Most people approach AI video generation backwards. They type a sentence, wait for a clip, and then judge whether the tool is any good. In reality, the tool is rarely the bottleneck. The prompt is. A vague instruction produces a vague clip; a well-structured instruction produces something you can actually drop into an edit.
This guide is a compact, practical course on designing prompts for AI video. It walks through the structure of a strong prompt, the camera and lighting vocabulary that models actually respond to, how to keep characters consistent across shots, a repeatable iteration workflow, and the mistakes that waste the most time. No prior experience is assumed.
What a Prompt Really Controls
A video prompt is not a description of an idea. It is a bundle of production decisions compressed into text: who or what is on screen, what they are doing, where they are, how the camera behaves, how the scene is lit, and what visual language the whole thing speaks.
When a result feels wrong, it is usually because one of those decisions was left to chance. If you do not specify camera movement, the model picks one. If you do not specify a time of day, the model picks one. If you do not specify a lens or depth of field, the model defaults to whatever its training data considered average. Every unspecified decision is a coin flip you volunteered for.
Order also matters. Most video models weight the beginning of a prompt more heavily than the end. Subject and action usually sit at the front; style modifiers and technical details sit near the back. If a detail is critical, do not bury it at the end of a long sentence.
Finally, prompts control probability, not certainty. You are nudging a system toward a particular region of its output space. That is why iteration, not perfectionism, is the real skill.
The Six Building Blocks of a Video Prompt
Almost every effective video prompt can be assembled from six components. You do not need all six every time, but knowing them makes it obvious what is missing when a clip disappoints.
1. Subject
Describe who or what the shot is about with enough specificity to differentiate it from every other version of the same thing. "A woman" is weak. "A woman in her sixties with short silver hair and a canvas apron" is usable. Include clothing, build, hair, and any distinguishing object. If the subject is an animal, vehicle, or object, the same rule applies: define the variant, not just the category.
2. Action
Action is a verb, but a good action clause also implies pace and effort. "Walks" is fine. "Walks slowly, leaning into the wind" tells the model something about body mechanics, timing, and resistance. Avoid stacking multiple unrelated actions in one prompt; models handle one continuous action per clip far better than a sequence of events.
3. Setting
Location, era, and weather belong here. "A narrow side street" becomes more useful as "a narrow side street in a dense coastal town, low tide, overcast afternoon." Setting also establishes scale. Are we in a wide landscape or a cramped interior? Models resolve spatial contradictions poorly, so keep the environment internally consistent with the action.
4. Camera
This is the component beginners skip and professionals never do. Specify shot size (wide, medium, close-up), angle (eye level, low, overhead), movement (static, pan, dolly, tracking, handheld), and lens character (wide angle, telephoto, macro) when it matters. Camera language does more for perceived quality than almost any adjective.
5. Light and Color
The model needs to know where light comes from and what it feels like. "Late golden-hour side light, long shadows" and "soft overcast diffusion, muted greens" produce very different clips from the same subject. Color direction — warm, cool, desaturated, high contrast — keeps multiple shots feeling like they belong to the same project.
6. Style, Medium, and Pacing
Finally, define the visual language: cinematic realism, documentary handheld, stop-motion, painterly 2D, analog film grain, clean commercial polish. Style modifiers set the overall rendering. Pacing modifiers such as "slow, deliberate motion" or "energetic, quick movement" help control how the clip feels in time.
The blocks assembled
Here is a prompt built from all six blocks, in the order most models handle well:
Medium shot, a cyclist in a yellow rain jacket pedals slowly through a
flooded side street at dusk, low camera angle following from behind,
neon reflections on wet asphalt, shallow depth of field, cool blue tones
with warm signage highlights, cinematic realism, light rain, slow motion.
Notice there is nothing poetic in it. Every clause is a decision.
Camera and Lighting Vocabulary That Models Understand
You do not need film school, but a working vocabulary pays off immediately. These are the terms that show up most reliably in generated output.
Camera terms worth memorizing
- Shot size: extreme wide, wide, full shot, medium, medium close-up, close-up, extreme close-up
- Angle: eye level, low angle, high angle, overhead or top-down, Dutch tilt, over-the-shoulder
- Movement: static or locked-off, slow pan left or right, tilt up or down, dolly in, dolly out, tracking shot, crane up, handheld, orbit or arc around subject
- Lens character: wide angle with visible distortion, 35mm natural perspective, 85mm portrait compression, macro detail, anamorphic flare
- Focus: deep focus, shallow depth of field, rack focus from foreground to background
Combining two movement ideas in a single clip usually produces mush. Pick one primary movement and one supporting detail.
Lighting and color terms
- Direction: side light, backlight, rim light, top light, frontal soft light
- Quality: hard shadows, soft diffusion, dappled shade, bounced light
- Time: golden hour, blue hour, harsh midday, overcast afternoon, night with practical sources
- Atmosphere: haze, fog, dust in the air, steam, rain, volumetric light shafts
- Palette: warm amber, cool teal, monochrome, pastel, high-contrast, desaturated documentary tones
If you only add one lighting detail, add direction. Flat light is the fastest way to make an AI clip look synthetic.
Text-to-Video or Image-to-Video? A Simple Decision Framework
Most modern video tools support both starting modes, and choosing the right one is a bigger deal than choosing the right adjectives.
When to start from text
Use text-to-video when the shot does not exist yet and you want the model to explore. It is ideal for mood boards, concept shots, establishing landscapes, abstract transitions, and anything where you are still deciding what the scene should look like. Expect to generate several variations and keep the best one.
When to start from an image
Use image-to-video when you already have a reference: a still you generated earlier, a product photo, a character sheet, or a frame you want to animate. This mode locks composition, color, and identity, so your prompt shifts from describing the scene to describing motion, camera behavior, and atmosphere. A typical image-to-video prompt is much shorter:
Slow dolly in, subject turns head slightly toward camera, hair moves in
the breeze, warm side light, natural handheld micro-movement, subtle
film grain.
The practical rule
If consistency across shots matters, generate a still first, approve it, then animate it. Text-to-video is for discovery; image-to-video is for production. Mixing the two modes inside one project without a reference image is the single most common cause of scenes that do not cut together.
Character and Style Consistency Across Shots
Consistency is where beginner projects fall apart. Shot one has a woman in a red coat, shot two has a woman in a maroon jacket, shot three has a different face entirely. The fix is procedural, not magical.
Write a character lock. Create a reusable block of text describing your subject in fixed terms: age range, hair, build, wardrobe, and one distinctive feature. Paste that identical block into every prompt that features the character. Do not paraphrase it. Small wording changes produce visible identity drift.
Anchor with a still. Approve one image of the character and reuse it as the starting frame for every shot in which they appear. This does more for consistency than any wording trick.
Write a style lock. Create a second reusable block covering palette, lighting quality, lens, and rendering style. Append it to every prompt in the project. This is what makes separate clips feel like scenes from one film.
Control continuity details deliberately. Time of day, weather, wardrobe state, and props should either stay identical or change for a clear narrative reason. Audiences forgive a lot, but they notice a coat that changes color between cuts.
Keep a prompt log. Store every approved prompt alongside the settings you used. When a later shot needs to match, you copy instead of guessing. A simple spreadsheet with columns for shot number, mode, prompt, and notes is enough.
A Repeatable Iteration Workflow
Professional-looking results come from repetition, not from one perfect prompt. Here is a loop that scales from a single clip to a full scene.
- Write the shot on paper first. One sentence describing what the audience needs to see and feel. If you cannot write that sentence, the prompt will not fix it.
- Assemble the six blocks. Subject, action, setting, camera, light and color, style. Keep it under roughly sixty words for a first attempt.
- Generate a low-cost draft set. Produce three or four variations with small changes — different camera movement, different lighting direction — rather than one attempt with everything cranked up.
- Judge against the sentence from step one. Not against how impressive the clip looks. Impressive but off-brief is a discard.
- Change one variable at a time. If the framing is right but the light is wrong, keep every word about framing and edit only the lighting clause. Changing three things at once teaches you nothing.
- Lock and document. Once a shot works, save the prompt, the seed if the tool exposes one, and the reference image. That combination is your asset, not the clip alone.
- Move to adjacent shots. Reuse the style lock and character lock. Only the action, setting, and camera clauses should change.
Expect the first shot of a project to take the longest. After the locks exist, subsequent shots often take a fraction of the time.
Common Mistakes and How to Fix Them
Overloading a single prompt. Ten ideas in one prompt produces a confused clip. Fix: one action per clip, one camera move per clip. Build sequences from multiple shots.
Vague emotional adjectives. "Beautiful," "epic," and "amazing" carry almost no visual information. Fix: translate the feeling into concrete choices — low angle, backlight, slow push-in, muted palette.
Contradictory camera instructions. "Static camera with a sweeping orbit" cannot both be true. Fix: choose one primary movement.
Ignoring duration. A prompt describing a complex three-part action will not fit into a short clip. Fix: match the complexity of the action to the length of the output.
Changing too many variables between attempts. You lose the ability to learn from the result. Fix: single-variable iteration.
Fighting the model's strengths. Some tools handle photoreal humans well and stylized animation poorly, or vice versa. Fix: test one short clip per style category before committing a project to a tool.
Forgetting negatives. If unwanted artifacts keep appearing — text overlays, distorted hands, extra limbs — state what you do not want, or use the tool's negative prompt field. Models increasingly listen to explicit exclusions.
Practice Drills for Your First Week
Drill one: the same scene, six ways. Take one subject and setting. Write six prompts that change only the camera or lighting. Compare the results side by side. This trains your eye faster than any tutorial.
Drill two: noun specificity. Write the same shot with a generic subject and with a highly specific subject. Note how much identity detail survives.
Drill three: motion only. Using a fixed still as input, generate five clips that differ only in movement description. Learn which movement verbs your tool respects.
Drill four: the 15-second scene. Build a three-shot sequence with a shared style lock. Cut them together. Continuity problems become obvious in the edit.
Drill five: reverse engineering. Find a short clip you admire and write the prompt that would produce it. You will learn more from this than from generating randomly.
Choosing Tools Without Tool-Hopping
Every video model has a temperament. Some favor photoreal human motion, some excel at stylized worlds, some are strongest at animating existing stills, and some are best for quick concept iteration. Rather than chasing every new release, evaluate tools against three questions.
First, does it handle your primary subject type well? Test with one clip, not a whole project. Second, does it accept reference images and maintain consistency between them? If yes, it can carry a multi-shot scene. Third, how much of your prompt does it actually honor — camera, lighting, and exclusions? A tool that ignores half your prompt costs more time than it saves.
Pick two tools: one for exploration, one for production. Learn them deeply enough that you can predict output before you generate it. Predictability is worth more than novelty, because it turns a lucky accident into a repeatable process.
Frequently Asked Questions
How long should a video prompt be?
For most models, thirty to eighty words is the practical sweet spot. Shorter prompts leave key decisions to chance; much longer prompts dilute the important clauses and increase the chance of internal contradictions. Start short, then add specificity where the output was wrong.
Do keywords matter as much as they used to?
Less than they did. Modern video models interpret natural, descriptive sentences well, so a grammatically plain sentence with precise visual information often beats a comma-separated keyword pile. That said, established technical terms like "dolly in" or "golden hour" still function almost as keywords, because they map to very specific visual patterns in training data.
Why does my result look nothing like my prompt?
Usually one of four reasons: the prompt contained contradictions, the action was too complex for the clip length, a critical detail was buried at the end, or the model simply handles that subject style poorly. Regenerate with a simplified prompt that keeps only the essential blocks, then add detail back one clause at a time.
Should I write prompts in English?
English remains the most reliable option for most video models because it dominates their training data, especially for camera and lighting terminology. If you are more comfortable in another language, write your creative notes in that language and translate the final prompt. Many tools accept multiple languages, but technical vocabulary can drift in translation.
How many variations should I generate per shot?
Three to five for a first attempt, with deliberate single-variable changes. Generating twenty random variations is not iteration; it is gambling. You want each attempt to teach you something about how the model interprets a specific clause.
Can I reuse the same prompt across different video tools?
The structure transfers well, but the wording usually needs adjustment. Some models respond to film terminology, others to plain descriptive phrasing, and some have their own preferred prompt templates. Keep your six blocks and your locks, then re-tune the phrasing per tool.
What about seeds and settings?
If your tool exposes a seed, save the seed alongside the prompt for any shot you plan to revisit or extend. Settings such as motion strength, aspect ratio, and resolution affect the result as much as wording, so document them together.
Do I need to learn to edit video too?
It helps enormously. A rough cut reveals continuity errors and pacing problems that are invisible in isolated clips. Even basic cutting and audio work will make an AI-generated sequence feel far more intentional.
Where to Go From Here
Prompt design for AI video is a craft with a short learning curve and a long mastery curve. The beginners who improve fastest are not the ones with access to the most models — they are the ones who write down what they did, change one thing at a time, and keep reusable locks for characters and style.
Start with one scene, one subject, and one camera movement. Generate six variations. Pick the best. Then build a second shot that matches it. The moment two clips cut together seamlessly, you have stopped experimenting with AI video and started directing it.
