Orelon logoOrelon
料金

AI Video Prompt Writing: A Practical Workflow Guide

2026年11月6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 up: shot structure, camera and lighting language, motion cues, iteration loops, and fixes for common failures.

Most disappointing AI video clips fail before the model even runs. The prompt was underspecified, so the model invented a subject, an action, a camera, a lens, a light source, and a style on its own — and it made those decisions differently from the way you would. Prompt writing for video is not a hunt for magic keywords. It is the practice of taking back the decisions you actually care about and leaving the rest to the model's strengths.

This guide is a working method rather than a list of tricks: what each part of a prompt really controls, how to build a structure you can reuse across projects, how to iterate without burning an afternoon, and how to tell the difference between a prompt problem and a tool problem.

Start with the shot, not the sentence

Writers are trained to build sentences. Video prompts work more like shot lists. Before you type anything, answer four questions in plain language:

  • Who or what is on screen? A person, a product, an animal, a landscape, an abstract texture.
  • What changes between the first frame and the last frame? A turn of the head, a door opening, steam rising, a camera pushing in.
  • Where is the camera, and what is it doing? Locked off, handheld, tracking, orbiting, drifting.
  • What does the light look like? Hard midday sun, soft window light, neon practicals, overcast diffusion.

If you can answer those four questions, you already have a prompt that most models can execute. If you cannot, no amount of stylistic adjectives will rescue the clip — the model will simply fill the gaps with its own defaults, which tend to be generic and slightly over-lit.

The single most common cause of a "bad" generation is a prompt that describes a mood but not an event. "Cinematic, moody, atmospheric" tells a model almost nothing. "A woman in a rain-soaked coat steps under an awning, camera holds at chest height, sodium streetlight behind her" gives it a scene it can actually build.

The four layers of a video prompt

A reliable prompt carries four kinds of information. Thinking in layers makes it easy to debug: when a clip comes out wrong, you usually know which layer failed.

Layer one: subject and setting

Be concrete and short. "A ceramic mug on a wooden counter" outperforms "a beautiful artisanal coffee scene" because the model knows exactly what geometry to render. Include one or two identifying details — material, color, age, condition — and skip the rest. Every extra adjective competes for the model's limited attention.

Layer two: action and change

Text-to-video models are motion engines, not still-image engines. Describe what happens and how fast. "Slowly turns toward the window" is a different generation from "snaps toward the window." If nothing needs to change, say so: "static subject, only fabric moves in the wind" prevents the model from inventing an unnecessary action to justify the clip length.

Layer three: camera and framing

This is the layer most beginners skip and the layer professionals obsess over. Shot size (wide, medium, close-up), height (low, eye level, overhead), angle, and movement all change the emotional read of the same scene. A close-up with a slow push feels intimate; the identical subject in a wide locked-off shot feels observational.

Layer four: look and light

Style, palette, contrast, grain, and lighting direction belong here. Keep this layer to a handful of specific choices. "Warm tungsten key from the left, deep shadows, fine grain" does more work than ten aesthetic adjectives stacked together.

A reusable prompt structure

You do not need a rigid template, but a consistent order of information helps you compare generations fairly. A structure that works across most text-to-video tools:

[Subject + one detail] + [action, with speed] + [shot size and camera move] + [lighting] + [look or finish]

A concrete example:

A weathered fisherman in a wool sweater coils rope on a dock, hands moving slowly, medium shot at eye level, camera drifting left to right, overcast morning light with soft shadows, muted cool palette, fine 35mm grain.

Notice what is absent: no "4K, masterpiece, ultra-detailed, trending on social media" stack. Those tokens rarely change composition and often pull a model toward a generic, over-sharpened look.

Write the negative as a sentence, not a chant

Many tools accept a negative prompt, and many people fill it with a long list of words. A better approach is to name the two or three failures you actually saw. If your last clip had warped hands, say so. If it kept adding a second person, say "single figure only." Negatives should answer real observed problems, not imagined ones.

Keep a spine sentence

When you generate ten variants, the spine sentence — subject, action, camera — should stay nearly identical. Style can flex. If you change everything between attempts, you learn nothing about which change mattered.

Camera and motion vocabulary that changes output

Models respond strongly to certain movement words, and weakly to others. Vague verbs like "dynamic" or "cinematic camera work" produce inconsistent results. Use physical descriptions instead:

  • Push in / dolly in — camera moves toward the subject; increases tension and intimacy.
  • Pull out / dolly out — reveals context; good for endings.
  • Tracking / following — camera moves alongside a subject moving through space.
  • Orbit / arc — camera circles the subject; strong for products and character reveals.
  • Handheld — small, irregular motion; reads as documentary or urgent.
  • Locked off / tripod — no camera motion at all; lets subject motion carry the shot.
  • Tilt up / tilt down — vertical reveal; useful for scale.
  • Rack focus — attention shifts between foreground and background.

Pair each move with speed. "Slow push in" and "fast push in" are different tools with different emotional jobs. For anything under five seconds, slow is usually the better default: fast moves have less time to read and tend to look like a glitch rather than a choice.

Motion inside the frame

Do not forget subject motion. Wind in hair, steam, dust, rippling fabric, and drifting smoke are cheap, reliable ways to make a static composition feel alive. If the environment is doing the moving, you often do not need camera movement at all — and a locked-off shot with strong internal motion usually looks more controlled than a wandering camera.

Light, lens, and texture as precision tools

Lighting is where amateur prompts and professional prompts diverge most. A few directional choices cover most needs:

  • Golden hour backlight — warm rim, soft flare, romantic.
  • Hard midday sun — short shadows, high contrast, unsentimental.
  • Overcast diffusion — even, flat, ideal for skin and product detail.
  • Practical sources — neon signs, screens, lamps in frame; strong color and believable motivation.
  • Single-source key — one dominant light, deep falloff, dramatic.

Lens language adds control that style words cannot. "Wide lens, close to subject" produces distortion and a sense of space; "long lens, compressed background" isolates a figure. Mention shallow depth of field only when you actually want the background blurred — for a landscape, it usually hurts.

Texture words are the finishing layer: fine grain, clean digital, slight halation, soft contrast. Use one. A prompt with a single considered texture term reads more intentional than one with five.

The three-pass method for a single shot

Professionals rarely get a usable clip on the first attempt, and they rarely get it by randomizing. They converge in passes.

Pass one: block the shot. Write only subject, action, and camera. Ignore style entirely. Generate two or three clips and check whether the composition and motion are correct. If the blocking is wrong, style is irrelevant.

Pass two: fix the light. Once the movement reads, add lighting and a palette. Regenerate. You are now adjusting a variable you know is the only thing that changed.

Pass three: finish it. Add texture, lens character, and any negative prompt for the specific artifacts you saw. This is where you decide whether the clip is good enough or needs a different approach entirely.

This costs a few more generations than a single big prompt, but each generation teaches you something. The alternative — stacking twenty adjectives and hoping — produces clips you cannot reproduce or explain.

Diagnosing output that misses the brief

When a clip fails, classify the failure before you rewrite. Most problems fall into five buckets, and each has a different fix.

What you see Likely cause Fix
Subject morphs or duplicates Overloaded prompt, contradictory details Cut the prompt to subject, action, camera; add "single figure"
Motion is frantic or jittery No speed specified, or too much happening Add "slow" or "subtle"; remove one action
Camera ignores instruction Movement described vaguely Use a physical verb: dolly, orbit, tilt, handheld
Image is flat and over-bright Style layer too thin or generic Add a specific light direction and a contrast term
Composition is fine but boring No event in the frame Add internal motion: wind, steam, a turn of the head

If the same prompt produces wildly different results across two tools, that is not a prompt problem — it is a model-character problem. Some models favor realistic physics; others favor stylized, high-contrast imagery. Matching the tool to the shot is part of the craft. If you are comparing engines, it helps to read side-by-side breakdowns such as Orelon vs Kling AI before rewriting a prompt that was never going to work on that model.

Mistakes that quietly ruin otherwise good prompts

Adjective stacking. "Epic, stunning, breathtaking, hyper-realistic" adds no spatial information. Every term like this dilutes the ones that matter.

Describing a still. If the prompt could describe a photograph, the model has nothing to animate and will invent motion you did not want.

Ignoring clip length. A four-second clip cannot contain a full character arc. Write for the duration you actually have: one action, one camera move, one beat.

Changing everything at once. If attempt two changes subject, lens, light, and style, you cannot tell which change fixed or broke the shot.

Fighting the model's defaults. Every engine has a look. If a model consistently renders soft, warm, shallow images, do not write a cold, deep-focus prompt and blame the tool. Either work with the grain or switch tools.

Forgetting the first frame. Many workflows start from a reference image. If you are generating from a still, the prompt should describe what happens next, not re-describe the image.

A worked example: a 30-second product teaser

Say you are building a short teaser for a ceramic pour-over coffee set. Six shots, roughly five seconds each. Here is how the prompt work maps to the edit.

  1. Establishing texture. "Macro shot of unglazed ceramic rim, water droplets beading, camera slowly drifting right, soft north window light, neutral palette, fine grain."
  2. The pour. "Hot water spirals into a ceramic dripper, steam rising, medium close-up at eye level, locked off camera, warm tungsten key from the right, shallow depth of field."
  3. The hands. "Two hands lift the carafe, slow motion, tight close-up, low angle, soft diffused light, muted warm palette."
  4. The room. "Wide shot of a quiet kitchen counter at dawn, single figure in the background, camera pulls out slowly, cool ambient light, gentle contrast."
  5. The detail. "Condensation on glass, tiny motion, extreme close-up, hard side light, high contrast, clean digital finish."
  6. The close. "The full set on a wooden board, camera orbits slowly left, soft window light, warm neutral palette, fine grain."

Each prompt is one action and one camera behavior. Nothing is asked to do two jobs. That discipline is why the sequence cuts together: the shots share a palette and light logic because those were specified deliberately rather than left to chance.

You can speed this up by starting from a video template or a saved prompt library entry and editing the specifics, rather than writing from a blank field every time.

Choosing the tool around the prompt

Prompt skill travels; model character does not. Before committing to a workflow, test the same well-built prompt on two or three engines and compare:

  • Motion realism — do people move like people, or like animated mannequins?
  • Instruction adherence — does the camera do what you asked, or something adjacent?
  • Duration handling — does the clip stay coherent for the full length you need?
  • Style range — can it do both grounded realism and stylized looks?
  • Iteration speed — how quickly can you test the same spine sentence again?

If you mostly make product and lifestyle footage, prioritize light and texture control. If you make character-driven narrative, prioritize motion realism and face stability. If you make abstract or stylized work, prioritize palette range. A short list of AI video generator alternatives is useful here, but test with your own prompts rather than someone else's demo reel.

FAQ

How long should a video prompt be? Long enough to specify subject, action, camera, and light — usually one to three sentences. Beyond that, extra words compete rather than clarify. If you need more control, add a reference image instead of more text.

Do style keywords like "cinematic" still help? They help slightly, but they are vague. Replace them with a specific light direction, lens behavior, or palette. "Cinematic" is a goal; "hard side light, shallow focus, cool shadows" is an instruction.

Why does the same prompt give different results each time? Generation is stochastic. If you need consistency, freeze as many variables as possible: same seed when available, same aspect ratio, same duration, and one reference image. Then change one element per run.

Should I write prompts differently for image-to-video? Yes. Describe motion rather than appearance. The image already carries the subject, framing, and palette; the prompt should say what moves, how fast, and what the camera does.

How do I stop hands and faces from warping? Reduce competing detail, avoid extreme close-ups of hands unless the model handles them well, shorten the clip, and specify a single subject. Sometimes the fix is a different engine rather than a better sentence.

Can I reuse one prompt across a whole campaign? Yes — that is the point of a spine sentence. Keep subject, action, and camera fixed, and swap only the look layer to produce a visually consistent series.

Write the shot, then let the model shoot it

Prompt writing is a small craft with a fast feedback loop. Decide what is on screen, what changes, where the camera stands, and how the light falls. Say those four things clearly, test them in passes, and keep notes on which words actually moved the output. Everything else — the adjectives, the trend words, the long tails of modifiers — is noise you can safely delete.

When you are ready to put the method to work, generate a video with Orelon, start from a saved structure in the prompt library, and iterate on one variable at a time. Cinematic ideas move faster when the prompt stops guessing and starts directing.