Orelon logoOrelon
料金

AI Video Prompts That Work: A Practical Prompting Guide

2026年11月7日 · Orelon Team 著

AI動画テンプレートを見る

着想のためにコミュニティ作品をいくつか閲覧し、任意のテンプレートを開いて Orelon で作成を続けましょう。

Learn how to write AI video prompts that hold character, camera, and lighting together, with structures, examples, and fixes for common failures.

Most disappointing AI video clips are not model failures. They are briefing failures. The model rendered something; it just wasn't the thing you pictured, because the prompt described a mood instead of a shot. "Cinematic drone shot of a city at night, dramatic" gives a generator a thousand valid answers. "Low drone push over wet rooftops at dusk, sodium streetlights, 24mm, slight downward tilt, rain mist catching the light" gives it one.

That difference is learnable. Prompting for video is a craft with repeatable parts: a subject, a single action beat, a camera position, a light source, an environment, and a technical finish. Once you can name those parts, you can debug a bad clip instead of re-rolling and hoping.

This guide covers the structure, the vocabulary, and the workflow — including the parts most prompt lists skip, like keeping a character consistent across six shots and turning a script into a shot list you can actually generate.

Why the prompt decides the shot more than the model

Generative video resolves ambiguity by averaging. If your prompt leaves the shot size open, the model picks something generic. If it leaves the time of day open, you get flat mid-afternoon light. If it leaves the action open, you get a slow drift with no event — technically video, emotionally nothing.

Three kinds of ambiguity cause most weak clips:

  • Spatial ambiguity. No stated shot size, angle, or lens, so the model defaults to a medium-wide view.
  • Temporal ambiguity. No stated action beat, so nothing happens inside the clip's duration.
  • Atmospheric ambiguity. No stated light source or palette, so the grade lands in the middle of everything.

Model choice matters, but it rarely rescues a vague prompt. A thin prompt produces thin results everywhere; a precise prompt travels well between engines. That portability is the real argument for learning structure — the skill survives every version upgrade. You can test it directly in the AI video generator and watch how one added camera clause changes the whole frame.

The six building blocks of a usable video prompt

Every prompt that reliably works contains six pieces of information, even if the sentence order varies.

Block Vague version Specific version
Subject "a woman" "a woman in her thirties, olive raincoat, wet hair stuck to her forehead"
Action beat "walking" "steps off a curb and stops as a bus passes"
Camera "cinematic shot" "medium shot, 50mm, eye level, slow push in"
Light "dramatic lighting" "single sodium streetlamp from camera left, hard shadow falloff"
Environment "city street" "narrow brick alley, steam vent, puddles reflecting signage"
Finish "high quality" "2.39:1, shallow depth of field, fine grain, no text overlays"

Subject and action belong together

Describe the subject with the details that would survive a costume change: age range, build, wardrobe, hair, and one identifying prop. Then attach exactly one action beat. One clip should contain one event — "she lifts the lid and looks inside" is a beat. "She discovers the truth about her family" is a synopsis, and the model will answer it with a slow zoom on a face.

If the beat needs two movements, split it into two clips. A prompt describing "she enters, crosses the room, opens the drawer, and finds the letter" asks the model to compress four beats into a few seconds, which usually produces a smear somewhere in the middle.

Camera, lens, and movement

Camera language is the highest-leverage addition to any prompt, because it changes composition rather than surface. Name four things: shot size, angle, lens, and movement. "Wide shot, low angle, 24mm, slow crane up" is a complete camera instruction. "Close-up, 85mm, static, slight handheld sway" is another.

Keep movement motivated. A push-in implies growing attention; a pull-back implies withdrawal or reveal; a handheld track implies urgency. Unmotivated camera moves read as stock footage, which is the exact feeling most creators are trying to avoid.

Light, palette, and finish

State where the light comes from and what quality it has. "Soft window light from behind, overcast, cool shadows" produces a very different image from "hard key from camera right, warm tungsten, deep falloff." Add a palette only if you care about it — two or three colors, not a mood board.

Finish covers the technical frame: aspect ratio, depth of field, grain, and what you don't want. This is also where you mention text: generated signage and labels are unreliable, so either specify the exact wording or ask for a frame with no legible text.

A prompt skeleton you can reuse for every project

The most practical habit is a fixed order. It makes prompts faster to write and much faster to debug, because when a clip fails you can point at the clause that caused it.

[shot size] of [subject with wardrobe and identifying detail],
[action beat],
[environment with one or two texture details],
[light source and quality],
[camera: angle, lens, movement],
[finish: aspect ratio, depth of field, grain, avoid list]

Filled in:

Medium close-up of a welder in his fifties, soot on his cheekbones, canvas apron, lifting a mask with one gloved hand, inside a dim fabrication shop with sparks drifting past, hard amber work lamp from below camera, 65mm, slight handheld, 2.39:1, shallow focus, fine grain, no text.

Two rules keep this skeleton honest. First, the opening clause carries the most weight, so lead with whatever matters most — usually subject or shot size. Second, keep prompts between roughly 35 and 80 words. Under 35, you're leaving decisions to the model. Over 80, clauses start fighting each other and the model drops the ones in the middle.

When a clip fails, change one clause at a time. Changing three variables at once is how people conclude that prompting is random.

One idea, three prompt levels: a worked example

Take a simple idea: a cyclist crossing a bridge at dawn. Here it is at three levels of control.

Level 1 — concept only. "A cyclist crossing a bridge at dawn, cinematic." Expect a wide, slow, generic shot with warm haze. Useful only for exploring whether the idea has legs.

Level 2 — structured. "Wide shot of a cyclist in a yellow jacket crossing an empty steel bridge, seen from behind, sunrise fog, warm rim light, slow tracking shot at bike pace, 2.39:1." Now the frame is legible and the viewer knows where they are.

Level 3 — production grade. "Wide shot from a low bridge rail, cyclist in a faded yellow rain jacket and black cap riding left to right across a steel truss bridge, fishing boats below in mist, sunrise backlight creating a rim on the jacket, 35mm, smooth lateral tracking at cycling speed, 2.39:1, cool shadows with warm highlights, light grain, no text, no logos."

Level 3 is not automatically the right choice. If you're still deciding what the scene is, stay at level 2 and vary one element per take — time of day, lens, or direction of travel. Precision is for shots you've already committed to. Exploration needs room.

Camera and lens language that changes the frame

Most prompt writers use three camera words. There are more, and each one does a specific job.

Term Effect
Wide establishing Shows geography and isolation; subject small in frame
Medium shot Default for dialogue and readable body language
Close-up Forces attention on face or object; kills context
Insert / macro Texture and proof — hands, tools, screens, fabric
Low angle Power, threat, monumentality
High angle Vulnerability, overview, smallness
Dutch tilt Instability; use once per project or it becomes a tic
Dolly in Rising focus, realization
Truck / lateral track Travel, parallelism, momentum
Crane up Release, scale reveal, endings
Handheld Immediacy and documentary honesty
Rack focus Redirects attention without cutting
Slow motion Elevates a small action into a moment

Two failure modes dominate. The first is contradiction: "static handheld whip pan" asks for three incompatible things and produces mush. The second is over-stacking: three camera moves in a five-second clip means none of them complete. Pick one primary move per shot and let the rest be framing.

Keeping characters and props consistent across shots

A single beautiful clip is a demo. A sequence is a project, and sequences live or die on continuity. Generators don't remember your character between prompts unless you give them something to remember.

Build a character sheet before you generate anything. Write down five anchors: face shape and age range, hair, signature wardrobe item, one prop, and posture or gait. Then reuse those anchors verbatim in every prompt. Words like "the same woman in the olive raincoat, dark hair tied back, silver hoop earrings" do more continuity work than any stylistic adjective.

Where the tool supports reference images, use them. Generate a clean character still in the AI image generator, then feed it as a reference for each shot. Reference images lock identity; text alone drifts. Keep the wardrobe constant and change only the camera and environment — audiences read a wardrobe change as a different person, or a different day.

For props, describe material and wear: "a scuffed brass lighter with a chipped enamel crest" stays recognizable across angles in a way that "a lighter" never will. Keep the prop in the same hand, too. Switching hands between shots is the continuity error viewers notice without knowing why.

Turning a script into a shot list of prompts

Prompting gets much easier when it stops being a creative act performed one clip at a time and becomes a translation step inside a workflow.

  1. Write a beat sheet in plain language. Six to eight beats, one sentence each, no camera language. This is the story.
  2. Convert each beat into a shot line. Subject, action, shot size. Nothing else yet.
  3. Assign coverage. For each beat, decide whether you need an establishing frame, a medium, a close-up, or an insert. Most scenes need two of the four, not all four.
  4. Write prompts from shot lines. Fill the skeleton. Keep continuity anchors identical across the whole sequence.
  5. Generate in batches of three or four per prompt. Keep the best take, and note in one line what failed in the others. That note is your debugging trail.
  6. Edit for usable seconds. Plan around short segments; assemble rhythm in the edit rather than trying to generate a single long take.

If you're starting from a format rather than a story — a product reveal, a title sequence, a social cutdown — begin with an existing video template and rewrite its prompt rather than starting from a blank field. Templates encode structure you'd otherwise have to invent.

Common failures and how to fix them

Symptom Likely cause Fix
Face morphs mid-clip Subject under-described, or too much action in one beat Shorten the beat, add reference image, add wardrobe anchors
Extra limbs or hands merging Complex interaction described in one clause Simplify to one gesture; use inserts for detail
Camera drifts or re-frames No movement specified, or two movements stacked Name one primary move and one framing
Action ignored Beat buried mid-prompt Move the action to the first clause
Flat, gray lighting No light source stated Specify direction and quality of the key light
Garbled signage or labels Text requested implicitly by the scene Ask for no legible text, add copy in the edit
Pacing feels rushed Too many beats in a short duration Split into two clips
Wrong frame shape Aspect ratio omitted State it in the finish clause every time

A useful diagnostic: if the clip is beautiful but wrong, your atmosphere was specific and your action wasn't. If it's right but ugly, the reverse.

FAQ

How long should a video prompt be?

Most well-formed prompts land between 35 and 80 words. Shorter prompts hand composition decisions to the model; much longer prompts start losing clauses. If you need more detail, add it as a separate reference image rather than more sentences.

Do negative prompts help?

Moderately, and only when short. "No text, no logos, no extra fingers" is useful. A twenty-item avoidance list usually dilutes the positive instruction. Describe what you want first; constrain second.

Why does the same prompt give different results each time?

Generation involves sampling, so variation is normal and often desirable. Reduce it by making the prompt more specific and, where the tool allows, reusing a seed or a reference image. Treat the first four takes as a search, not a verdict.

Can one prompt generate a whole scene?

Usually not well. A prompt describes a shot, not a scene. Build scenes from multiple short clips that share continuity anchors, then assemble them in the edit. Planning for three-to-five-second usable segments keeps expectations realistic.

Should I write prompts in English if my project isn't in English?

Prompt in whichever language the model handles best, and treat dialogue and on-screen copy as a separate post-production step. Language in the prompt affects rendering quality; language in the finished video is an editorial choice.

How many takes should I generate before changing the prompt?

Three or four. If none of them are close, the prompt is the problem, not the luck. Change one clause — usually the shot size or the light — and try again.

Start with one shot

Pick a single image you already know you want: a specific person, in a specific place, doing one specific thing, lit by one specific source. Write it in the six-block order, generate four takes, and change exactly one clause between rounds. That loop — brief, generate, diagnose, revise — is the entire skill.

When you're ready to put it into practice, Orelon gives you a place to turn those briefs into moving frames, with a prompt library of reusable structures to build from and more workflow breakdowns on the Orelon blog. Start with one shot. Get it right. Then build the sequence.