Learn how to structure prompts for AI video: shot design, camera language, motion, lighting, character consistency, and a repeatable iteration workflow.
A prompt is not a wish — it is a shot list compressed into a paragraph. Most disappointing AI video output is not caused by a weak model. It is caused by a prompt that describes a mood but never describes a shot. The fix is craft rather than magic words: decide what the camera sees, how it moves, how long it holds, and what must not appear. That is the entire discipline.
This guide is a working method for prompt engineering in AI video. It covers how to structure a prompt, how to write differently for text-to-video and image-to-video, how to hold a character or location together across shots, and how to iterate without burning an afternoon on rerolls. Every section ends in something you can apply to the next clip you generate, whether you work inside Orelon's AI video generator or anywhere else.
Why structure beats length in video prompts
Long prompts do not automatically produce better video. They produce averaged video. When a single prompt contains three lighting setups, two camera moves, and five style adjectives, the model does not prioritize — it blends. The result is a soft, generic clip that technically contains everything you asked for and emotionally contains nothing.
Think of a prompt as a stack of priorities. The order matters, because most systems weight early tokens more heavily than late ones:
- Subject — who or what the shot is about.
- Action — the single physical thing happening.
- Camera — framing, angle, lens, movement.
- Light and palette — the source, direction, and color logic.
- Texture and style — grain, stock, rendering character.
- Constraints — what must not appear.
A useful rule: if you cannot point at a phrase in your prompt and say which of those six slots it fills, delete it. Decorative language feels productive to write and is mostly noise to the model.
There is also a real difference between prompting a still image and prompting motion. Image models reward texture, material, and style vocabulary — they are judging a single frame. Video models reward temporal language: verbs, speed, direction, cause and effect, and how long the shot lasts. A prompt that wins in an image generator often produces a static, drifting clip in a video generator, because nothing in it told the model what changes between the first frame and the last.
The anatomy of a cinematic prompt
Every strong video prompt contains five functional parts. You do not need all five in every shot, but you should know which ones you are leaving out and why.
Subject and action
Cast the scene instead of describing a category. "A ferry captain in his fifties, oilskin slick with rain" gives the model something to build. "A man" gives it a default. Then give exactly one continuous physical action: he coils a wet rope, she lifts a lantern, the car rolls to a stop. One action, not a story arc. A ten-second clip cannot carry a beginning, middle, and end; it can carry one gesture, and the gesture is what makes it feel alive.
Camera and lens
Specify shot size, angle, lens, and movement as separate decisions. "Low-angle medium shot, 35mm, slow dolly in" is actionable. "Cinematic camera work" is not. Two cautions: combine at most one movement with one framing adjustment, because stacked moves (dolly in and orbit and crane up) usually collapse into a wobble; and choose a lens that matches the emotional distance you want. Wide lenses exaggerate space and unease, long lenses compress and isolate.
Light, palette, and texture
Name one dominant light source and its direction. "Single sodium streetlamp from frame left, cold blue ambient fill" gives the model a physics problem it can solve. Name a palette with two or three anchors rather than a mood: "wet asphalt, sodium orange, deep teal shadows" outperforms "moody and beautiful." Texture belongs here too — 16mm grain, anamorphic flare, matte skin, condensation on glass. Texture is what separates an AI clip that reads as generated from one that reads as shot.
Motion, pace, and duration
Motion is not only the camera. It is hair, steam, fabric, water, dust, breath, and background traffic. Name two or three secondary motions and the shot stops feeling like a photograph being pushed around. Pace adjectives do real work here: languid, urgent, unhurried, abrupt. And state duration implicitly by describing how much happens — a slow reveal needs fewer events than a busy handheld shot.
Constraints and exclusions
Tell the model what to avoid. Extra fingers, floating objects, warped faces in the background, on-screen text, watermarks, sudden cuts, morphing. Keep the exclusion list short and specific. A fifteen-item negative list tends to bleed into the positive prompt and make the output timid.
A reusable prompt skeleton
Instead of writing from scratch, keep a skeleton and fill the slots. This is the single biggest time saver in prompt engineering, because it turns a creative act into an editing act.
[shot size + angle], [lens] — [subject with two identifying details]
[one continuous action] while [secondary motion 1] and [secondary motion 2]
[camera movement, single] ; [pace]
lighting: [one source + direction], palette: [2-3 color anchors]
texture: [film stock / grain / material quality]
avoid: [2-4 specific failure modes]
Filled in, it reads like this:
Low-angle medium shot, 35mm — a ferry captain in his fifties, oilskin slick with rain,
coils a heavy wet rope while steam rises from the deck vents and gulls cross the far background
slow dolly in ; unhurried
lighting: single sodium lamp from frame left, cold blue ambient fill
palette: wet asphalt, sodium orange, deep teal shadow
texture: fine 16mm grain, soft halation on highlights
avoid: on-screen text, warped hands, background figures merging
That is roughly ninety words and does more than a five-hundred-word paragraph of atmosphere. Save three or four skeletons — a character portrait, a landscape establishing shot, a product close-up, an action beat — and you will rarely start from a blank page again.
Text-to-video versus image-to-video
These are two different crafts wearing the same name, and confusing them is the most common reason a prompt "stops working" when someone switches modes.
Text-to-video: invent everything, then constrain
When there is no starting frame, you own composition, subject design, light, and motion. Write the full skeleton. Expect the first render to be a blocking draft rather than a finished shot — your job on the first pass is to check that the framing and the action are right, not that the skin is beautiful.
Image-to-video: describe only the difference
The still image already answers "what does this look like." Your prompt should answer "what happens next." Do not re-describe the frame. If you write "a woman in a red coat standing in a snowy street" as the prompt for an image that already shows exactly that, the model may re-render the scene and drift away from the composition you carefully built. Write the delta instead: "she turns her head toward camera, coat hem lifting in the wind, slow push in, falling snow drifting left to right."
This is why a strong pipeline often starts with a locked frame. Generate or select the still, approve it, then animate it — Orelon's AI image generator exists for exactly that handoff. A frame you approved is a constraint; a prompt alone is a suggestion.
Directing motion without a storyboard
Motion is where AI video most often fails, and it fails in a predictable way: the model animates the subject but forgets that the world is physical. A coat should move when the person walks. Steam should rise rather than slide sideways. A hand should stop when it reaches a surface.
Three habits fix most of it:
- Give the camera a motivation. "Push in as she reads the letter" beats "push in." Motivation gives the model a reason for the timing.
- Describe cause and effect, not just states. "He sets down the glass, and the liquid ripples" creates a chain the model can animate. "A glass of water" does not.
- Keep one dominant motion per clip. A dolly, a turn, a fall, a gust — pick one. Two dominant motions compete and the output looks like it is being pulled in two directions.
Holding characters and locations consistent
Consistency is not a model setting; it is a writing discipline. Build a fixed descriptor block for each recurring element and paste it verbatim into every prompt in the sequence. Do not paraphrase it, do not "improve" it between shots.
For a character, lock four things: age and build, one garment with a visible material, one distinguishing feature, and one habitual posture. For a location, lock three anchor objects and the light direction. Then each new shot changes only the shot size, the action, and the camera movement.
A character block might read:
woman, late thirties, cropped dark hair, faded olive field jacket with a torn left cuff,
thin scar through the left eyebrow, shoulders slightly forward
Used identically across six prompts, that block does more for continuity than any amount of retouching. Where a project supports reference frames or saved styles, pair the block with a locked still — text alone drifts, text plus a frame holds.
The iteration workflow: three passes, one variable at a time
Random rerolling is the most expensive habit in AI video. Replace it with a pass structure.
Pass one — blocking
Ignore beauty. Check only framing, subject placement, and whether the action reads. If the composition is wrong, rewrite the shot-size and angle clause; do not add adjectives.
Pass two — performance and camera
Now fix timing and movement. Shorten the action, slow the pace word, or narrow the camera move. Change exactly one clause per attempt so you learn which clause caused the change.
Pass three — polish
Only now touch light, palette, and texture. This is where you add grain, adjust the color anchors, and tighten the exclusion list.
Keep a log while you work: prompt version, what changed, verdict in one line. After twenty clips you will have a personal library of what your tooling actually responds to, which is worth more than any generic list of tips. Browsing a curated prompt library is a fast way to seed that log with structures that already work.
Mistakes that waste renders
| Mistake | What you see | Fix |
|---|---|---|
| Stacked camera moves | Wobbly, drifting frame | One movement per clip |
| Contradictory lighting | Flat, sourceless gray | One source, one direction |
| Emotion words instead of behavior | Stiff, posing subject | Describe the gesture, not the feeling |
| Re-describing an input image | Composition drifts from your frame | Prompt only the change |
| Style adjective pile-up | Muddy, over-processed look | One texture, one stock reference |
| No secondary motion | Clip looks like a sliding photo | Add hair, steam, fabric, dust |
| Wrong aspect framing | Cropped heads, dead space | State orientation and shot size early |
Read the table as a diagnostic: when a clip disappoints, find the row that matches the symptom before you rewrite anything else.
Working efficiently at volume
If you are producing a sequence rather than a single clip, efficiency comes from reuse, not speed. Three practical habits:
- Freeze the skeleton. One template per shot type, with only the variable slots changing.
- Approve frames before animating. Stills are cheaper to iterate than motion, and a locked frame removes half the uncertainty from the video prompt.
- Define "good enough" before you generate. Write two acceptance criteria per shot — for example "action reads clearly" and "light direction consistent with shot one." When both are true, stop, even if the clip is not perfect. Sequences are built from consistent shots, not exceptional ones.
Starting from a pre-built structure also shortens the blank-page phase; Orelon's video templates are organized around exactly this kind of reusable shot pattern.
FAQ
How long should a video prompt be? Usually 60 to 120 words. Long enough to cover subject, action, camera, light, texture, and constraints; short enough that no clause contradicts another. If you pass 150 words, check whether you have described two different shots.
Do negative prompts actually work? Yes, in moderation. Two to four specific exclusions (on-screen text, warped hands, morphing faces, extra limbs) help. Long negative lists tend to suppress detail across the whole frame.
Why do faces drift over a longer clip? Because identity is being re-inferred every frame from a short description. Fix it with a verbatim character block, a locked reference frame, shorter clips that you extend, or framing that keeps the face at a consistent scale.
Should I write prompts in English? English usually has the most training coverage, but the more useful rule is consistency: keep one language per project so your fixed descriptor blocks stay byte-identical between shots.
How many rerolls before I change the prompt? Two. If two attempts with the same prompt miss, the prompt is the problem, not the seed. Change one clause and test again.
Can I reuse a prompt across different video tools? The structure transfers well; the vocabulary does not always. Motion phrasing, negative-prompt syntax, and duration handling vary. Keep your skeleton and adjust the dialect — a comparison of AI video generator alternatives is useful when you need to know which conventions a given tool expects.
Write the shot, not the mood
Prompt engineering for video is not a hunt for secret words. It is the ordinary craft of shot design moved into a text field: choose a subject, give it one action, place a camera, light it once, add physical motion, and exclude what ruins the frame. Do that consistently and the model stops feeling unpredictable, because you have finally given it something specific to be predictable about.
The fastest way to internalize this is to write three prompts with the skeleton above and render them back to back. Start in Orelon's AI video generator, pull a structure from the prompt library if you want a head start, and keep the version that reads as a shot rather than a slideshow. Cinematic ideas in motion begin as precise sentences.