Learn how to craft AI video prompts that control camera, light, motion, and pacing, with iteration workflows, worked examples, and common mistakes to avoid.
A prompt is not a wish. It is a shot list compressed into a paragraph, and the model treats it that way: every clause is a constraint, and every missing clause is a decision handed back to the model's training data. That single idea explains most of the gap between people who get usable AI footage on the second try and people who generate twenty clips and keep none of them.
This guide is a practical method for writing prompts that direct rather than suggest. It covers how the model reads your words, how to structure a prompt in layers, which camera and lighting vocabulary actually changes output, how to sequence clips into a scene, and how to iterate without losing track of what worked. Every section ends in something you can apply to the next clip you generate.
Start With the Shot, Not the Sentence
Most weak prompts are written as sentences. Strong prompts are written as decisions. Before typing anything, answer five questions:
- Subject: who or what is on screen, described specifically enough to be visualized (age range, wardrobe, material, species, object condition).
- Action: one verb, one direction, one pace. "Steps off a curb" beats "is walking around."
- Environment: location, time of day, weather, and what is behind the subject.
- Camera: framing, angle, lens feel, and movement — or deliberate stillness.
- Light and palette: source, direction, hardness, and two or three dominant colors.
Compare these two prompts:
A beautiful cinematic video of a woman walking in a city, stunning, 4k, masterpiece.
Wide 24mm shot, low angle: a woman in a mustard raincoat steps off a curb into a shallow puddle, neon signage reflected in the water, overcast dusk, slow dolly-in.
The first prompt contains no action verb the model can animate, no camera, no light, and no time of day. Words like "beautiful" and "stunning" carry almost no visual information — they describe your reaction, not the frame. The second prompt is longer by a dozen words and resolves a dozen ambiguities. Note that it also picks one motion (dolly-in) instead of stacking three.
A useful habit: write the prompt, then read it back and ask which words could be removed without changing the image. Those words are noise. Then ask which visible detail is still undecided. Those gaps are what the model will fill with something generic.
What the Model Actually Reads
AI video generation is usually a two-stage system, and knowing which stage handles what changes how you write.
The language layer
A language model parses your prompt into concepts and relationships: subject, attributes, spatial prepositions, and implied motion. This stage is where specificity pays off. "A brass compass resting on wet slate" resolves into distinct objects with material properties. "A cool object" resolves into almost anything.
The language layer is also where contradictions get silently resolved rather than flagged. If you ask for a locked-off static shot with handheld energy, the system picks one and drops the other, and you will not be told which.
The pixel layer
A diffusion-style generator then synthesizes frames, and a motion component keeps them consistent across time. Motion quality depends heavily on how familiar the requested action is. Common actions — walking, turning a head, water flowing, smoke rising, hair moving — have strong learned priors and tend to look convincing. Rare or compound actions, such as a person juggling while descending a ladder, need more descriptive support and more attempts.
What training data means for your prompt
Every model has seen more of some things than others. Cinematic sunsets, neon city streets, and slow-motion smoke are overrepresented, which is why they look effortless and also why they look identical across projects. Conversely, specific industrial equipment, regional clothing, and niche architecture may need explicit description — materials, colors, proportions — because the model has thin coverage.
Practical consequence: when a subject looks wrong, do not just add adjectives. Break the subject into visible parts. Instead of "traditional Polish folk costume," describe the embroidered white blouse, red ribbon, and floral skirt, then let the model assemble it.
The Prompt Stack: Six Layers That Do the Work
A reliable prompt has a repeatable order. Constraints first, polish last.
| Layer | Question it answers | Example phrase |
|---|---|---|
| 1. Subject | Who or what is on screen? | a courier in a navy jacket, mid-thirties |
| 2. Action | What changes during the clip? | sets down a crate and straightens up |
| 3. Environment | Where and when? | narrow alley, early morning fog |
| 4. Camera | How is it framed and moved? | medium shot, 50mm, slow tracking right |
| 5. Light and color | What illuminates it, in what palette? | soft side light, cool blue with warm sodium accents |
| 6. Style and technical | What medium and format? | documentary realism, 16:9, subtle grain |
Two rules make this stack work. First, put the most important constraint earliest; many models weight the opening tokens more heavily. Second, keep each layer to one idea. "Medium shot, 50mm, slow tracking right" is one camera decision. "Medium shot, wide angle, handheld, locked off, orbiting" is four conflicting decisions.
You can reuse the stack across an entire project by keeping layers 3, 5, and 6 fixed and only changing subject and camera. That is how a set of clips starts to feel like one film instead of a moodboard.
For teams producing volume, saving the stack as a reusable template — subject blank, everything else pre-filled — cuts iteration time dramatically. Browsing a prompt library of layered examples is a fast way to internalize the ordering before writing your own.
Camera, Light, and Motion Vocabulary That Changes Output
Abstract praise does nothing. Technical vocabulary does, as long as it is consistent.
Framing and lens
- Extreme wide / wide: establishes geography; subject small in frame.
- Medium: waist-up; the workhorse for dialogue and action clarity.
- Close-up / extreme close-up: emotion or detail; risky for hands and faces in fast motion.
- 24–28mm feel: exaggerated perspective, environment-heavy.
- 50mm feel: neutral, close to human perception.
- 85mm+ feel: compressed background, subject isolation.
Angle and height
Low angle increases scale and threat. Eye level reads neutral. Overhead flattens and turns people into shapes, which is useful for patterns and crowds.
Movement
Name one primary move: dolly in, dolly out, tracking left or right, crane up, orbit, handheld drift, or locked-off. Add speed only if it matters — "slow" and "fast" are usually enough. Handheld plus locked-off is a contradiction; a zoom plus a dolly is a specific stylistic choice that most models handle poorly, so use it deliberately or not at all.
Light
Describe source and direction before mood. "Backlit by a low sun through haze" tells the model where the light comes from. "Cinematic lighting" does not. Useful shorthand: soft window light, hard noon sun, overcast diffusion, practical neon, single bare bulb, bounced fill.
Color
Two or three dominant colors plus one accent outperform long palettes. "Desaturated teal shadows, warm amber highlights" gives the grader something to hold onto.
Style
Style words should describe a medium or technique, not a compliment: 35mm film grain, documentary handheld, stop-motion, architectural render, watercolor. Stacking five style words usually produces mush, because they compete for the same pixels.
Negative Prompting: Subtracting the Noise
Negative prompts are supported in different ways across tools, so treat this section as conditional. Where they exist, they are most useful for artifacts and unwanted overlays rather than for creativity.
Good candidates for the negative field:
- text, captions, watermarks, logos
- extra limbs, warped hands, duplicated faces
- flicker, frame tearing, morphing backgrounds
- jump cuts, strobing
- split screen, borders, collage layouts
Poor candidates: "ugly," "bad quality," "amateur." These are vague and can flatten the image along with the problem. Also avoid negating something you explicitly asked for — listing "no camera movement" while your positive prompt requests a dolly will produce inconsistent results across attempts.
Keep the negative list short and stable across a project. If you keep adding to it after every bad clip, you lose the ability to tell which change fixed what.
From Single Clip to Scene: Sequencing Prompts
A scene is not one long prompt. It is four to six short clips that share continuity anchors. Build the sequence in a beat sheet before generating anything:
- Establish — where are we, what time, what mood.
- Introduce — the subject enters or is revealed.
- Complicate — the action that creates tension or interest.
- Payoff — the reaction or result.
- Exit — a closing frame that resolves the movement.
Each beat becomes one prompt that repeats three anchors unchanged: wardrobe or object description, light and palette, and lens family. Only subject position, action, and camera move. This is the cheapest way to get visual consistency without reference images.
Where the tool supports image-to-video, generate a hero still first and animate from it. A still is faster and cheaper to iterate than video, and it locks composition before motion introduces new variables. The AI image generator can serve as that pre-visualization step, and the resulting frame often becomes the first frame of the clip.
Duration planning matters more than people expect. Most clips read best at three to five seconds when the action is simple. If a beat needs eight seconds of story, split it into two prompts rather than stretching one. Transitions — cuts, match cuts, wipes — belong in the edit, not in the prompt, unless the tool explicitly supports transition generation.
The Iteration Loop: Test, Compare, Lock
Prompting is a loop, not a lottery. The loop has four steps and one rule: change one variable at a time.
- Baseline. Generate one clip from the full prompt. Do not judge it yet; label it.
- Diagnose. Name the single biggest problem — composition, motion, lighting, or subject fidelity.
- Adjust. Change the layer responsible. Composition problems are camera fixes. Motion problems are action or duration fixes. Lighting problems are source and direction fixes.
- Lock. When a clip works, freeze its prompt in a document and note what made it work.
Keep a prompt log with columns for clip ID, full prompt, what changed, and verdict. After ten clips you will have a usable record of which phrases your chosen model responds to, which is more valuable than any generic list of magic words.
Comparison discipline helps too. Generate two variants with a single difference and watch them side by side, not from memory. Memory is generous; it remembers the version you intended rather than the version you got.
If you are still choosing a tool, evaluate it on prompt responsiveness rather than feature lists: does the same prompt produce consistent results across attempts, does camera language land, does the model respect aspect ratio and duration? Side-by-side comparisons such as Orelon vs Runway are useful because the differences usually show up in how literally a prompt is interpreted.
Common Mistakes and How to Fix Them
| Mistake | Why it hurts | Fix |
|---|---|---|
| Adjective pile-up ("stunning, epic, masterpiece, 8k") | Adds no visual constraint; may trigger clichéd output | Replace each adjective with a visible detail |
| Two actions in one clip | Motion model splits attention, produces mush | One verb per clip; split into two prompts |
| Contradictory camera instructions | Model silently drops one | Choose one framing and one movement |
| No environment or time of day | Model defaults to generic daytime | State location, hour, weather |
| Neglecting aspect ratio | Composition breaks in the edit | Set it before generating, not after |
| 150-word prompts with five style references | Competing constraints cancel out | Trim to the six-layer stack, one style |
| Judging on a single attempt | Variance is normal | Generate three, then compare |
One more mistake is worth calling out because it is invisible: prompts written for the wrong stage. Asking a text-to-video model to fix a composition problem that a still generator could solve in one attempt wastes time. Match the tool to the problem.
A Worked Example: A 30-Second Cinematic Teaser
Suppose the brief is a teaser for a fictional documentary about night-shift workers. Four clips, twenty-eight seconds, consistent palette.
Anchors (repeated in every prompt): documentary realism, 2.39:1 framing, cool teal shadows with warm sodium highlights, 35mm grain, 40mm lens feel.
Shot 1 — Establish. Wide 28mm shot from across an empty street: a lit laundromat window at 2 a.m., one figure folding clothes inside, wet asphalt reflecting the sign, locked-off camera, slow ambient motion only.
Shot 2 — Introduce. Medium 40mm shot at eye level: a woman in a grey work shirt loads a tray of cups into an industrial dishwasher, steam rising, hard overhead fluorescent light, slight handheld drift.
Shot 3 — Complicate. Close-up 85mm shot: her hands pause on a wet cup, water dripping, sodium street light through the window behind her, shallow depth of field, no camera movement.
Shot 4 — Exit. Wide 28mm shot from behind: she steps out of the laundromat into rain, pulls her hood up, walks out of frame right, camera locked, streetlight flares in the lens.
The variable that usually needs the most retries is shot 3 — hands plus water plus shallow focus is a hard combination. If it fails twice, simplify: replace hands with a face, or remove the drips and keep the pause. Do not add more adjectives; remove competing constraints.
The sequence also illustrates a useful budget rule: spend prompt complexity where the story is, and keep transition shots boring. A clean, simple establishing clip is more valuable than an ambitious one that never resolves.
FAQ
How long should an AI video prompt be?
Most clips work best at 30–60 words across the six layers. Under 20 words leaves too much to chance; over 80 often introduces contradictions. Length is not the goal — coverage of the six decisions is.
Do camera terms like dolly and 85mm actually matter?
Yes, when used consistently. Models respond more reliably to framing family and movement direction than to exact focal lengths, so treat lens numbers as guidance for perspective rather than literal math.
Why does the same prompt give different results?
Because generation involves randomness. Set a seed where the tool allows it, then change one variable at a time. Three attempts per prompt is a reasonable baseline before concluding the prompt is wrong.
Should I write prompts in my own language?
Prompt adherence is usually strongest in the language the model was primarily trained on, which for most tools is English. If your project language differs, write the prompt in English and keep the script and edit in your own language.
How do I keep characters consistent across clips?
Repeat an exact wardrobe and feature phrase in every prompt, keep the same lens and palette wording, and use image-to-video from a fixed reference frame where supported. Consistency comes from identical text, not from similar text.
Can one prompt generate a full scene?
Rarely well. Models handle one beat per generation. A scene is built in the edit from several short clips, each with a single action and a clear camera decision.
Direct Your Next Clip on Orelon
The prompt is the direction. Decide the subject, the action, the environment, the camera, the light, and the style before you generate — then change one thing at a time until the clip matches the frame in your head. Everything else is iteration discipline.
Orelon is built for that workflow: an AI video generator for cinematic ideas in motion, where layered prompts translate into controlled camera movement, deliberate lighting, and footage that cuts together. Start with the AI video generator, borrow a starting structure from video templates, and browse the Orelon blog for more breakdowns of shot design, pacing, and post-production. Your first usable clip is one well-structured prompt away.

