Text-to-video generation stopped being a novelty the moment its output became usable in real edits. The bottleneck is no longer access to a model — it is describing a shot precisely enough that the model builds what you imagined instead of a plausible average of everything it has seen. Prompting for video is a separate discipline from prompting for stills, because every prompt now has to survive motion: subjects move, cameras pan, light shifts, and continuity has to hold across seconds instead of a single frame.
This guide covers the craft layer above the tool: how to structure a video prompt, how to speak in the vocabulary of a camera crew, how to feed a model more than text, and how to build a workflow that produces repeatable results rather than lucky accidents.
Why most text-to-video prompts underperform
Almost every weak result traces back to one of three failure modes.
The first is vagueness by omission. A prompt like "a woman walking through a city at night" hands the model hundreds of valid interpretations. It will pick one at random, and the random pick is usually the most statistically common version of that sentence — flat lighting, generic street, average framing. Nothing in the prompt tells the model which of the ten thousand possible shots you actually want.
The second is over-stuffing. Creators who learn the first lesson often add forty adjectives, three competing camera moves, and an emotional arc in a single sentence. The model then averages the contradictions: a slow dolly that also orbits, a sunny scene with storm clouds, a joyful face with melancholic posture. Contradictions do not create complexity; they create mush.
The third is ignoring the time axis. A still-image prompt describes a moment. A video prompt describes a change — something must be different at the end than at the beginning. If nothing in your prompt implies motion, transformation, or progression, the model fills that space with the cheapest available motion: drifting limbs, flickering backgrounds, and the notorious slow zoom that looks like a mistake.
The fix for all three is the same: treat the prompt as a shot description written for a crew, not a wish written for a genie.
The anatomy of a video prompt that actually works
An effective video prompt has a predictable skeleton. Once you internalize the skeleton, you can vary it freely without losing control of the output.
The five core slots
Most reliable prompts cover five slots, usually in this order:
- Subject — who or what, with two or three concrete visual anchors (age range, wardrobe, material, texture).
- Action — what changes during the shot, expressed as a verb with a direction and a speed.
- Environment — location, time of day, weather, background activity, depth cues.
- Camera — shot size, angle, movement, lens character.
- Look — lighting, palette, film stock or render style, mood.
Order matters less than completeness, but keeping a consistent order in your own prompts makes it far easier to debug a bad result. If the lighting is wrong, you know exactly which line to edit.
Structure versus creative latitude
A common worry is that structured prompts kill creativity. In practice the opposite happens. Structure absorbs the decisions you do not care about so you can spend your attention on the decisions you do. If the wardrobe is irrelevant to the story, specify something neutral and move on. If the exact shade of the lantern light matters, spend three words there instead of one.
The practical rule: be specific where the shot fails without precision, and deliberately loose everywhere else. "A weathered fisherman" is loose enough for the model to add character detail you would never have thought to write. "A 60-year-old fisherman with a scar through his left eyebrow, salt-crusted beard, and a wool sweater the color of wet slate" leaves nothing to chance — useful for continuity, wasteful for a one-off mood shot.
A worked example
Here is a prompt that produces consistent, editable results:
Medium close-up, eye level, slight handheld drift.
A woman in her early thirties, cropped dark hair, damp from rain,
wearing a charcoal wool coat with the collar up.
She turns her head slowly from left to right, scanning the empty street,
then exhales visibly.
Rainy city street at blue hour, shallow puddles reflecting neon signage,
blurred pedestrians far in the background, no readable text.
Anamorphic lens look, soft highlights, cool cyan and amber palette,
subtle film grain, naturalistic lighting from practical signs.
Compare that with "a woman in the rain looking around a city." The difference is not verbosity — it is that every extra phrase removes an ambiguity the model would otherwise resolve by guessing. The turn from left to right is the shot. The exhale is the emotional beat. The cyan-and-amber palette is the grade you will not have to fight in post.
Cinematic control: camera, lens, lighting, motion
The single fastest quality upgrade available to any prompt writer is learning to write like a director of photography. Models have absorbed a enormous amount of film language, and film language maps onto technical parameters far more reliably than emotional adjectives do.
Camera vocabulary that models understand
Shot size: extreme wide, wide, full, medium wide, medium, medium close-up, close-up, extreme close-up, insert, macro.
Angle: eye level, low angle, high angle, overhead, top-down, dutch tilt, over-the-shoulder, point of view.
Movement: static lock-off, slow push in, slow pull out, dolly left, truck right, pan, tilt, crane up, arc around subject, tracking follow, whip pan, handheld drift, gimbal glide.
Lens character: wide-angle distortion, telephoto compression, anamorphic flare, macro shallow focus, deep focus, 24mm, 50mm, 85mm.
Two rules make this vocabulary work. First, name exactly one primary movement per shot. A push in and an arc are different shots; asking for both produces neither. Second, pair movement with a speed word — slow, gradual, deliberate, brisk — because unqualified movement defaults to whatever the model considers average, which is usually too fast.
Lighting and color as instructions
Lighting descriptions do double duty: they set mood and they constrain the render. Useful terms include soft key light, hard rim light, practical sources, motivated light, overcast diffusion, golden hour backlight, neon spill, firelight flicker, volumetric haze, and silhouette against a bright background.
Color is often better handled as a palette than as a mood word. "Warm nostalgic" is ambiguous. "Amber highlights, deep teal shadows, muted mid-tones" is a grade. Naming two or three colors keeps the model inside a coherent world and makes multi-shot sequences look like they belong to the same film.
Motion and pacing
If you want a shot to feel calm, say so through physics: slow movement, steady framing, gentle motion, long take. If you want energy, use short shot duration, handheld drift, faster subject action, and tighter framing. Pairing a fast camera move with a slow subject, or vice versa, is one of the most effective ways to create tension on purpose — but it only works when both speeds are explicitly stated.
Beyond words: multimodal inputs
Text is the entry point, not the ceiling. Modern video workflows accept image references, video references, and increasingly audio context, and each of these solves a specific problem that words handle poorly.
Image references and character consistency
The classic text-to-video weakness is identity drift: a character's face, wardrobe, or proportions shift between shots. A single reference image anchored into the prompt dramatically reduces this. The workflow looks like this:
- Generate or photograph a clean, well-lit reference of the subject — neutral expression, simple background, sharp detail.
- Attach it as a reference and describe only what should change: action, camera, environment.
- Avoid re-describing the subject's appearance in text; conflicting text and image descriptions force the model to compromise.
- Test one short shot first. If identity holds, expand to a sequence using the same reference every time.
For multi-character scenes, keep separate references and describe positioning explicitly ("she stands screen left, he enters from the right") so the model does not blend features.
Audio-driven prompting
Audio context is underused. Feeding a music bed or a voice clip alongside the prompt helps the model match pacing, rhythm, and emotional register. A cut that lands on a downbeat is a prompt decision as much as an editing decision. If your tool supports it, describe the sync intention in text too — "the subject turns on the beat" — because explicit timing language reinforces what the audio implies.
Layering references without confusing the model
Two or three references are usually helpful: one for identity, one for environment or palette. Five is chaos. When you combine references, add a short line that resolves priority: "Match the character from the reference image; take the environment and lighting from the written description." Explicit priority statements prevent the model from averaging two incompatible visual worlds into something that satisfies neither.
Model-specific prompt strategies
Different model families reward different prompt habits. You do not need to memorize parameter names, but you do need to recognize which category you are working with.
Realism-first models respond best to photographic language: lens choice, sensor behavior, naturalistic lighting, film grain, shallow depth of field. They punish fantasy vocabulary and abstract adjectives. Write them like a camera report.
Stylized and animation-leaning models reward art-direction language: illustration style, cel shading, painterly texture, graphic shapes, bold palettes. Photographic terms still help with framing, but color and texture carry more weight than lens realism.
Fast draft models are for structure, not finish. Prompt them for composition and motion only, and accept soft detail. Use them to test blocking and timing, then re-render the approved shot on a slower, higher-fidelity model with the same prompt.
Image-anchored workflows care most about the reference, not the prose. In those pipelines, your text should describe change and camera, nothing else.
The practical takeaway: keep two versions of every prompt — a long, fully specified version for final renders and a short, slot-based version for drafts. Moving between models becomes a copy-paste operation instead of a rewrite.
A shot-by-shot workflow from script to final render
Good prompts come from good preparation. This sequence keeps prompt writing fast and keeps the resulting footage editable.
Step 1 — Beat sheet
Write the story as five to ten beats in plain text. No visuals yet. What changes between the first beat and the last?
Step 2 — Shot list
Convert each beat into one or more shots. For each shot, note only four things: shot size, movement, subject action, and the single most important visual element. This is your storyboard in words.
Step 3 — Prompt blocks
Expand each shot into a full prompt using the five slots. Keep a shared header block with recurring details — palette, grade, film look, environment rules — and paste it into every prompt in the sequence. Shared blocks are what make separate clips feel like one film.
Step 4 — Test renders and continuity check
Render every shot at low quality first and watch them in order, not individually. Problems that are invisible in isolation — a lighting shift, a wardrobe change, a subject who suddenly faces the wrong direction — become obvious in sequence.
Step 5 — Finish and assemble
Re-render only the shots that need it. Then do the work that makes AI footage feel intentional: cut on motion, add sound design, apply one consistent grade across all clips, and keep shot durations roughly proportional to their importance. A single grade and a coherent sound bed do more for perceived quality than any prompt refinement.
Common mistakes and how to fix them
Flicker and texture crawl. Usually caused by an under-specified style or too many competing texture words. Fix: state one render style, one grain level, and one palette.
Rubber limbs and morphing faces. Caused by asking for complex action in a tight frame. Fix: simplify the action, widen the shot, and anchor identity with a reference image.
Everything happens too fast. Movement without a speed qualifier. Fix: add slow, gradual, or minimal effort, and reduce the number of things moving at once.
Text and logos appear as gibberish. Generative models default to plausible-looking glyphs. Fix: add "no readable text, no signage, no logos" whenever signage might enter frame.
Contradictory lighting. Two light sources described as equally dominant. Fix: name a key and a secondary, and say which one motivates the scene.
Every shot looks like a different film. No shared header block. Fix: centralize palette, grade, lens family, and environment rules, then reuse them verbatim.
The subject never changes expression. Prompts describe appearance but not performance. Fix: describe one physical beat per shot — a turn, a breath, a hand movement, a shift in posture.
Building a prompt library you can reuse
Once a sequence works, the value is in the parts, not the whole clip. Break winning prompts into reusable fragments and store them by function: lighting setups, lens packages, environment templates, performance beats, negative constraints. A library of thirty fragments lets you assemble a new shot in a minute while keeping visual consistency with everything you have already made.
Tag each fragment with what it solved. "Soft rim light, practical spill — used for night interiors" is far more useful six months later than a folder of untitled text files. Also keep a short list of phrases that consistently caused problems in your specific models, and treat that list as a checklist before every render.
Frequently asked questions
How long should a video prompt be?
Long enough to remove ambiguity, short enough to stay coherent. In practice that is usually 50 to 120 words plus a shared style block. If you cannot tell which sentence caused a specific problem in the output, the prompt is too dense.
Do I need to write prompts in English?
Many models perform slightly better with English prompts because of training data distribution, but that is not a reason to abandon your working language. A reliable compromise is to keep technical terms — lens, lighting, movement — in English, and describe story and emotion in your own language.
Can one prompt generate a whole scene?
Reliably, no. One prompt produces one shot. Scenes are built by combining shots that share a style block. Trying to compress a scene into a single paragraph usually produces a soft, directionless clip that has to be reshot anyway.
How do I keep a character consistent across many shots?
Use a single high-quality reference image, describe only the action and camera in text, and reuse a shared style header verbatim in every prompt. Check identity drift in a low-quality sequence pass before committing to final renders.
Is negative prompting worth the effort?
Yes, but keep it short and literal. A handful of targeted exclusions — no text, no logos, no distortion, no crowd — solves specific recurring problems. Long lists of vague negatives mostly dilute the positive description.
A pre-render checklist
Before you commit to a final render, confirm that the prompt names one primary camera movement with a speed, one dominant light source, a defined palette, one clear action beat, a stated shot size and angle, an explicit reference priority if images are attached, and a short set of negative constraints. Then render a low-quality pass, watch it in sequence with neighbouring shots, and only promote what survives that test.
That discipline is what separates prompting as a hobby from prompting as a production method. The tools will keep changing and the model names will keep rotating, but the underlying craft — describing a shot so precisely that motion, light, and continuity stay under your control — is the part that carries over. Master the description, and every new model becomes an upgrade rather than a new learning curve.
