Learn how to write ChatGPT image prompts with clear subject, style, and camera language, then adapt them into cinematic AI video shots.
Most creators arrive at video generation holding a folder of image prompts that already work. The stills land: the product sits on the right surface, the character has the right face, the light falls where it should. Then the same words go into a motion tool and everything drifts — the camera lurches, the subject morphs, the mood evaporates two seconds in.
The gap is rarely the model. It is the grammar of the prompt. An image prompt describes a moment; a video prompt has to describe a moment, plus what changes, plus how the camera witnesses that change. Learning the translation is the highest-leverage skill in AI filmmaking right now, because it lets one idea travel across stills, shots, and full sequences without being rebuilt from scratch each time.
Why a Chatbot Is Still the Best Storyboard Scratchpad
A common point of confusion: ChatGPT does not render pixels. It writes text, and that text becomes the instruction another model follows. Used well, it is the cheapest pre-production tool you own — a writing partner that can expand a vague idea into a shot list, compress a rambling description into twelve sharp words, and expose your own ambiguity before you spend a single generation.
What a text model is genuinely good at
- Compression. Feed it a paragraph and ask for the same scene in under 25 words, keeping the lens and the light. The result usually fits tight input limits without losing the idea.
- Variation. Ten rewrites of one scene with different eras, palettes, or weather conditions give you a mood board in text form.
- Precision naming. "Vintage" is a dead end; "1970s Kodachrome editorial, slight halation, warm shadows" is a direction. A language model is good at proposing those specific terms.
- Continuity notes. Ask for a character sheet — wardrobe, hair, distinguishing marks, posture — and reuse it verbatim across shots.
- Critique. Paste a prompt and ask what a literal-minded reader would misread. The answer is usually the bug.
What it cannot do for you
It cannot see the output. Judgment about whether a frame actually works still belongs to you, and that judgment is what separates a portfolio from a folder of experiments. Treat the text model as a collaborator with strong vocabulary and no eyes.
The Anatomy of a Prompt That Survives Motion
Every prompt that holds up over several seconds tends to contain the same four layers. Missing one is survivable. Missing two is where the drift begins.
Subject and continuity
Say who or what, in what state, wearing or holding what. For recurring characters, fix the details once and paste them unchanged: "late-30s woman, short copper hair tucked behind the left ear, grey wool coat with a torn cuff." Small specifics act as anchors. When a video model has to invent a face from "a woman," it invents a new one at every frame boundary.
Style and grade
Style words fall into three families, and mixing families carelessly is what makes prompts feel random:
| Family | Examples | What it controls |
|---|---|---|
| Medium | 35mm film, digital cinema camera, watercolor, cut-paper | Texture and material logic |
| Era or movement | 1970s editorial, Italian neorealism, brutalist graphic | Reference frame for taste |
| Grade | teal shadows, low-contrast pastel, high-key bleach bypass | Color and tonal response |
Pick one from each family at most. Three grade words fight each other; one grade plus one medium reads cleanly.
Camera and lens
For stills this is flavor. For video it is structure, because the camera implies movement. "85mm, shallow depth of field, static tripod" tells a model to hold still. "24mm handheld, slight drift" tells it to move. Naming the movement — push in, pull out, orbit, crane down, tracking left — is far more reliable than hoping the model infers it.
Light and atmosphere
Light is the fastest way to make two unrelated shots feel like one film. Write it as a physical setup rather than a mood: "single window camera-left, overcast bounce, dust in the air" beats "moody lighting" every time.
A Reusable Prompt Template
A fill-in structure keeps you from forgetting a layer under deadline pressure:
Subject + state + action, in location, at time of day. Medium and grade. Lens and camera movement. Light setup. Continuity anchors. Negative constraints.
Two worked examples:
Product, 6 seconds. "Matte black water bottle standing on wet slate, condensation beading, rotating slowly a quarter turn. Studio macro photography, neutral grade, 100mm lens, locked-off tripod with a 5% push-in. Single softbox from top-right, hard rim light behind. Label faces camera at start. No text overlays, no hands, no studio reflection."
Character, 8 seconds. "Woman in her late thirties, copper hair tucked behind left ear, grey wool coat, walking toward camera through a rain-slicked night market. Anamorphic 40mm, shallow focus, slow tracking backward. Practical neon from stall signs, cool ambient fill. Coat and hair unchanged throughout. No crowd close to camera, no dialogue."
Notice that both prompts name what must not happen. Negative constraints are cheap insurance in motion, where a model has dozens of chances per second to invent something wrong.
Where Image Prompts Break the Moment Motion Starts
Three failures repeat often enough to plan for.
Ambiguity compounds. A still has one frame to be wrong in. A clip has a hundred and fifty. Any word with two readings — warm, modern, clean — will get both readings inside the same shot.
Nouns without verbs. Image prompts are noun piles. Video needs a change of state: opening, turning, arriving, settling, dissolving. If your prompt has no verb a camera could observe, the model invents motion for you, and it usually picks drift.
Physical impossibilities that read fine as stills. Two light sources that contradict, a hand with an unclear grip, a reflection that does not match the subject. On a still you crop around it. In motion it becomes the thing viewers notice first.
Converting a Still Prompt Into a Shot Sequence
The translation is mostly editorial work: deciding what the audience needs to see, in what order, and how long each beat holds.
The three-beat rule
Give every short clip a beginning state, a change, and an end state. Establish the bottle standing; it rotates; it settles facing the label. Establish the woman mid-stride; she passes a stall; she continues out of frame. Three beats in six to eight seconds is comfortable. Five beats is a fight with the model.
Motion verbs the model can actually read
Prefer verbs with a visible physical result: turns, lifts, slides, opens, folds, pours, steps, reaches, exhales. Avoid interior verbs — remembers, realizes, decides — unless you translate them into something the camera can capture, like a pause, a glance down, a hand closing.
Continuity anchors between shots
Write a short block that never changes and paste it into every prompt in the sequence: wardrobe, hair, key prop, lens family, grade, light direction. Sequences fall apart not because any single shot is bad, but because shot four has a different coat.
Then assemble in a tool that holds the look steady across generations. The Orelon AI video generator is built for exactly this pass, and pairing it with the image generator lets you lock a hero frame before animating it.
Prompt Patterns for Common Video Types
Product and packshot
Priorities: geometry, label legibility, controlled reflections. Use macro or 100mm framing, a locked camera with one small deliberate move, and explicit "label faces camera" language. Keep the background simpler than you think — motion plus busy background equals mush.
Character scenes
Priorities: face stability and wardrobe. Front-load the character sheet, keep the action single, and avoid crowds unless the crowd is the point. If the shot needs a spoken line, plan a separate audio pass; generated video rarely lands lip sync without extra work.
Environment and establishing shots
Priorities: depth and parallax. Give the camera a reason to move — a slow crane down over a valley, a lateral drift past a warehouse — and describe foreground, midground, and background separately. Flat descriptions produce flat plates.
Abstract and motion graphics
Priorities: rhythm and color discipline. Two colors and one repeating shape beat five colors and chaos. Describe the transition, not just the elements: "bands of color slide left, overlapping, then collapse into a single line."
Iterating Without Burning Your Week
Iteration discipline is worth more than prompt cleverness. A workable loop:
- Lock the frame first. Generate stills until composition and grade are right. Failing here is fast and cheap.
- Change one variable at a time. Camera, then light, then action. Changing three things at once teaches you nothing about which one worked.
- Test short. Six seconds at low resolution tells you whether the motion idea holds.
- Judge muted, at normal speed. If the shot does not read silent, the prompt is doing too little and the music is doing too much.
- Log the winner. Copy the exact prompt into a library with a one-line note about what it produced.
Browse the prompt library when you are stuck; reading how other people phrase camera moves is faster than deriving them from first principles.
Mistakes That Quietly Ruin Output
- Overloading the prompt. Twenty details means the model honors twelve and improvises eight. Cut to what the shot is about.
- Style stacking. "Cinematic, filmic, movie-like, Hollywood" is one instruction repeated four times, and it dilutes the useful words around it.
- Unfixed subjects. Any recurring character without a written character sheet will change faces between generations.
- Ignoring aspect ratio. A composition built for 16:9 dies in 9:16 unless you re-block it for vertical, subject higher in frame.
- No negative constraints. Without them, models add lens flares, floating text, extra fingers, and slow drift.
- Chasing a shot that will not resolve. After three failed passes the prompt is not the problem — the concept is. Storyboard it differently.
Building a Prompt Library That Compounds
Save prompts the way you save code: named, versioned, annotated. A simple structure works — character_mara_v3, product_bottle_toplight_v2 — with three fields per entry: the full prompt, the aspect ratio and duration it was tested at, and one line of honest commentary.
Tag entries by intent rather than by tool: establishing, product-hero, handheld-dialogue, transition. When a new project starts, you shop from your own shelf instead of starting blank. Within a few months this becomes the real asset — more valuable than any single render, because it encodes your taste in a form you can reuse. Browsing templates is a good way to see how other creators structure that shelf before building your own.
Pair each entry with its still prompt. The still is the reference frame; the motion prompt is the instruction for how that frame behaves. Keeping them side by side makes future adaptation trivial.
FAQ
Do I need separate prompts for images and video? Usually yes, but they should share a spine. Write the still prompt first to lock composition and grade, then rewrite the same idea around action and camera movement.
How long should a prompt be? Long enough to fix subject, style, lens, light, and one action — typically 40 to 80 words. Beyond that, extra clauses start competing with each other.
Can one prompt generate a whole scene? Rarely well. Generate shots, not scenes, then cut them together. A sixty-second piece is usually ten to fifteen short generations, not one long one.
Why does my character's face change between shots? Because the description was general. Write a character sheet with hair, wardrobe, age, and one distinguishing feature, then paste it unchanged into every prompt in the sequence.
How do I stop the camera drifting when I want it still? Say so explicitly: "static tripod, locked frame, no camera movement," and add "no zoom, no drift" to the negative constraints.
Turn the Prompt Into a Moving Frame
The prompt is not the product; the shot is. Write the frame, name the change, anchor the continuity, then let the model do the part it is good at. If you want a place to run that loop end to end — stills, motion, and a library that keeps your best work within reach — start with Orelon and build the next sequence from a prompt you can actually repeat.

