Orelon logoOrelon
料金

AI Video Prompting: Write Prompts That Direct a Scene

2026年11月11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 structure AI video prompts with subject, action, camera, and light cues, plus negative prompts, seed iteration, and shot chaining.

A prompt is not a wish. It is a shot list compressed into a paragraph. When you type a sentence into an AI video generator, the model does not read your intent — it reads your words, weighs them, and resolves them into pixels, motion, and time. That is why two people can describe "the same" scene and get results that feel like different films.

Good prompting is less about memorizing magic phrases and more about thinking like a director, an editor, and a lighting designer at the same time. Below is a practical system: the structure that keeps a shot coherent, the vocabulary that controls camera and light, how to iterate without losing what worked, and how to chain prompts into sequences that read like scenes rather than random clips.

Why a prompt directs a scene instead of describing it

A description answers "what is in the frame." A direction answers "what happens, from where, and how it feels." Most weak AI video output comes from descriptions: "a woman in a forest, cinematic." The model has no reason to choose a lens, a time of day, or a camera height, so it averages everything and produces something generic.

Direction closes those gaps. Compare:

  • Description: "a knight in a forest."
  • Direction: "Low-angle medium shot of a knight in dented plate armor pushing through waist-high ferns, overcast dawn light, slow handheld push-in, cool desaturated grade."

The second version is not longer because it is fancier. It is longer because every clause removes a decision from the model and makes it on the model's behalf. That is the entire job.

There is a second reason direction matters: video is temporal. A still image only has to be right at one moment. A clip has to be right across hundreds of frames, which means motion, continuity, and physics all need a described intention. If you do not specify how something moves, the model invents movement — often drifting, warping, or sliding in ways that break the illusion.

The four-part frame: subject, action, camera, light

A reliable prompt answers four questions in order. Expand each part as needed, but do not skip any of them.

Subject: the specific thing the frame is about

Name one primary subject, then add two or three identifying details that survive motion. Details that survive motion are structural — hair shape, clothing silhouette, posture, a held object — rather than micro-detail that smears at frame rate.

Weak: "a woman." Stronger: "a woman in her thirties with close-cropped silver hair, wearing a waxed canvas jacket."

Avoid stacking more than one hero subject per shot. Models handle two interacting subjects well when their relationship is explicit ("a man and a child walking hand in hand"), but ambiguous crowds turn to mush. If a scene needs a crowd, describe it as texture: "a blurred crowd of commuters in the background, shallow depth of field."

Action: verbs with a beginning and an end

Video prompts need verbs that imply a change of state: pushes open, turns toward, lifts, settles, steps back, exhales. Static verbs ("is standing," "is looking") give the model nothing to animate, so it fills the gap with drift.

Where possible, describe a small arc rather than a pose. "She lifts the letter, reads it, and her expression drops" gives the generation a shape to build across the clip. On short durations, keep it to one arc: one action, one reaction.

Camera: position, lens, and movement

Camera language is the highest-leverage vocabulary in AI video prompting, because it changes composition and motion simultaneously. Useful building blocks:

  • Shot size: extreme close-up, close-up, medium, medium-wide, wide, establishing.
  • Height and angle: low angle, eye level, high angle, overhead, dutch tilt.
  • Lens feel: 24mm wide with edge distortion, 50mm neutral, 85mm compressed portrait, macro.
  • Movement: locked-off static, slow push-in, pull-back reveal, lateral tracking, orbiting arc, handheld follow, crane up.

Combine two at most. "Slow push-in on an 85mm" reads clearly. "Push-in while orbiting while craning while zooming" reads as chaos and usually produces a warped frame.

Light: source, quality, and color

Lighting is where amateur prompts lose their cinematic feel. Specify three things: where the light comes from, how hard it is, and what color it is.

  • Source: window light, practical neon, golden-hour sun, overcast sky, single bare bulb.
  • Quality: hard shadows, soft wraparound, dappled, smoky volumetric beams.
  • Color: warm amber, cool moonlight blue, sodium orange, neutral daylight.

"Late afternoon sun raking through blinds, hard shadow bars across the wall, warm amber grade" will consistently outperform "good lighting."

Style, format, and grade: controlling the look

Once the four-part frame is in place, style tells the model which visual tradition to draw from. It is not decoration; it is a constraint that keeps output consistent between takes.

Reference attributes, not studios

Instead of naming a specific studio or living director, describe the attributes you actually want: "1970s Kodak film stock, visible grain, warm halation on highlights, slightly muted greens." Attribute-based style descriptions are more portable and less likely to produce an imitation you did not ask for.

Format terms worth knowing:

  • Aspect ratio: 16:9 landscape, 9:16 vertical, 2.39:1 anamorphic feel.
  • Texture: film grain, digital clean, VHS softness, 16mm bounce.
  • Grade: teal-and-orange, bleach bypass, warm pastel, high-contrast monochrome.
  • Motion feel: cinematic cadence, sports smoothness, slight shutter blur.

Match style intensity to duration

Short clips reward strong, single-idea styling. Longer generations reward restraint, because every flourish has more time to contradict itself. For a five-second shot, a bold grade and a distinctive lens are assets. For a longer sequence, keep the style block identical across prompts and let only the content change.

Negative prompts: removing what keeps appearing

Negative prompts are subtractive control. They are most useful as a reaction to a repeating failure rather than as a standing list.

A practical approach: run two or three takes first, note what recurs, then add only those items. Common additions include:

  • Artifacts: extra fingers, warped faces, duplicated limbs, text overlays, watermarks.
  • Motion problems: jitter, strobing, morphing background, sliding feet.
  • Aesthetic problems: oversaturated neon, plastic skin, stock-photo smiles, flare spam.

Keep the list short. Every excluded term consumes attention that could go toward what you do want, and long negative lists tend to flatten images toward blandness. When a scene needs a specific mood, replacing a negative with a positive works better: instead of "no harsh light," write "soft diffused window light, gentle falloff."

Seeds, takes, and iteration: refining without starting over

Most good AI video comes from the third or fourth take of a good prompt, not the first take of a perfect one. Iteration is a discipline, and it has rules.

Change one variable at a time

If you change the camera, the light, and the wardrobe in the same revision, you learn nothing about which change helped. Keep a working prompt stable and adjust a single clause. That is slower per attempt and dramatically faster overall.

Lock a seed when something works

Seed values let you reproduce the underlying noise pattern of a generation. Once a composition lands, fixing the seed and editing only the prompt gives you controlled variations of the same shot rather than new shots entirely. If your tool does not expose a seed, keep a written record of the exact prompt and settings that produced the good result — reproducibility by documentation still works.

Evaluate against a checklist, not a vibe

Before accepting a take, check:

  • Is the subject consistent from first frame to last?
  • Does the motion resolve, or trail off into drift?
  • Is the light direction stable across the clip?
  • Does the camera move as described?
  • Would this cut against the shot before and after it?

Any "no" is a prompt problem to fix, not a reason to reroll blindly.

Prompt chaining and shot lists: from clips to scenes

Single clips impress; sequences tell stories. Prompt chaining means writing a series of prompts that inherit a shared style block so the results cut together.

A simple template:

  1. Write the style block once — lens family, grade, grain, aspect ratio, time of day.
  2. Write each shot as subject, action, camera, and light, reusing the style block verbatim.
  3. Vary shot size and camera movement between adjacent shots so edits have rhythm.
  4. Generate a coverage set: one wide, one medium, one close-up, one insert.

The last point matters more than most people expect. If every prompt produces a medium shot, your edit has no rhythm. Deliberately generating an establishing wide, a medium for dialogue, and a close-up for reaction gives you something that can actually be cut.

Keep a shot list document with columns for shot number, prompt, seed, duration, and status. It sounds bureaucratic until you are twenty generations deep and cannot remember which version produced the good take. If you build sequences often, saving a few reusable starting points in a prompt library shortens the blank-page phase considerably.

Prompt patterns for common video types

Different genres reward different emphasis. A few starting patterns:

Product and object shots. Prioritize surface, light behavior, and a single controlled movement: "macro shot of a brushed steel watch face, hard key light from the left, slow 15-degree orbit, shallow depth of field, reflections sliding across the bezel."

Character and performance. Prioritize face, eyeline, and micro-action: "close-up of a man in his fifties, eye level, soft window light from the right, he exhales and looks down, minimal camera movement, muted grade."

Landscape and environment. Prioritize scale, atmosphere, and slow motion: "wide establishing shot of a fog-filled valley at sunrise, layered ridges receding into haze, very slow crane up, cool blue shadows with warm rim light on the peaks."

Abstract and motion graphics. Prioritize rhythm and material: "liquid chrome ribbons flowing against black, rim-lit edges, continuous slow rotation, high-contrast grade, no background detail."

For teams producing repeatable formats, starting from an existing video template and rewriting the prompt inside it is usually faster than starting from an empty field. If you also need stills for storyboards or thumbnails, generate them in an AI image generator first and use them as visual reference for the video prompt.

Mistakes that flatten AI video output

Most disappointing generations share a handful of causes:

  • Prompt symphonies. Ten adjectives competing for attention produce an average of all of them. Cut to the two that define the shot.
  • No motion intention. A beautiful description with no verbs yields a drifting, lifeless clip.
  • Contradictory light. "Golden hour" plus "neon interior" plus "harsh overhead" gives the model an irreconcilable brief.
  • Over-stuffing one shot. Two locations, three characters, and a costume change cannot fit in five seconds.
  • Ignoring continuity. Prompts written independently produce clips that cannot be cut together, no matter how good each one looks.
  • Rerolling instead of reading. If three takes fail the same way, the prompt is the problem.

A practical workflow from idea to final clip

Here is a repeatable loop that works for solo creators and small teams alike:

  1. Write the shot in plain language. One sentence about what the audience should feel and see.
  2. Convert it to the four-part frame. Subject, action, camera, light.
  3. Add a style block. Lens, grade, texture, aspect ratio.
  4. Generate two takes at short duration. Cheap, fast, and enough to reveal problems.
  5. Diagnose, then edit one clause. Fix the biggest failure first.
  6. Lock the seed and extend. Once composition works, push duration or variations.
  7. Assemble and cut. Judge the clip in the timeline, not in isolation — many shots that look mediocre alone cut beautifully.

When you are ready to run the loop yourself, you can generate a video from a prompt and iterate against the same project, and the Orelon blog has more breakdowns of specific styles and formats.

FAQ

How long should an AI video prompt be?

Long enough to answer subject, action, camera, and light — usually 25 to 60 words. Past that, additional clauses tend to dilute rather than refine. If you need more control, split the idea into multiple shots instead of one enormous prompt.

Do negative prompts actually change the output?

Yes, but they work best as targeted fixes. Add them after you observe a recurring artifact, not before. A short, specific list of three to six items outperforms a long generic one.

Should I describe camera movement in every prompt?

Describe it in most prompts, even if the answer is "locked-off static." Explicit stillness prevents the model from inventing drift, and a stated movement gives the clip a reason to exist as video rather than a slideshow.

Why does my character's face change between takes?

Consistency comes from specificity plus constraints: a stable character description, a fixed seed where available, the same lens and light block, and similar shot sizes. Changing camera distance between takes while keeping the description identical is one of the most common causes of identity drift.

Can I reuse the same prompt for images and video?

Yes, with edits. Still prompts can carry more texture and micro-detail; video prompts need motion verbs, a camera instruction, and a duration that the described action can plausibly complete. Trim detail, add movement.

How many takes should I expect before a usable clip?

Plan on three to five for a controlled shot and more for anything with complex motion or multiple subjects. Batching takes and evaluating them side by side is faster than judging each one in isolation.

Direct your next scene with Orelon

Prompting is a craft of subtraction and intention: describe less, direct more, and change one thing at a time until the shot behaves. The creators who get consistently cinematic results are not using secret words — they are running a tighter loop.

Orelon is an AI video generator built for cinematic ideas in motion, so you can write a directed prompt, iterate quickly, and assemble clips into something that cuts. Start with a single shot, apply the four-part frame, and see how much of the scene you can control before you ever touch an editor.