Learn how to write AI video prompts that hold together: shot structure, camera language, lighting, motion, negatives, and iteration habits that deliver usable clips.
A prompt for an AI video generator is not a wish. It is a compressed shot list: subject, action, lens, light, motion and mood, written in the order a camera would meet them. Most disappointing AI clips are not a model failure — they are the predictable result of a prompt that described a vibe and hoped for a film. The distance between 'cinematic city at night' and a usable five-second shot is mostly vocabulary and structure. This guide covers both, plus the iteration habits that turn one lucky generation into a look you can repeat across an entire project.
Start With the Shot, Not the Sentence
Beginners write prompts the way they write captions. The result reads like a description of a video rather than direction for one. Before typing, decide two things: what the camera is doing, and what changes between the first frame and the last. A clip without a change is a still image with drift.
Compare these two prompts.
Weak: cinematic drone shot of a city at sunset.
Strong: Slow aerial push over a coastal city at golden hour, camera tilted twenty degrees down, traffic streaming along a seafront road, long shadows across rooftops, 24mm equivalent, deep focus, warm haze, gentle film grain.
The second version is not longer for the sake of length. Every clause maps to a control the model can act on: movement type, angle, subject detail, time of day, lens character, focus behaviour, and finish. If you removed any clause, the output would change in a specific, predictable way. That is the test for whether a phrase belongs in your prompt.
A useful habit is to say the prompt out loud as if you were briefing a camera operator standing next to you. If a sentence would make that operator ask a question, the model will guess — and it will guess generically.
The Six Slots Every Video Prompt Needs
Order matters less than coverage. Strong prompts usually touch six areas, and skipping one is what produces the vague, floaty results people blame on the tool.
Subject and wardrobe
Who or what is on screen, with two or three concrete details. A woman in her thirties, cropped wool coat, damp hair beats a person. Specific materials — wool, nylon, wet cotton, brushed steel — give the renderer something to model light against.
Action and the turn
One primary action per clip, plus the change that makes it a clip. She turns from the window and steps toward the door is a beat. She walks around the city thinking is a synopsis and will produce mush.
Setting, time, weather
Place, hour, and atmosphere. Overcast noon and blue-hour drizzle are different films. Name the weather only if it does work: rain adds reflections, wind adds cloth motion, fog compresses depth.
Camera and lens
Shot size, angle, movement, and focal length. Medium close-up, eye level, slow push in, 50mm is a full instruction. Without it, models default to a neutral mid-shot that reads as stock footage.
Light and palette
Where the light comes from, how hard it is, and what colours dominate. Single warm practical behind her, cool ambient fill, deep shadows communicates more than moody lighting.
Texture and finish
Grain, haze, contrast, and the overall render feel. Fine 35mm grain, slight halation on highlights, muted contrast tells the model you want photochemical texture rather than glossy CGI.
Filling all six slots takes about forty words. That is the sweet spot for most generators: enough constraint to be specific, short enough that nothing gets diluted.
Camera Language the Model Actually Reads
Camera vocabulary is the highest-leverage part of a video prompt because it changes composition, not just mood. Four dials are worth learning.
Shot size. Extreme wide, wide, medium, close-up, macro. Shot size controls how much context the viewer gets, and models respond to it more reliably than to adjectives.
Angle. Eye level, low angle, overhead, Dutch tilt. A low angle makes a subject dominant; an overhead flattens a scene into geometry. State it explicitly or you will get eye level by default.
Movement. Locked-off, slow push in, pull out, dolly, truck, crane up, orbit, handheld follow, whip pan. Pick one primary movement per shot. Combined movements like static handheld orbit contradict each other and produce warping.
Lens and depth. 24mm for wide environmental drama, 50mm for natural perspective, 85mm for compressed portraits, macro for texture. Add shallow depth of field or deep focus to control how the background behaves.
If you are unsure which movement fits a beat, describe what the audience should learn. Push in as she reads the letter tells you the movement is a slow push. Camera language is a translation of intent, not decoration.
Light, Colour and Texture
Lighting prompts fail when they rely on shorthand. Cinematic lighting means nothing specific; single hard key from camera left, deep falloff, cool bounce from wet pavement means a great deal.
Describe three properties: source, quality, and direction. Source is the practical origin — window, streetlamp, neon sign, overcast sky, phone screen. Quality is hard or soft, and it decides shadow edges. Direction is where it lands: backlight for silhouettes, rim light for separation, side light for texture.
Colour is easier to control as a relationship than as a list. Instead of naming a preset, say what contrasts with what: warm tungsten interiors against blue evening exteriors. That single sentence gives you the separation most people are chasing when they type a look-up name.
Texture is the finish pass. Halation, grain, bloom, slight lens breathing, chromatic aberration at the edges: used sparingly, these push a clip from clean render toward photographed image. Used everywhere, they read as a filter.
Motion Is the Part Most Prompts Skip
A video prompt needs two kinds of motion: what the camera does and what the world does. Most people specify the first and forget the second, then wonder why the frame feels frozen inside a moving shot.
Add environmental motion in plain terms. Hair lifting in wind, coat fabric shifting, steam curling off a cup, rain streaking across a window, a crowd shifting in the background. These small movements are what convince a viewer the scene is alive.
Then state speed. Slow and deliberate read differently from quick and urgent, and pacing affects how much the model attempts within the clip duration. A five-second clip comfortably holds one action and one camera move. Ask for three and you get morphing.
Finally, protect the parts that should not move. Background architecture remains static is a legitimate and often necessary instruction when you are placing a subject in a real-feeling location.
What to Leave Out
Negative prompts are useful for artifacts, not for ideas. They work best against a short list of recurring failures: extra fingers, melting faces, warped text, floating limbs, watermark-like smudges, jitter between frames. Keep the list short and specific; a long negative list dilutes attention the same way a long positive prompt does.
What to avoid in the positive prompt is equally important.
- Stacked adjectives. Three mood words do less than one concrete noun.
- Multiple subjects with separate actions. Two characters doing different things in one short clip is a continuity gamble.
- Brand and celebrity names. Describe the qualities you want instead of borrowing a name.
- Contradictions. Different times of day, incompatible lenses, or two camera moves fighting for the same frame.
- Synopses. Story beats belong in your outline, not in a single generation.
Iterating Without Losing the Shot
Random regeneration is the most common time sink. Treat prompting as controlled experimentation instead.
Change one variable per generation. If you adjust the camera move and the lighting at the same time, you learn nothing about which one fixed the frame. Keep a written log of prompt, settings, and result quality — even a three-column note file is enough, and it compounds fast.
Use a reference image when composition matters more than description. Image-to-video gives you the framing you already chose, and the text prompt then handles motion, pacing, and atmosphere. You can build reference stills quickly with an AI image generator before committing to video generation.
Finally, judge results against a checklist rather than a feeling. Does the subject match? Is the camera move the one you asked for? Is the lighting direction consistent across the clip? Are hands and faces stable? A clip can look impressive and still fail all four.
A Workflow You Can Repeat
Sketch the beat sheet first
Write the sequence in plain language before writing any prompts: what the viewer sees, in what order, and what each shot has to accomplish. Three to six beats is enough for a short piece. Only then convert each beat into a prompt using the six slots.
Generate three variants, not one
Run the same prompt with small deliberate variations — a tighter lens, a different time of day, a slower push. Comparing three options teaches you the model's dialect faster than studying documentation. If you want a starting point, browse a prompt library to see how other people structure the same elements.
Lock the look, then vary the action
Once a shot works, treat its lighting, palette, lens, and finish clauses as a fixed block. Reuse that block across the sequence and change only subject and action. This is how you get visual continuity without a custom-trained model.
Assemble on a timeline early
Drop clips into an editor before you have the full set. Continuity problems — mismatched colour temperature, inconsistent motion energy, jarring shot sizes — are obvious in sequence and invisible in isolation. Importing early also shows you which shots are missing, so you can generate with purpose instead of in bulk.
Common Mistakes and What They Look Like
| Symptom | Likely cause | Fix | | --- | --- | | Soft, generic frames | Mood words instead of concrete detail | Add shot size, lens, light direction | | Characters morph mid-clip | Too many actions for the duration | One beat per clip, shorten the ask | | Warped architecture | Movement fighting a detailed setting | Simplify geometry or slow the camera | | Style drifts between clips | No reusable look block | Freeze light, palette, and finish clauses | | Text and signage look wrong | Text is hard to render reliably | Frame around it or add a negative for text |
A subtler mistake is assuming prompts transfer unchanged between tools. Each generator has its own dialect: some weight camera terms heavily, others respond more to subject description. When you switch models, keep the structure and re-test the vocabulary on a short clip.
Reusable Prompt Patterns
Patterns save time because they encode a proven structure. Adapt the details, keep the skeleton.
Product beauty shot. Macro push across brushed aluminium surface, single softbox from top left, gradual falloff to black, fine dust particles catching light, shallow depth of field, subtle rotation, no background detail.
Character entrance. Wide shot, eye level, figure in a long coat walks from background haze toward camera, slow dolly back at matching pace, backlit by cold window light, warm practical on the left edge, 35mm, handheld micro-movement.
Establishing landscape. High aerial glide over a river valley at first light, low mist in the treeline, long shadows, 24mm, deep focus, slow lateral truck, muted greens with a warm horizon band.
Seamless loop. Static locked-off shot, overhead, liquid surface rippling in a slow repeated cycle, even diffuse light, no cuts, symmetrical composition, gentle grain.
Notice that each pattern names a shot size, a movement, a light source, and a finish. Those four elements carry most of the reliability. Everything else is subject matter.
FAQ
How long should an AI video prompt be?
Most models respond best to roughly 30–60 words of dense, specific instruction. Beyond that, extra clauses compete for attention. If you need more control, split the idea into two shots rather than one longer prompt.
Do I need to describe the camera in every prompt?
Yes, if composition matters. Without a stated shot size and angle, generators fall back on a neutral mid-shot. You can leave movement out for genuinely static shots, but say locked-off so the model does not invent drift.
Why does my character change between clips?
Text alone rarely holds identity across separate generations. Use a reference still, keep wardrobe and lighting descriptions identical word for word, and generate shots in one session so settings stay consistent.
Are negative prompts worth using?
They are worth a short, artefact-focused list. They are not a fix for a weak positive prompt. If the frame is wrong conceptually, rewrite the description rather than adding more exclusions.
How many generations should a good shot take?
With a structured prompt, three to six attempts is normal for a shot you will actually use. If you are past ten, the prompt is usually asking for two incompatible things — split the shot instead.
Can I use the same prompt across different tools?
Keep the structure and expect to retune the vocabulary. Camera and lighting terms transfer reasonably well; stylistic shorthand does not. Test one short clip before committing to a whole sequence.
Put the Prompt to Work
Good prompting is a craft of specifics: one subject, one beat, one camera move, one light source, and a finish that holds the look together. Write the shot before you write the sentence, keep a log of what worked, and reuse the parts that survive. Do that consistently and generation stops feeling like a slot machine.
When you are ready to test a prompt against a real render, open the Orelon AI video generator, start from a template if you want a structure to adapt, and build your first sequence shot by shot. More workflow breakdowns live on the Orelon blog.

