Orelon logoOrelon
料金

The Power of the Prompt: Writing AI Video Prompts That Work

2026年11月12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: camera language, motion, lighting, iteration loops, shot continuity, and the mistakes that flatten results.

Most dull AI video clips are not a model problem. They are a writing problem. A line like a cinematic shot of a woman walking through a city at night hands the system hundreds of equally valid interpretations, and it will confidently pick one you did not want. The frame comes back soft, the movement drifts, the light is generic, and the creator concludes the tool is not ready.

The tool is ready. It needs a better brief. A prompt is not a magic phrase you recite until something good appears — it is a shot document compressed into a paragraph. Written the way a director writes a shot list, it becomes repeatable, editable, and fast. Written as a mood, it becomes a slot machine.

This guide covers what a prompt actually controls, how to build one layer by layer, how to iterate without burning a whole afternoon on re-rolls, and how to move from a single attractive clip to a sequence that cuts together.

Start with the frame, not the feeling

The most common failure in AI video is describing an interior state instead of a visible image. She feels nervous cannot be rendered. Her eyes flick left, jaw tight, fingers tapping the table edge can. The model has no access to intention; it has access to pixels and the words that predict them.

That single shift — from feeling to frame — fixes more clips than any list of style keywords. Ask yourself after every prompt: if a camera operator read this, would they know where to stand? If the answer is no, the model does not know either.

A useful prompt answers four questions in order:

  • What is in the frame? Subject, state, one action, one or two physical details.
  • How is it captured? Shot size, lens feel, camera behavior, framing angle.
  • How is it lit? Source, direction, contrast, palette, time of day.
  • How does it finish? Grain, sharpness, texture, and the artifacts you want excluded.

When a render disappoints, diagnose which question went unanswered rather than rewriting everything. That habit alone is the biggest speed gain available in an AI video workflow, because it converts random re-rolls into a controlled search.

The levers that actually move a render

Not every word carries equal weight. Some phrases reshape the whole frame; others nudge the finish. Knowing the difference stops you from padding prompts with decoration.

Lever Influence on output Notes
Shot size and framing Very high The single strongest control you have
Camera movement High Keep moves simple and single-purpose
Subject and action Very high One action per clip, always
Lighting direction Very high Mood words alone produce flat frames
Palette Medium-high Two or three colors beat a color story
Lens specifics Medium Plausible focal lengths only
Texture and finish Medium Cheap to add, noticeable in the cut
Style references Medium One reference, not five
Exact on-screen text Low Avoid relying on it
Numeric timing Low Control duration in the tool, not the prose

Two consequences follow. First, specificity is generative: precise language narrows the probability space to something that looks intentional. Second, specificity is cumulative and fragile: six descriptors that contradict each other produce mush. The craft is choosing descriptors that reinforce one visual idea rather than sampling from every idea you like.

Style references are seasoning, not structure. One reference plus concrete visual language outperforms four references with no visual language at all.

Building a prompt in five passes

A reliable prompt reads like one dense sentence or a short paragraph with clear internal order. Build it in passes, adding one layer at a time.

Pass one: subject and action

Name the subject, its state, and the micro-action that fills the clip. Clips are short, which makes one action per clip a hard rule. A cyclist rounds a wet corner is a clip. A cyclist rides through the city, stops for coffee, and meets a friend is three clips and will render as a muddle.

Add one or two anchor details the camera could actually see: a soaked yellow rain shell with the hood up, water beading on the fabric; a flour-dusted apron; a scuffed boot. Anchors give the model something to render beyond a silhouette.

Pass two: camera and lens

State shot size first, then behavior. Use standard vocabulary — extreme wide, wide, medium, close-up, macro, over-the-shoulder, low angle, high angle — and then define movement: locked-off tripod, slow push in, lateral dolly, orbit, handheld follow, crane up, or static with subject motion only.

Lens language adds realism when it is physically plausible. A 24mm wide exaggerates depth and stretches edges. An 85mm compresses the background and flatters faces. A macro lens produces a razor-thin focal plane. Impossible combinations — a 14mm tilt-shift macro with telephoto compression — confuse the render rather than enriching it.

Pass three: light and palette

Describe the source and the direction, not the mood. Single hard key from camera left, deep falloff into black produces a specific image. Moody produces a coin flip. Pair the lighting with two or three colors: teal shadows, sodium-orange highlights, desaturated midtones.

Time of day and weather are efficient shorthand for entire setups. Golden hour, blue hour, overcast noon, and post-rain dusk each imply contrast ratios and color temperatures that models recognize consistently. When a clip looks flat, the missing light direction is almost always the cause.

Pass four: motion, pacing, texture

These three get merged constantly, so separate them. Motion is what moves inside the frame. Pacing is how fast the clip feels. Texture is the finish: fine grain, halation on highlights, clean digital sharpness, archival softness.

A slow push into a static subject reads as patient. A whip pan across a busy street reads as urgent. Say which one you want, and match the pacing to the emotional function of the shot rather than to whatever looked impressive in a demo.

Pass five: constraints

Negative constraints are useful when they target a recurring artifact: no text overlays, no logos, no extra limbs, no fish-eye distortion. Keep the list short and specific. Long prohibition lists dilute the positive description that actually drives the image.

A finished prompt built this way usually lands between 25 and 60 words. Beyond that, contradictions creep in faster than detail accumulates.

One idea, three drafts

Watch the same concept improve across passes.

Draft one: A cinematic video of a coffee shop in the morning. The result is a generic, over-lit interior with ambiguous motion. Nothing is wrong with it and nothing is memorable.

Draft two: Medium shot of a barista pulling an espresso shot in a small coffee shop, morning light through the window, warm tones, shallow depth of field, slow push in. Better. The subject is identifiable, the light has a direction, the camera has intent — but it still looks like stock footage.

Draft three: Medium close-up, 50mm, of a barista's hands locking a portafilter into a chrome espresso machine, steam curling upward through a hard shaft of low morning sun from camera right, dust visible in the light, dark wood counter, muted green wall behind, slow ten percent push in, fine grain, warm highlights with cool shadow fill, no text, no logos.

Now the clip has a subject, a scale, a light source, a color relationship, a motion, and a texture. Crucially, it is editable: if the push-in feels wrong, change one clause. If the wall color clashes with a brand palette, change one clause. That editability is the real payoff of structure — not the first render, but the fifth.

The iteration loop that saves renders

Most creators iterate by rewriting the entire prompt, which makes it impossible to learn what caused the improvement. Treat renders like a controlled experiment.

  1. Lock the subject clause. Do not touch it while you test camera or light.
  2. Run two variants differing in exactly one dimension — 35mm versus 85mm, hard key versus soft window light.
  3. Judge against intent, not against which frame looks prettier in isolation.
  4. Log the result. Save the winning text plus one line on why it worked.

A log is unglamorous and enormously valuable. After twenty clips you own a personal library of phrases that reliably produce the looks you like — the same principle a shared prompt library applies at scale. Iterate on composition with an AI image generator first when stills are enough; a frame you already love is a much easier thing to animate.

One more rule: change the prompt before you change the tool. Creators routinely switch apps when the real problem was a missing light direction.

Prompt shapes for common shot types

Different shot types reward different emphasis.

Establishing shot. Prioritize geography and scale: the location feature, the weather, the time of day, and one dominant line — a road, a river, a row of windows. Keep camera movement slow or absent so the viewer can read the space.

Character beat. Prioritize face, hands, and one gesture. Longer lens, tighter framing, single light source, background reduced to a few recognizable shapes. Describe the gesture as a sequence of small physical events rather than an emotion.

Product macro. Prioritize surface and light. Describe materials honestly — brushed aluminum, matte ceramic, condensation on glass — plus one specular highlight and a slow orbit or rack focus. Skip marketing claims; the camera cannot render a value proposition.

Action insert. Prioritize direction and speed. State the vector (left to right, toward camera) and choose crispness or motion blur deliberately. Short single-purpose clips cut together far better than one clip attempting the whole sequence.

Abstract transition. Prioritize color and motion continuity. These clips exist to bridge scenes, so match the palette of the shots on either side and specify a continuous movement that motivates the cut.

Continuity: from single clips to a sequence

A beautiful clip is a demo. A sequence is a video. Continuity is where most AI workflows fall apart, and prompts are the fix.

Write a continuity block — a short paragraph describing recurring elements — and paste it unchanged into every prompt for that scene. It should cover wardrobe, location architecture, time of day, palette, and film texture. Local variation goes after the block; the block itself stays fixed. Consistency comes from repetition of description, not from one clever phrase.

Then build a shot ladder: wide to establish, medium to orient, close to emotionalize, insert to pace. Five to eight clips usually cover thirty seconds. Plan the ladder before generating anything. Generating first and editing later produces sequences that cut badly and need to be remade.

Match motion across cuts. If a shot ends on a rightward pan, the next shot beginning with rightward movement feels seamless. If a shot ends static, cutting to a fast orbit feels jarring. Audiences cannot explain why these choices work, but they feel them immediately.

Mistakes that flatten output

Stacking contradictory styles. Photorealistic anime watercolor with documentary grit satisfies nothing. Pick one visual world.

Describing emotion instead of behavior. Convert every feeling into an action, a gesture, or a lighting choice.

Overloading a single clip. Six actions in one prompt produce six half-rendered actions. Split them across shots.

Ignoring aspect ratio and duration. A vertical social clip and a widescreen hero shot need different framing logic. Decide delivery format before writing a word.

Neglecting the first and last frame. The opening frame is your thumbnail; the closing frame is your cut point. Describe both when they matter.

Skipping sound. Silence makes good visuals feel like a test render. Plan ambience from the start — room tone, weather, one diegetic sound matching the visible action.

Re-rolling without changing anything. Three identical attempts is gambling, not iteration.

Where prompting sits in the production workflow

A prompt is one stage in a pipeline, not the whole job. A practical order of operations:

  1. Write the concept in one sentence.
  2. Break it into a shot ladder with approximate durations.
  3. Draft prompts per shot using the layered structure.
  4. Generate two variants per shot at low commitment.
  5. Select winners, then refine only those.
  6. Assemble on a timeline, cut to a scratch track, then finish sound.
  7. Grade for consistency with a single adjustment layer.

Teams that skip step two spend their day re-rolling clips. Teams that do it spend their day editing, which is faster and makes better films. When evaluating tools, judge how well they support this pipeline rather than how impressive an isolated demo looks; side-by-side comparisons such as Orelon vs Runway are mostly useful for workflow fit, not for spectacle.

Frequently asked questions

How long should a prompt be?

Long enough to specify subject, camera, light, and texture — usually 25 to 60 words. Past that, returns diminish quickly and contradictions creep in. One dense sentence plus a short constraints clause is a strong default.

Do style keywords like shot on 35mm actually matter?

Yes, but as texture rather than structure. They influence grain, contrast, and highlight behavior. They will not rescue a vague subject or a missing camera direction.

Why do my clips look flat?

Usually because no light direction was specified. Add a source and a side: hard key from camera left, deep shadow falloff. Flatness is nearly always a description problem.

How many variants should I generate per shot?

Two or three, with a changed variable between them. More than that without edits is just rolling dice.

Can one prompt move between different tools?

Mostly yes, with adjustment. Camera and lighting language transfers well; style references and parameter syntax do not. Keep a tool-agnostic core prompt and append tool-specific modifiers at the end.

How do I keep a character consistent across shots?

Use a fixed continuity block with wardrobe, hair, age range, and build, repeated verbatim. Keep the same lens family and lighting setup across the scene. Repetition does the work.

What is the fastest way to improve?

Log every prompt and its result, then review weekly. Pattern recognition beats tips, because your reliable phrases will be specific to the subjects and looks you actually shoot.

Should I write prompts before or after choosing a template?

Choose the delivery format first — duration, aspect ratio, and whether the piece needs titles — then write. Borrowing structure from ready-made video templates early saves a rewrite later.

Start with one shot, written properly

The distance between a vague idea and a finished clip is mostly a writing problem. Describe the subject and its single action, name the camera and its movement, place the light, choose two colors, add texture, and exclude the artifacts you keep seeing. Then change one thing at a time until the frame matches the picture in your head.

Orelon is built for that rhythm — cinematic ideas in motion, with fast iteration so a prompt can become a shot and a shot can become a sequence. Start in the AI video generator when you are ready to render, browse the Orelon blog for workflow patterns, and keep your winning prompts in one place so the next project starts three steps ahead. The prompt is not a technical formality. It is the script, the storyboard, and the direction, compressed into a paragraph — and it is the part only you can write.