Orelon logoOrelon
Precios

Text to Video Prompts: Directing Cinematic AI With Words

11 nov 2026 · Por Orelon Team

Explora plantillas de video con IA

Echa un vistazo a algunas creaciones de la comunidad para inspirarte y abre cualquier plantilla para seguir creando en Orelon.

Write prompts that behave like a shot list: subject, action, lens, light, and continuity rules that turn one AI video clip into a finished sequence.

A text prompt is the cheapest piece of equipment in your production and the easiest one to waste. Type a wish — “cinematic drone shot of a city at sunset” — and you get something that technically matches the sentence but plays like stock footage with nobody behind the camera. The creators getting shots that feel directed are not typing better wishes. They are writing shootable instructions: a specific subject doing a specific thing, in a place with texture, seen through a named lens, at a chosen moment in time.

That shift from wishing to instructing is the entire skill. It applies whether you are generating a five-second product reveal with an AI video generator or assembling a sixty-second narrative open from a dozen separate clips. What follows covers the decisions every prompt has to make, how to plan shots that cut together, how to describe camera and light in language that actually lands, how to diagnose a weak take instead of rerolling blindly, and which habits reliably produce forgettable video.

A Prompt Is a Shot List, Not a Spell

The lottery mindset treats generation as chance: type something, pull the lever, hope. That is fine for a single experiment and useless the moment you need consistency, revision, or a second shot that matches the first. A director does not describe a feeling and walk away. A director decides what the frame contains, where the camera sits, what changes between the first and last second, and what the audience is supposed to notice.

Every one of those decisions can be expressed in text — if you stop writing adjectives and start writing behavior.

Mood versus instruction

  • Wish: “A lonely man in a big city, cinematic, moody.”
  • Instruction: “A man in a wet grey overcoat stands still on a crowded crosswalk while pedestrians blur past him in both directions. Shoulders tight, jaw set, gaze fixed slightly past the lens. Overcast daylight, desaturated blue-grey palette, faint rain on the glass.”

The second version is not more poetic. It is more shootable. It describes only things a camera could record, which is precisely the domain these systems work in. Loneliness is not a setting you can switch on; it is a byproduct of stillness inside motion, a tight jaw, and cold light.

What the model will not infer for you

Generation models are pattern completers, not collaborators. They will not ask what you meant, will not preserve a character’s wardrobe because it was mentioned once, and will not know that your scene is supposed to happen in the same location as the last one. Anything you leave out gets filled with the average of everything the model has seen — and the average is exactly what “looks like stock footage” means.

Two consequences follow. First, state what matters. Second, delete what does not. A prompt that lists nine adjectives gives the model nine weak signals; a prompt that lists three strong, compatible signals gives it a direction.

The Six Layers of a Prompt That Shoots

Reliable prompts stack decisions in a predictable order. Skipping a layer produces generic output; contradicting yourself inside a layer produces mush.

  1. Subject — who or what, with at least one identifying detail: age band, build, wardrobe fabric, material finish, species, wear and tear.
  2. Action — a verb with a beginning and an end. “Lifts the cup and drinks” beats “is drinking.”
  3. Space — location, time of day, weather, background activity, plus one sensory detail such as steam, dust, or reflected signage.
  4. Optics — shot size, angle, lens feel, depth of field, and one camera movement.
  5. Light — source, direction, hardness, color.
  6. Constraints — aspect ratio, duration, motion cadence, palette limits, and anything that must not appear.

A skeleton you can reuse

“[Subject with one detail] [single action with an end state] in [space with time, weather, and one sensory detail]. [Shot size], [lens feel], [one camera movement]. [Light source and direction], [palette of two or three words]. [Duration], [aspect ratio], no on-screen text.”

Fill it in once, badly, then rewrite only what you can picture more clearly. The skeleton is not a formula for sameness — it is a checklist that stops you from forgetting the camera, which is the single most common omission.

Filling it in

“An older ceramicist with clay-dusted forearms shapes the rim of a bowl on a spinning wheel, then stops the wheel with her palm. A narrow studio with north-facing windows, shelves of unfired pots behind her, dust hanging in the light. Medium close-up, 50mm feel, shallow depth of field, slow push in. Soft directional daylight from camera left, warm clay tones against cool grey walls. Six seconds, vertical, no on-screen text.”

Nothing there is exotic. It simply answers the questions a camera operator would ask before rolling.

Writing Action That Survives Generation

Most morphing, warping, and melted anatomy comes from asking for too much motion in too little time.

Verbs with edges

Choose actions with a clear start and a clear finish: lifts, turns, opens, sets down, steps through, switches off. Avoid continuous states — running, dancing, fighting — unless that motion is the entire point of the clip, and avoid stacked actions such as “she walks in, sits down, opens a laptop, and smiles” crammed into four seconds. Four to six seconds is roughly one gesture, one reaction, or one reveal.

Micro-blocking beats description

Instead of “she looks surprised,” write what the body actually does: “her hand stops halfway to the cup, eyes lift, breath held.” Instead of “he is nervous,” write “his thumb taps the steering wheel twice, then stops.” Behavior is visible; emotion is not. Every time you catch yourself naming a feeling, ask what a camera would see in its place.

Decide where the clip begins and ends

State the entry and exit conditions before generating. “Begins with the door closed, ends with it open and the room lit” is a tiny story the model can complete. Without it, you get a loop with no arc — technically moving, dramatically inert, and hard to cut against anything else.

Camera Language That Changes Perceived Quality

Camera vocabulary is the highest-leverage text in the prompt because it controls how the viewer’s eye travels through the frame. Use conventional industry terms; they carry stable visual meaning and are far more useful than poetic description.

Shot size: extreme wide, wide, full, medium wide, medium, medium close-up, close-up, extreme close-up, macro.

Angle: eye level, low angle, high angle, overhead, Dutch tilt, over-the-shoulder, point of view.

Lens feel: 24mm with visible distortion, 35mm natural, 50mm neutral portrait, 85mm compressed background, macro with razor-thin focus.

Movement: locked off, slow push in, pull out, orbit, crane up, tracking alongside, handheld drift, tilt reveal, dolly through a doorway.

One move per clip

Push in and orbit and crane usually resolves into a wobble that reads as a mistake rather than a style. Pick one primary movement and, if you want texture, add a modifier: “slow push in, subtle handheld drift.” Locked-off shots are underrated — they let performance and light carry the frame, and they cut cleanly against movement elsewhere in the sequence.

Depth cues earn the cinematic look

Shallow depth of field, foreground occlusion — a doorway edge, a plant, a shoulder — and haze between camera and subject all create separation. A flat frame with no foreground and no falloff looks like a screensaver. Say “foreground blur of a hanging plant, subject sharp at mid-distance” and the image instantly gains layers.

Compose for your aspect ratio

Vertical favors single subjects, stacked depth, and strong vertical leading lines; widescreen favors environment, lateral movement, and negative space to one side where titles can live. Decide this before generating anything. Reframing a vertical composition into widescreen in the edit usually costs you the subject’s head or the punchline.

Light Is Geometry, Not Mood

“Beautiful lighting” tells the model nothing. “Soft window light from camera left, falling off into shadow on the right” tells it everything it needs.

Describe light in three steps: direction (where it comes from), quality (hard or soft, and how large the source is), then color (warm tungsten, cool daylight, greenish fluorescent, blue-hour ambient). Direction comes first because it defines the geometry of the face and the room.

Usable light phrases

  • Hard noon sun from above, sharp shadows on the pavement.
  • Single practical lamp on the desk, everything beyond it falling to near-black.
  • Backlit rim through window haze, face two stops under.
  • Bounced fill off a white wall, soft and nearly directionless.
  • Neon spill from an off-screen sign, magenta on one cheek, cyan on the other.

Contrast is a decision

State the contrast level explicitly: high-contrast with crushed shadows, or low-contrast and milky. Then state the palette in two or three compatible words — “muted teal and rust,” “bone white and charcoal.” Three palette words that fight each other produce the muddy middle ground that reads as amateur no matter how good the subject is.

Say what must not appear

Negative instructions are cheap insurance: no on-screen text, no logos, no extra people, no lens flare, no slow motion. Treat in-frame typography as unreliable in any case; generate clean areas and add titles in the edit, where they stay legible after compression and can be revised without regenerating footage.

Plan a Sequence, Not a Clip

A single clip is a demo. A sequence is a film. The difference is coverage planned before you generate anything.

The five-beat spine

Write five plain-language beats, then expand each into a full prompt:

  • Establish (3–5s): wide or extreme wide, environment and mood, static or very slow movement.
  • Introduce (3–5s): medium shot, subject enters the story, one clear action.
  • Detail (2–4s): close-up or macro, texture — hands, stitching, steam, a mechanism.
  • Turn (3–5s): something changes. A look, a step, a light switching on.
  • Resolve (3–5s): wider again, or a clean frame with room for titles.

Continuity lives in the text

Models do not remember your previous clip. Continuity has to be pasted in every time. Build four reusable blocks and keep them identical across the sequence:

  • Character block: age band, build, hair, wardrobe fabric and color, one distinctive feature.
  • Palette block: the same three color words and the same contrast level.
  • Optics block: the same lens feel and depth of field.
  • Time block: the same hour, weather, and light direction.

Change one block per clip, never four. If shot two switches to golden hour while shot one was overcast, the cut reads as two different films spliced together by accident.

Design transitions in camera

Instead of relying on editing tricks, generate matching movement: end clip one on a push toward a doorway, open clip two on a push through it. Or match shapes — a circular bowl rim cutting to a round window. Matching motion and shape inside the footage makes the edit feel intentional and hides the seams between separately generated clips.

Iterate by Diagnosis, Not by Rerolling

Nobody lands a final take on the first attempt. The productive loop has three moves: draft, diagnose, refine.

Diagnose the failure type

Symptom Likely cause Fix
Wrong wardrobe or age Subject layer too thin Add fabric, fit, and one distinctive detail
Warped hands, melting objects Too much simultaneous motion Simplify to one gesture, shorten duration
Character drifts between clips Continuity block missing or edited Paste the block unchanged, reduce camera movement
Flat, lifeless image No light direction, vague praise words Add source, direction, contrast
Cluttered composition Multiple subjects and actions Remove one subject, state shot size and “single subject, centered”
Garbled in-frame text Asking for readable typography Remove text, add graphics in the edit

One variable per pass

When a frame is close, stop regenerating the whole scene. Test lighting on a locked composition, then motion, then duration. Where the tool supports it, reuse the same starting frame so composition stays fixed while you change the variable under test. Two clean comparisons teach you more than twenty random attempts.

Keep a prompt log

A plain text file listing each version, the single change made, and a one-line verdict becomes your personal prompt library. It is the reason your fifth project takes a fraction of the time of your first, and it stops you from rediscovering the same settings twice.

Workflows That Fit Real Projects

Product teaser without a studio

Work macro-first. Generate three texture close-ups — stitching, brushed metal, a liquid pour — then one hero medium shot with a named light source, then one clean wide for the end card. Limit the palette to two colors plus a neutral. Never describe the label text; reserve that area for real typography added in the edit.

Explainer and educational sequences

Abstract ideas need physical metaphors. “Data moving through a network” becomes “hundreds of small glass spheres rolling along thin copper rails, some branching left, some right, lit warm from below.” Keep the same metaphor family and palette across every clip so the sequence teaches rather than dazzles. A reusable visual grammar is worth more than any single clever shot.

Social-first hooks

The first second decides whether anyone sees the second. Open mid-action, in a frame that would stop a thumb: an unusual angle, an unexpected scale, a face very close to the lens. Keep clips short, cut on movement, and design frame one to work as a still thumbnail. A quiet establishing shot is an excellent second beat and a poor first one.

Pre-visualization for live shoots

Before a real shoot, generate rough versions of key shots to test lens choice, blocking, and light direction. Arguing with a generated frame is faster and cheaper than arguing with a rented lighting package, and the frames make a clear mood board for a crew. Start from video templates if you want to see how a strong shot is structured before adapting it to your own scene.

Habits That Make Output Look Generic

  • Stacking praise adjectives. “Cinematic, epic, stunning, masterpiece” pulls in conflicting directions and decides nothing.
  • Describing inner states. Invisible to a camera. Replace with behavior.
  • Crowding four seconds. Two subjects and three actions produce blended anatomy.
  • Omitting the camera. No shot size and no movement defaults to a neutral middle distance.
  • Deciding aspect ratio and duration after generating.
  • Expecting reliable dialogue or lip sync from a visual model; plan voice and sound as a separate layer.
  • Never saving what worked.

A seven-point pre-flight check

  1. Is there exactly one primary action with a clear start and end?
  2. Does the prompt name a shot size and one movement?
  3. Is the light source described with a direction?
  4. Are palette words limited to three and mutually compatible?
  5. Are aspect ratio and duration stated?
  6. Is there at least one exclusion for what must not appear?
  7. Would a camera operator understand this sentence on set?

Seven yeses and you are already ahead of most generated video published today.

FAQ

How long should a prompt be? Long enough to cover the six layers, rarely more than 60–90 words. If it grows past that, you are usually describing two shots. Split them.

Do longer prompts produce better video? Not reliably. Specificity helps; volume does not. Ten precise words about lens and light beat forty adjectives about mood.

How do I keep the same character across clips? Write a fixed character block — age, build, hair, wardrobe, one distinctive feature — and paste it unchanged into every prompt. Keep lens and framing consistent, and avoid large camera moves that reveal angles you never described.

Why does my footage look like stock video? Usually because the prompt named a subject and a place but no point of view. Add a specific camera position, an unusual angle, and a reason the camera is standing there. Perspective creates authorship.

Can I get readable text inside the video? Treat in-frame typography as unreliable. Generate clean, low-detail areas and set titles, labels, and captions in your editor, where they survive compression and can be corrected later.

Should I storyboard before prompting? For anything longer than one clip, yes. A five-line shot list in a notebook saves more time than any prompt trick, and it is the difference between a sequence and a pile of clips.

What is the fastest way to improve? Pick one variable per week — camera language first, then lighting, then continuity — and compare outputs side by side. Deliberate comparison beats random experimentation.

Direct With Words, Then With Frames

The gap between a forgettable clip and a cinematic one is rarely the model. It is the thinking in front of the model: a subject with detail, an action with edges, a camera with intent, light with a direction, and a sequence that cuts together. Write those five things down and you have effectively directed your own footage without a crew, a permit, or a lighting truck.

Start with one shot you can already see clearly in your head. Write it as behavior, not mood. Add a lens and a light source. Generate, diagnose, change one thing, and go again. When you are ready to move from a single clip to a finished sequence, Orelon gives you the canvas for cinematic ideas in motion, and the Orelon blog has more workflows to borrow when your shot list gets ambitious.