Prompt Writing for AI Video: Notes Into Cinematic Shots

A video model does not read your intentions. It reads your words, in order, and resolves them into pixels. Two people can describe the same scene and get completely different footage, because one wrote a mood and the other wrote a shot. That gap is not talent. It is craft, and it is learnable in an afternoon.

This guide is about that craft: how to write prompts that behave like a director's instructions instead of a wish list. You will get the structure behind clear video prompts, a worked example that turns a story beat into a shot, a workflow for producing many shots without losing continuity, and the mistakes that quietly ruin otherwise good footage.

If you want to follow along in a tool while you read, keep the AI video generator open in a second tab. The exercises are short enough to try in real time.

Why prompt writing decides what your video looks like

Most disappointing AI video is not a model failure. It is a specification failure. The prompt asked for "a beautiful cinematic scene of a woman in a city at night" and the model delivered something technically correct and emotionally empty, because nothing in that sentence told it where the camera stands, what the light is doing, or what changes between the first frame and the last.

Professional shot language solves this. A camera department does not receive a vibe; it receives a shot size, a lens, a movement, a lighting plan, and a duration. The same specificity works on a model. When you write:

Wide shot, slow dolly forward, rain-slick street, sodium streetlamps flaring in puddles, a woman in a grey coat stops walking, holds still, looks up at a lit window.

the model has decisions to make — and constraints to obey. The result may still surprise you, but it will be a surprise inside your frame rather than outside it.

There is a second reason specificity matters. Video generation is expensive in time. A vague prompt usually costs you four or five renders before you admit the idea was underspecified. A structured prompt costs you ninety seconds of writing and often lands on the first or second attempt. Once you internalize the structure, your iteration speed roughly triples, and speed is what makes a multi-shot sequence feasible at all.

What pen instructions actually mean to a video model

Think of a prompt as a stack of specifications, layered from general to specific. Models tend to weight the opening of a prompt most heavily, so the first clause should establish the frame, not the mood. Everything after that refines.

A useful mental model is the shot card that a first assistant director hands out on set. It answers five questions in a fixed order: what are we looking at, where is it happening, how is it photographed, what moves, and how is it finished. When any one of those is missing, the model improvises — and improvisation is where continuity dies.

Two mechanics are worth knowing because they change how you write.

Negations are unreliable. "No people, no cars, no text" often summons exactly the thing you excluded, because the tokens are still present in the conditioning. Describe the positive state instead: "an empty street at dawn, bare asphalt, closed shutters." The absence is implied by what you filled the frame with.

Adjectives compound weakly; nouns compound strongly. "Beautiful, stunning, epic, cinematic" adds almost nothing — every prompt contains those words, so they carry no information. "Anamorphic flare," "handheld 24mm," "overcast daylight through frosted glass," and "35mm film grain" are specific enough to steer the render. Replace three adjectives with one concrete noun and watch the output tighten.

A third mechanic is timing. Most models generate a fixed short duration, so a prompt should describe a beat, not an arc. "She crosses the room, picks up the letter, reads it, cries, and leaves" is four shots crammed into one prompt. Ask for one action and let the edit do the rest.

The five blocks of a prompt that renders correctly

You do not need to write in this order every time, but you should check that every block is present before you render. Most failed prompts are missing two or more.

Subject and action

Name the subject with enough specificity to be castable: age range, wardrobe, distinguishing detail, and what they are doing at this instant. "A woman" is a placeholder. "A woman in her late thirties, damp hair, olive raincoat, standing still" is a casting decision. State one action, in present tense, with a clear beginning and end.

Setting and light

The location carries more visual information than any other block. Give the space, the time of day, and the source of light. "A narrow kitchen, late afternoon, hard sunlight through a half-open blind cutting across the counter" tells the model where shadows fall. Light direction is the single highest-leverage detail you can add to a scene that looks flat.

Camera and lens

Shot size, angle, lens character, and movement. "Medium close-up, eye level, 50mm, locked off" and "wide, low angle, 18mm, slow push in" produce entirely different films from the same subject. If you only add one thing to your prompts this week, add shot size and movement.

Motion and timing

What moves, in what direction, at what pace. Movement can live in the camera, the subject, or the environment — wind, rain, traffic, drifting smoke. Specify a pace word: slow, deliberate, sudden, unhurried. Without it, models default to a generic medium drift that reads as stock footage.

Style and grade

Finish the image: color palette, contrast, texture, and any reference to a photographic tradition rather than a named artist. "Muted teal shadows, warm practicals, soft halation, fine grain" is more controllable than a director's name and far less likely to produce a pastiche.

A worked example: from a story beat to a shot prompt

The story beat: a woman discovers a letter was written by the father she never met.

The first instinct is to write the emotion.

A woman is shocked and emotional as she realizes the letter is from her father.

That prompt gives the model nothing to photograph. Shock is not visible; it is inferred from behavior, framing, and timing. So convert the interior state into observable evidence.

Close-up, eye level, 85mm, shallow focus. A woman in her late thirties holds a folded letter at chest height in a dim hallway. She reads, stops, lowers the paper slightly, and her jaw tightens. Warm light from a doorway behind her rims her shoulder; the rest of the frame falls into soft shadow. Slow, minimal camera drift.

Same beat, but now every department has a job: casting, lighting, camera, performance. If you want the moment to land harder, do not add more adjectives — split it into two shots and cut between them.

Shot A: extreme close-up on the letter, fingers trembling slightly, shallow depth of field. Shot B: medium shot from behind, her shoulders rising once in a slow breath, hallway dark, doorway light behind her.

The cut does the emotional work, exactly as it would on a real set. This is the core insight of prompt writing: you are not describing a feeling, you are designing the evidence from which a viewer infers the feeling.

From script to shot list: a workflow that scales

One good prompt is a demo. Twenty consistent prompts are a film. Here is a workflow that survives contact with a real project.

Draft in plain language first

Write the sequence as prose with no formatting rules, the way you would explain it to a friend. Do not think about camera language yet. Getting the story straight before the terminology prevents the very common failure of beautifully specified shots that add up to nothing.

Convert each beat into a shot card

Give every shot a short ID (S01, S02, S03) and fill in the five blocks. Keep a single spreadsheet or document with columns for shot ID, subject, setting, camera, motion, style, and status. This becomes your continuity ledger, and it is the difference between a sequence that feels intentional and one that feels assembled from strangers.

Change one variable at a time

When a render misses, resist the urge to rewrite everything. Change the camera block, re-render, compare. Change the light, re-render, compare. Systematic iteration teaches you what the model actually responds to, and that knowledge carries across projects. Random re-rolling teaches you nothing and burns an afternoon.

Batch by location and light

Render all shots that share a set and a time of day in one sitting. Prompts written back to back in the same context tend to drift toward each other, which is usually what you want for continuity. It also makes mismatches obvious immediately rather than three days later in the edit.

Build a reusable prompt library

When a prompt works, save it with a note about why. Over time you accumulate a personal vocabulary of light, lens, and movement phrases that reliably produce the look you want. Starting from a curated prompt library shortens the distance between idea and first render considerably, and it stops you from re-solving the same problems on every project.

Keeping characters and locations consistent across shots

Consistency is the hardest part of AI video and the part most guides skip. A model has no memory of the last shot unless you give it one.

Anchor your characters in words. Write a fixed character block — age, hair, build, one wardrobe item, one distinguishing feature — and paste it verbatim into every prompt that includes that person. Do not paraphrase it. "Olive raincoat, short dark hair, silver ring on right hand" repeated exactly will hold far better than three variations on the same description.

Anchor your locations the same way. A location block should include architecture, dominant color, light source, and one recurring object. If a hallway has a broken wall sconce in shot one, mention the broken wall sconce in shot seven.

Use reference images when identity matters. A still frame as a visual anchor does more for facial consistency than a paragraph of description. Generate a character still first with the AI image generator, then use it as your anchor across the sequence.

Track continuity in writing. Note which side of frame a character exits, what they are holding, and what the light is doing. If she exits left in a wide shot, she should enter right in the reverse. These small disciplines are what make an AI sequence read as directed rather than generated.

Common prompt mistakes and how to fix them

Adjective soup. Six mood words and no camera. Fix: one subject, one action, one shot size, one light source.

Contradictions. "Soft natural light, harsh neon, golden hour, midnight." Fix: choose one lighting plan per shot. Contradictions do not average out; they produce mush.

Multiple actions in one prompt. Fix: one beat per render, joined in the edit.

Camera confusion. "Drone shot in a small room, handheld locked-off." Fix: pick a single camera logic and stay inside it for the duration of the shot.

Crowded frames. Five characters and a busy street means the model will lose someone's face. Fix: stage two subjects maximum, and put the rest in the background as silhouettes.

Expecting reliable on-screen text. Legible signage and titles are still fragile. Fix: generate the plate clean and add typography in post, where you have full control.

Ignoring aspect ratio. A vertical prompt judged against a widescreen reference will always disappoint. Fix: decide the delivery format before you write a single prompt, and use video templates sized to that format.

Re-rolling instead of revising. Ten random attempts teach you nothing. Fix: three deliberate edits, each targeting a named block.

Prompt-only vs reference-image workflows

Both approaches are legitimate, and the right choice depends on what has to stay stable.

Use a prompt-only workflow when the project is about atmosphere, montage, or abstract imagery — title sequences, music visuals, dream sequences, B-roll. Writing speed is high and continuity requirements are low, so a well-structured prompt per shot is enough.

Use a reference-anchored workflow when identity, product, or set must survive across shots. Start from a still, a frame grab, or a concept image, and write the prompt as a modification of that reference rather than a description from nothing. This costs a step up front and saves a dozen renders later.

A practical hybrid, and the one most working creators settle into: prompt-only for exploration, reference-anchored for the final sequence. Explore broadly and cheaply, then lock the two or three frames that define the look and rebuild the sequence around them.

FAQ

How long should a video prompt be? Long enough to cover the five blocks and no longer. Most effective prompts run two to four sentences. Beyond that you start adding competing instructions that dilute each other.

Does word order matter? Yes. Put the frame first — shot size and subject — then setting and light, then movement, then style. Early tokens generally carry more weight, so front-load what must be true.

Can I copy prompts that worked for someone else? You can, and they are useful starting points, but treat them as vocabulary rather than recipes. The same sentence behaves differently with a different reference image or a different duration. Adapt, then test.

Why does my footage look like stock video? Usually because the movement block is missing or generic. Add a specific pace and direction: a slow push in, a lateral tracking move, a handheld drift. Specificity in motion is what separates authored footage from library footage.

How many renders should a shot take? Two to four is normal for a well-written prompt. If you are past six, the prompt is probably contradictory rather than unlucky. Stop rendering and re-read it block by block.

Should I write dialogue into prompts? Describe performance and mouth movement, but plan to record or synthesize audio separately. Writing a line of dialogue into a prompt usually produces mouth motion that does not match the words you later place over it.

What is the best way to learn this quickly? Take one scene you love, describe it shot by shot in plain language, then rewrite each description using the five blocks. Do that ten times and the structure becomes automatic.

Write the shot, not the feeling

The shift from vague prompting to shot-based prompting is the same shift a director makes when a scene moves from the page to the floor: you stop asking for an emotion and start building the conditions that produce it. Light, lens, movement, and a single clear action do more emotional work than any adjective you could add.

Start with one scene you already know well. Write the five blocks for three of its shots, render them, and cut them together. The moment you see a sequence rather than a clip, the workflow clicks — and everything after that is refinement.

Open the Orelon AI video generator, write your first shot card, and let the camera language do the work. Cinematic ideas in motion start with a prompt that knows where to stand. If you want to see how other creators structure their sequences, the Orelon blog has more breakdowns to borrow from.