Orelon logoOrelon
요금

How to Write Cinematic AI Video Prompts in Chat Assistants

2026년 11월 5일 · Orelon Team 작성

AI 동영상 템플릿 둘러보기

영감을 위해 커뮤니티 창작물 몇 개를 둘러본 다음, 템플릿을 열어 Orelon에서 계속 만들어 보세요.

Learn how to structure cinematic AI video prompts for chat-based generators: camera language, lighting, motion, iteration loops, and a reusable shot template.

Most people type a sentence into a chat assistant, get a four-second clip that looks almost right, and then spend the next hour guessing what to change. The problem is rarely the model. It is that the prompt described a subject but never described a shot. A camera, a lens, a light source, a movement, and a duration are not decoration — they are the instructions that make a generated clip feel filmed rather than rendered.

This guide is a practical system for writing video prompts inside chat-based tools, where you describe a scene in natural language and the model returns motion. You will get a reusable prompt skeleton, camera and lighting vocabulary that actually changes output, a repair loop for bad generations, and the workflow habits that separate a usable shot list from a folder of near-misses.

Why Chat-Based Video Prompting Feels Hard

A chat assistant invites conversation. Video generation rewards specification. Those two modes pull in opposite directions. When you talk to a person, you leave gaps because shared context fills them. When you prompt a video model, every gap is a coin flip — it will invent a lens, a time of day, a wardrobe, and a performance style, and its inventions may not match the ones in the next shot.

That is why the same sentence can produce a gorgeous clip on Monday and an unusable one on Tuesday. The model is not being unreliable; it is resolving ambiguity differently. Your job as a prompter is to reduce the number of decisions the model has to make on your behalf while leaving it enough room to make the image interesting.

The second difficulty is that video prompts are temporal. An image prompt describes a frozen frame. A video prompt describes a frame that changes. If your prompt contains no verb of motion and no duration, the model defaults to something safe: a slow push, a drifting subject, or a near-still shot. That default is fine once. It becomes a stylistic tic across ten clips.

The Six-Slot Prompt Framework

Instead of writing prose and hoping, write in six slots. You do not have to output them as labeled sections — you can fold them into two sentences — but every slot should be answered before you press generate.

1. Subject and action

Who or what is on screen, and what are they doing right now? Use a present-tense verb that implies a body in motion: lifting, turning, pouring, stepping, glancing. "A baker" is a subject. "A baker sliding a tray into an oven and straightening up" is a shot.

2. Environment and time

Where is this, and what time of day is it? Environment sets texture: steam, dust, rain, fluorescent hum. Time of day sets the entire color grade. "Kitchen" is not enough. "A narrow commercial kitchen at 6 a.m., grey pre-dawn light through a high window, flour dust in the air" gives the model a palette and an atmosphere.

3. Camera

Choose a shot size, an angle, and a movement. One of each is usually plenty. Shot size: wide, medium, close-up, macro. Angle: eye level, low angle, overhead, over-the-shoulder. Movement: locked-off, slow push in, handheld follow, orbit, tilt up. If you specify none, expect a generic medium shot with a gentle drift.

4. Light

Name the source and the quality. Source: window, practical lamp, neon sign, fire, overcast sky. Quality: hard or soft, warm or cool, high contrast or flat. Lighting is the single fastest way to make a generated clip look intentional, because it changes shadows, which change depth, which changes perceived production value.

5. Lens and texture

Emulate a real capture format rather than an adjective. "Shot on a 35mm lens with shallow depth of field" behaves differently from "cinematic." Add grain, halation, or a slightly desaturated grade when you want an analog feel. Keep this slot to one or two items; stacking five film references muddies the result.

6. Duration and pacing

State the length and the tempo. "A single unbroken 5-second shot, slow and deliberate" reads differently from "a 5-second shot with a quick whip pan midway." Pacing language tells the model where the energy belongs.

A filled-in prompt might look like this: Medium close-up, low angle, of a mechanic wiping grease from her hands with a red rag, inside a dim garage at night, lit by a single hanging work lamp that throws hard shadows, 35mm lens, shallow focus, slow handheld push in, five seconds, unhurried. Every slot is answered, and the sentence still reads naturally.

Camera Language That Actually Changes Output

Vague camera words are the most common source of disappointing results. "Dynamic camera" means nothing. "Cinematic camera" means less. Replace them with physical descriptions.

  • Locked-off / static frame: the camera does not move. Use this when the subject's motion should carry the shot.
  • Push in / dolly in: the frame tightens toward the subject. Implies realization, tension, or emphasis.
  • Pull out / dolly out: the frame widens, revealing context. Implies loneliness or scale.
  • Handheld follow: the camera trails the subject with slight instability. Implies documentary immediacy.
  • Orbit: the camera arcs around a subject. Implies product beauty or character reveal; use sparingly because it draws attention to itself.
  • Tilt up / tilt down: vertical reveal. Excellent for architecture, tall subjects, and text on signage.
  • Overhead / top-down: flattens the scene into graphic composition. Strong for food, craft, and process shots.

One movement per shot is the rule. Two movements in a five-second clip produces mush, because the model has to interpolate a transition it cannot see. If you need a push and then a pan, generate them as two shots and cut them together in the edit.

Also name the shot size explicitly. Models tend to default to a comfortable medium. If you want a true macro of water beading on glass, say "macro, extreme close-up, focus on the droplets, background fully blurred." If you want an establishing wide, say "wide shot, full environment visible, small figure in the lower third." Composition instructions — subject in the lower third, negative space on the left for text — are surprisingly effective and cost you six words.

Lighting, Color, and Texture as Directing Tools

Lighting vocabulary transfers almost directly from photography, which is why it is worth learning a handful of terms.

Hard light comes from a small or distant source: bare bulbs, midday sun, phone flashes. It creates crisp shadow edges and reads as tension, realism, or grit. Soft light comes from large diffused sources: overcast skies, bounce boards, curtains. It flatters faces and reads as calm or commercial.

Directional cues matter as much as quality. "Backlit" separates the subject from the background and creates rim light. "Side-lit" sculpts faces and reveals texture. "Front-lit" is flat and safe, and usually the least interesting choice.

Color should be named as a relationship, not a mood. Instead of "moody colors," try "cool blue shadows against a warm tungsten practical in the background." Instead of "vibrant," try "saturated teal and orange, high contrast." Naming two colors and their relationship gives the model a grade to apply, and it keeps consecutive shots in the same world.

Texture covers grain, haze, reflections, and surface detail. "Slight 16mm grain, soft halation around highlights" is a texture instruction. "Wet asphalt reflecting neon" is both texture and lighting, and it is one of the most reliable ways to make a night exterior look expensive.

Describing Motion and Time

Video models are conservative about motion because motion is where artifacts appear. If you want energy, you have to ask for it specifically.

Describe motion at three scales. Subject motion is what the character or object does. Secondary motion is what reacts: hair, fabric, steam, leaves, dust, liquid. Secondary motion is what makes a generated clip feel alive, so name one reactive element in almost every prompt. Camera motion is the frame itself.

Then describe tempo. Words like "slow," "deliberate," "abrupt," and "frantic" do real work. A shot of someone running can be a slow-motion hero moment or a panicked sprint; the subject is identical and the clip is not.

Finally, think about the moment you are capturing. A strong video prompt usually implies a before and an after: a hand reaching for a door handle, a glass tipping past the point of no return, a match striking. Choosing an action that is mid-transition gives the model a natural arc within a few seconds, which is far more watchable than a subject standing and smiling.

The Iteration Loop: How to Fix a Bad Generation

Never rewrite the whole prompt after a weak result. Change one variable, generate again, and note what moved. Three rounds of single-variable changes will teach you more than thirty random rewrites.

A simple diagnostic order works well:

  1. Is the subject wrong? Rewrite the subject and action slot only. Keep camera and light fixed so you can compare.
  2. Is the framing wrong? Change shot size and angle. Do not touch anything else.
  3. Is it boring? Add secondary motion and a stronger light source before you add more adjectives.
  4. Is it chaotic? Strip the prompt to subject, camera, and light. Remove every stylistic reference.
  5. Is it inconsistent with the previous shot? Copy the environment, light, and lens slots verbatim from the earlier prompt and change only the action.

Keep a text file of prompts that worked, with a one-line note about what each produced. That file becomes your personal style guide and is worth more than any generic prompt list, because it is calibrated to the way you describe things.

If you want a starting library to adapt rather than write from scratch, browse the prompt library and treat each entry as a template to modify slot by slot.

Three Prompt Patterns Worth Reusing

The product beauty shot. Overhead or 45-degree angle, single soft source with a subtle gradient background, macro detail on texture, slow orbit or slow push, five seconds. The key is one light and one movement — commercial footage is calm, not busy.

Extreme close-up, 45-degree angle, of condensation running down a glass bottle of amber liquid, dark stone surface, one soft key light from the upper left with a cool rim from behind, macro lens, shallow depth of field, very slow push in, five seconds, quiet and precise.

The dialogue-free narrative beat. Medium shot, eye level, a character in the middle of a decision, backlit window, handheld subtle drift, four seconds. The emotional weight comes from the pause, not the action, so keep the movement minimal and let the light do the work.

The seamless loop. Centered composition, locked-off camera, a cyclical action such as a turntable, a pendulum, or steam rising and dissipating, with matching start and end frames. Loops are ideal for social backgrounds and website heroes because they hide the cut.

Building a Shot List Instead of Single Clips

A single clip is a demo. A sequence is a video. Before you generate anything, write a shot list of five to eight lines in plain language: what the viewer sees, in order, and why each shot exists. Then assign one camera setup and one lighting setup per line.

The discipline that matters most is consistency across shots. Choose a small visual bible and repeat it: one lens family, one color relationship, one time of day, one movement style. When you change location, change one thing at a time so the audience feels continuity rather than randomness.

Then edit. Generated clips almost never cut together on their own. Trim the first and last frames where motion settles, match action direction so subjects move the same way across a cut, and add a sound layer — room tone, footsteps, a low pad — because audio does more for perceived quality than another round of regeneration.

Templates help here, especially for formats you produce repeatedly: a 15-second product teaser, a 30-second explainer opener, a vertical social hook. Starting from a structured video template keeps the pacing decisions from being reinvented every time, and lets you spend your attention on the imagery. When you are ready to try a full sequence, the AI video generator is built around cinematic phrasing, and still frames for your opening shot can be drafted in the AI image generator first so you know the composition works before you animate it.

Common Mistakes That Waste Generations

Writing an essay. Long prompts do not equal precise prompts. Four sentences covering six slots beats twelve sentences of atmosphere with no camera.

Stacking style references. "In the style of three famous directors plus a video game plus an anime" produces an average of everything, which is nothing. Pick one reference or describe the look directly.

Forgetting the negative space. If you plan to overlay a headline, say "negative space on the right" in the prompt. Fixing it in post with a crop is always worse.

Asking for on-screen text. Rendering legible words inside generated footage remains unreliable. Add typography in the edit where you control kerning and legibility.

Ignoring aspect ratio. Vertical, square, and widescreen compositions are different pictures. Decide the frame before you write the prompt, not after.

Overloading one clip. A five-second shot cannot establish a location, introduce a character, and deliver a punchline. Give each beat its own shot.

A Short FAQ

How long should a prompt be? Long enough to answer all six slots. For most shots that is 30 to 60 words. If you are past 90 words, you are probably repeating yourself or hedging.

Should I write prompts in full sentences or fragments? Either works. Comma-separated fragments are faster to edit and easier to diff between iterations, which is why many people keep a fragment version and a polished sentence version.

Why does my clip look like a stock video? Usually because the lighting is flat and the camera is static. Add a named light source with direction and one slow movement.

Can I reuse one prompt across different models? The six slots travel well, but each model weights terms differently. Expect to adjust camera phrasing and duration, and keep the subject, light, and color language identical.

How do I get consistent characters across shots? Lock wardrobe, hair, and light in the prompt text, describe the person the same way in every line, and generate a reference still first. Consistency is a system, not a single magic phrase.

What is the fastest way to improve? Generate the same ten prompts every month and compare results. Your vocabulary improves faster when the subject stays constant.

Start With One Shot, Then Build the Sequence

Cinematic prompting is not about finding secret keywords. It is about writing a shot the way a director would call it on set: subject, place, lens, light, movement, and time. Once those six slots become habit, the chat box stops being a slot machine and starts behaving like a camera you can aim.

Pick a scene you already have in mind, write it in six slots, and generate one five-second shot today. Then change exactly one variable — the light, or the camera — and generate it again. That single comparison will teach you more than another hundred prompts copied from the internet. When you are ready to put those shots in motion, Orelon is built for cinematic ideas that need to move, with a blog full of practical workflows to keep your next sequence sharp.