright night" forces a compromise you will not like.
  • Negation: "no people, no text" often summons both. Describe the empty state instead.
  • Vague counts: "people walking" produces crowds in flux; "two people walking" holds.
  • Style soup: five aesthetic references average into none. Two at most.
  • Aspect-ratio neglect: generating landscape and cropping to vertical destroys the composition you described.
  • Judging detail first: check motion and framing before you judge texture.
  • Silent revisions: changing three things at once teaches you nothing about any of them.

Pre-Flight Checklist

Before a serious generation, confirm: one subject, one action, one idea · location and time stated · shot size, angle, and lens chosen · one camera move with a speed · light source and direction described · palette limited to two or three colors · a single texture reference · aspect ratio and duration set · no negations or contradictions · prompt order consistent with earlier takes.

Ten seconds of checking saves a dozen wasted generations.

FAQ

How long should a prompt be? Most shots land between 40 and 80 words. Shorter works for simple visuals; longer only when blocking and camera movement both need control.

Do director or film references help? Only when they encode a visible trait. Name the grain, palette, or contrast you want, and the model has something concrete to act on.

Why does my subject look different in every clip? Each generation starts fresh unless it is anchored. Repeat the subject sentence word for word, reuse the style block, and begin from a reference still.

Should I write prompts in my first language? Write in the language you can be precise in, and keep camera and lighting vocabulary in one language throughout. Inconsistent terminology, not a mixed language, is what causes drift.

How do I keep the camera still? Say "locked-off static camera, tripod, no camera movement" and simplify the action. Busy scenes invite the model to invent motion to fill the frame.

What if the take is beautiful but off-brief? Save it as a style reference, then re-read your original logline and find which layer you left vague. That gap is the real problem, not the take.

Can one prompt cover a whole scene? Rarely. Split the scene into shots and generate each one. Sequences are built in the edit, not inside a single long paragraph.

Direct the Frame, Then Refine It

Prompting rewards specificity and patience: one idea per clip, one camera decision, one light source, one palette, one change per iteration. Do that consistently and generation stops feeling like a slot machine and starts feeling like a camera you control.

Orelon is built for that loop — cinematic ideas in motion, from an exploratory first frame to a finished sequence. Start with the shot you already see, describe it in plain language, and refine it in the Orelon AI video generator. If you would rather have structure than a blank page, the video templates are a quick way in.