Learn to write AI video prompts that hold together: shot structure, camera language, lighting, motion, character consistency, and a repeatable iteration loop.
Most disappointing AI video clips are not model failures. They are brief failures. A prompt like "a woman walking through a neon city at night, cinematic" hands the generator a subject and a mood, then leaves it to invent the shot — and it will invent the most average shot in its training data. Framing, lens, movement, pacing, and light direction all get defaulted. The result reads as generically "AI" not because the model is weak, but because nobody made a decision.
Prompt engineering for video is the discipline of making those decisions explicitly and in the right order. It is closer to writing a shot brief for a camera operator than to writing a caption for a photo. This guide covers what each part of a prompt actually controls, how to build reusable skeletons, how to keep a character recognizable across clips, how to debug a generation that missed, and how to turn single shots into a sequence that edits together.
Start with the shot, not the story
Text-to-video models generate clips, typically a few seconds long. They do not generate scenes. If your prompt describes a plot — "a detective discovers the truth and decides to run" — the model has to compress an arc into eight seconds and will usually produce a vague person doing a vague thing.
Write the unit of work as one shot:
- Who or what is on screen, described concretely enough to visualize.
- What happens in this clip only — one action, one beat.
- Where it happens, with enough environment detail to anchor depth.
- How it is filmed — framing, lens, camera movement.
- How it is lit — direction, quality, color.
A useful test: can a stranger sketch a storyboard frame from your prompt? If not, the model cannot either.
The single-beat rule
One clip, one beat. "She turns from the window, picks up the phone, and walks out" is three beats and will produce either a rushed blur or a static shot where nothing completes. Split it into three generations and cut them together. You get more control, and you can reshoot one beat without losing the other two.
Deconstructing a video prompt into layers
Every strong video prompt contains the same layers, even if they are written as one flowing sentence. Naming them separately makes them editable.
| Layer | What it controls | Weak version | Strong version |
|---|---|---|---|
| Subject | Identity, wardrobe, expression | "a woman" | "a woman in her thirties, cropped black hair, olive field jacket" |
| Action | What changes in the clip | "doing something" | "lifts a brass kettle and pours" |
| Environment | Depth, texture, background | "a kitchen" | "a narrow tiled kitchen, morning light through a half-open window" |
| Camera | Framing and movement | "cinematic angle" | "medium close-up, 50mm, slow push in" |
| Light | Direction, quality, color | "dramatic lighting" | "single hard window key from camera left, deep shadow on the right" |
| Grade | Overall color and texture | "film look" | "muted warm highlights, slightly lifted blacks, fine grain" |
| Pace | Speed of action and motion | "energetic" | "steady, unhurried motion, no camera shake" |
Most people write the first three layers and stop. The last four are where the visible quality lives.
Order matters more than length
Models weight the beginning of a prompt more heavily. Put subject and action first, then environment, then camera, then light and grade. If you bury the subject after four clauses of atmosphere, you will often get a beautiful empty room.
Length is not quality. A 45-word prompt that makes seven decisions beats a 120-word prompt that repeats the word "cinematic" four times and contradicts itself on lighting.
A reusable prompt skeleton
Here is a skeleton you can keep in a notes file and adapt per shot:
[subject + wardrobe detail], [single action in present tense],
[environment with one depth cue],
[camera framing], [lens], [camera movement],
[light direction and quality], [grade and texture],
[pace note], [negative: what should not appear]
Example: a cinematic product shot
A matte black mechanical watch resting on wet slate, a single drop of water runs down the crystal, shallow tide pools blur in the background, extreme close-up, 100mm macro, slow lateral slider move left to right, hard key light from behind at a low angle creating a bright rim on the bezel, cool blue shadows with warm specular highlights, fine grain, slow steady motion, no text, no logos, no hands.
Notice what is absent: no adjectives about the brand, no story, no emotion words. Every clause is either a visible object, a camera decision, or a light decision. Models handle that well.
Example: a dialogue-adjacent two-shot
Two people at a formica diner table, the man in a grey hoodie on the left, the woman in a red corduroy jacket on the right, she slides a folded napkin across the table toward him, he does not reach for it, static medium two-shot from across the booth, 35mm, slight handheld float, overhead fluorescent light mixed with cold blue window light from the right, desaturated grade with warm skin tones, natural conversational pace, no on-screen text.
This is the kind of prompt that produces something you can actually cut into a scene, because the composition is locked and the lighting source is named.
Camera and motion language that models understand
Camera vocabulary is the highest-leverage part of a video prompt, and the most commonly wasted. "Dynamic camera" means nothing. Specific terms mean something.
Framing terms that work: extreme wide, wide, full shot, medium shot, medium close-up, close-up, extreme close-up, over-the-shoulder, two-shot, insert, low angle, high angle, overhead, Dutch angle, point of view.
Movement terms that work: locked-off static, slow push in, pull back, dolly left, truck right, crane up, tilt down, orbit around the subject, handheld follow, whip pan, rack focus from foreground to background, crash zoom, aerial descend, drone rise.
Speed modifiers that work: slow motion, real-time, speed ramp, freeze at the end, steady and unhurried, quick and snappy.
Two rules make these reliable:
- Attach the movement to something. "The camera orbits the subject clockwise at chest height" is clearer than "orbit shot."
- Do not stack contradictory moves. "Push in while pulling back with a whip pan" gives you mush. Pick one primary move per clip and, if you want, one small secondary one.
If the generated clip has good framing but wrong motion, regenerate with the same prompt and a stricter movement clause rather than rewriting the whole thing. Motion is one of the easiest variables to isolate.
Lighting, lens, and grade vocabulary
Lighting is the second highest-leverage layer, and it is where most prompts collapse into vague words like "cinematic lighting." Replace mood words with physical descriptions.
Direction: from camera left, from camera right, backlit, top light, underlit, side key, rim light, bounce from below.
Quality: hard single source, soft diffused, overcast ambient, dappled through leaves, flickering practical, mixed sources.
Time and color: golden hour, blue hour, harsh midday, sodium streetlight, cool moonlight, warm tungsten interior.
Contrast: high contrast with crushed shadows, low contrast and flat, chiaroscuro, high-key and airy.
Lens language changes how space feels. A 24mm wide lens exaggerates depth and makes rooms feel bigger; an 85mm flattens a face into a portrait; a macro lens reveals texture; an anamorphic look adds oval bokeh and horizontal flares. Naming a focal length is a fast way to get a consistent visual identity across a whole project.
Grade language is the finishing layer: teal shadows with warm skin, bleach-bypass contrast, muted pastel, film print emulation, lifted blacks, fine 35mm grain, clean digital sharpness. Choose one grade phrase per project and reuse it in every prompt — it does more for sequence cohesion than any other single token.
Text-to-video versus image-to-video prompting
The biggest mistake with image-to-video is re-describing the still. The model already sees the frame. What it needs is motion.
When starting from an image, write prompts about change:
- What moves first, and in which direction.
- What the camera does (or that it stays locked).
- What enters or leaves the frame.
- How light shifts, if at all.
From the still: the camera holds static, the character's jacket fabric shifts in a light wind, hair moves left to right, background pedestrians walk past out of focus, no change to framing or lighting.
That prompt will hold the composition, which is exactly what you want when the still was carefully designed. Text-to-video gives you more freedom; image-to-video gives you control. Most professional workflows start by generating a strong keyframe with an AI image generator, then animating it.
Keeping characters and locations consistent across clips
Consistency is the hardest problem in AI video, and it is solved with reference discipline rather than clever adjectives.
- Build a character sheet first. Generate 4–6 images of the same person from different angles before you animate anything. Pick the one that reads best and use it as the reference for every clip.
- Name fixed traits once, then repeat them verbatim. Same wording, every prompt: "red corduroy jacket, silver hoop earring left ear, chin-length black hair." Paraphrasing erodes identity between shots.
- Change one variable at a time. Keep the character block identical and vary only the action and camera lines.
- Anchor locations the same way. "Narrow tiled kitchen, green cabinets, window camera left" reused across five shots gives you a room the audience believes in.
- Use negatives for drift. If a hairstyle keeps changing, add it to a negative clause. If a jacket changes color, add "no color change to wardrobe."
A practical shortcut: write the character block and the location block once, save them as reusable text, and paste them into every prompt in the sequence. Boring, but it works.
Building a shot list before you generate
Generating clips one at a time and hoping they edit together is how projects stall. Plan coverage like an editor:
- Establishing wide — where are we?
- Medium — who is here and what are they doing?
- Close-up — what does the moment feel like?
- Insert — what object matters?
- Reaction — how does the other person respond?
Five prompt slots per scene is usually enough for a 20–30 second sequence. Keep light direction consistent across all five: if the key light comes from camera left in the wide, it must come from camera left in the close-up. If a character exits frame right, they should enter the next shot from frame left.
You can speed this up by starting from an existing structure rather than a blank prompt box — video templates are useful as scaffolding even when the final content is entirely yours.
The iteration loop: how to debug a bad generation
Treat each generation as a test with one hypothesis. When a clip fails, classify the failure before rewriting.
| Symptom | Likely cause | Fix |
|---|---|---|
| Nothing happens | Action described as a state, not a verb | Rewrite as a present-tense action |
| Framing is generic | No camera or lens line | Add framing plus focal length |
| Subject drifts or morphs | Too many simultaneous actions | Reduce to one beat |
| Flat, washed-out look | No light direction or contrast term | Add direction and contrast |
| Motion is chaotic | Contradictory movement clauses | One primary camera move only |
| Wrong subject entirely | Subject buried late in the prompt | Move subject to the first clause |
| Clips do not cut together | Grade phrase varies per shot | Lock one grade and reuse it |
Then follow a fixed order of operations: fix composition first, then motion, then light, then grade, then detail. Do not re-roll twenty times hoping for luck. Each re-roll should change exactly one line.
Keep a simple log — a spreadsheet with columns for shot name, prompt version, seed if available, and verdict — and save every keeper with a versioned filename. Within a week, that log becomes a personal prompt library far more useful than any generic list, because it encodes what your specific subject matter needs.
Common mistakes worth fixing today
Adjective stacking. "Cinematic, epic, stunning, ultra-detailed, 8K, masterpiece" adds nothing measurable. Replace with one concrete camera and one concrete light decision.
Unknown proper nouns. Models do not reliably know brand names, minor celebrities, or invented terminology. Describe the visual instead.
Ignoring aspect ratio. Vertical social clips and widescreen narrative clips need different framing. Decide before you generate, not in the edit.
Expecting legible text. On-screen typography and signage remain unreliable. Add graphics in post.
Overwriting the negative prompt. A negative list of thirty items starts suppressing legitimate detail. Keep negatives to three or four specific drift problems.
Skipping the keyframe. If a shot matters, generate the still first, approve it, then animate. It is faster than three failed text-to-video attempts.
For a broader sense of how different engines respond to the same prompt, it helps to compare outputs side by side rather than switching tools on instinct. Browsing curated examples and alternatives makes those differences obvious quickly.
Frequently asked questions
How long should an AI video prompt be? Long enough to cover subject, action, environment, camera, light, and grade — usually 30 to 60 words. Beyond that, you are usually repeating yourself.
Do negative prompts actually work? Yes, but narrowly. They are effective for specific recurring problems like flickering, extra limbs, text overlays, or wardrobe changes. They are not a substitute for a clear positive description.
Why does my character's face change between clips? Because identity is being re-invented each time. Use a reference image, repeat the exact same descriptive block, and keep the wardrobe wording identical across prompts.
How many generations should a shot take? For a simple product or landscape shot, two to five. For a character performance with specific motion, ten or more is normal — which is why planning coverage in advance saves so much time.
Can I get an exact, repeatable camera move? You can get close, not exact. Name one primary movement, describe its speed, and describe its subject. If precision matters, generate a locked-off version and add the move in post for perfect repeatability.
Should I write prompts in my own language? Write in the language you think in, then test whether the model responds better to English camera vocabulary. Many filmmakers keep a hybrid: narrative description in their own language, technical terms in English.
Where to go from here
The method is simple even though the craft takes practice: decide the shot, write it in layers, keep the character and grade blocks fixed, change one variable per attempt, and log what worked. Do that for a dozen shots and you will have a personal prompt system that produces sequences, not just clips.
Orelon is built for exactly that rhythm — an AI video generator for cinematic ideas in motion, where a strong keyframe, a clear prompt, and a consistent grade come together into something that cuts. When you are ready to shoot your next sequence, start creating in Orelon, and keep the prompt library open beside you while you work.