Every generation of video models arrives with the same promise: the tool will do the thinking for you. It never quite works out that way. Rendering quality improves quickly, but the distance between a usable clip and an unusable one is still mostly explained by the prompt. Two creators using the same model on the same day will get wildly different results, and the difference is rarely luck.
A prompt is not a wish. It is a specification. The more of the specification you supply, the less the model has to invent on your behalf, and invention is where things break: faces drift, props teleport, the camera does something you never asked for, and the light changes between shots that are supposed to match.
The most useful mental shift is to stop thinking of yourself as someone describing a picture and start thinking of yourself as someone writing a shot for a crew. A camera operator needs to know where to stand. A gaffer needs to know where the light comes from. An editor needs to know how long the beat lasts. A video model needs the same information, expressed in a compact, unambiguous string.
Three failure modes account for most bad output:
Everything below is about reducing those three problems while keeping the creative idea intact.
Treat a video prompt as a stack of decisions rather than a sentence. The order is not magic, but keeping the order stable makes you faster and makes troubleshooting possible, because when something goes wrong you know which layer to blame.
The subject layer carries identity and behaviour. Be specific about age range, build, wardrobe, and emotional register, then commit to one verb that describes what changes during the shot.
Weak: a woman walking down a street.
Stronger: a woman in her late thirties in a charcoal wool coat walks toward the camera, hands in pockets, eyes down, and lifts her head at the three-second mark.
The second version gives the model an arc. It knows the shot has a beginning, a middle, and a change. That change is what makes a clip feel directed rather than sampled.
Environment is not decoration; it is the constraint that makes everything else believable. Name the location, the time of day, the weather, and the direction and quality of light. Practical sources matter more than you think, because they anchor the colour of the frame.
Example: a dim parking garage at dusk, cool blue ambient light from a stairwell behind the subject, a single warm sodium fixture overhead casting a hard pool of light on wet concrete.
That sentence does more work than five adjectives about mood.
This is the layer most beginners skip, and it is the layer professionals reach for first. Specify four things: framing, height, distance, and movement.
Add an equivalent focal length when you care about distortion and compression: a wide lens exaggerates space and makes movement read faster, a longer lens flattens faces and background and feels more intimate.
Video prompts have a dimension that image prompts do not: time. State the duration you are aiming for, the speed of the main motion, and what should stay still. Stillness is a tool. A locked background with one moving element reads as intentional; everything moving at once reads as chaos.
Useful phrasing includes: over four seconds, gradually, within the first second, holding steady, slow enough that the motion is barely perceptible.
Style should describe a look, not a person. Instead of leaning on a living filmmaker's name, describe what you actually see: documentary realism, desaturated teal shadows, soft highlight rolloff, 16mm grain, high-contrast neon night, muted pastel daylight. Descriptive style survives model updates and keeps you out of policy trouble.
Negatives are guardrails, not a second creative brief. Keep them short and specific: no on-screen text, no captions, no logos, no extra limbs, no face morphing, no camera shake, no scene cuts, no zoom. Ten focused negatives beat thirty vague ones.
Assembled, a single prompt might read:
Medium shot, chest height, 50mm look. A woman in her late thirties in a
charcoal wool coat walks toward the camera, hands in pockets, eyes down,
lifting her head at the three-second mark. Dim parking garage at dusk,
cool blue ambient light from a stairwell behind her, one warm sodium
fixture overhead. Slow push in, locked background. Documentary realism,
muted teal shadows, soft highlight rolloff. Four seconds. No text, no
logos, no camera shake, no scene cuts.
That is roughly ninety words. It is specific enough to direct and short enough to stay legible to the model.
A scene is not one shot. Take a simple dramatic beat: a woman confronts her brother in a parking garage about money he never repaid. That beat might become five shots:
Each shot has a job. The establishing shot gives geography. The tracking shot creates approach and tension. The close-up delivers the emotional turn. The over-the-shoulder establishes the power dynamic. The two-shot resolves it. If a shot has no job, cut it before you generate it, not after.
A shot card is a one-line-per-field record you keep next to the prompt. It turns prompt writing into a repeatable process instead of a memory test.
Shot ID: 03
Purpose: emotional turn
Framing: close-up, eye level, subject fills frame
Action: she stops, jaw tightens, eyes lift to off-screen right
Camera: locked off, no movement
Light: overhead sodium, hard shadow under brow
Duration: 2.5 seconds
Anchors: charcoal wool coat, dark hair tied back, silver hoop earrings
Negatives: no text, no shake, no cut
When a clip fails, you compare the card with the output and you can usually name the missing field in seconds.
A single beautiful clip is a demo. A sequence of clips that look like the same film is a deliverable. Consistency comes from repeating exact language and reusing visual anchors, not from hoping the model remembers.
Write a short block of subject descriptors and paste it verbatim into every prompt featuring that subject. The moment you paraphrase, features drift: the coat becomes a jacket, the hair changes length, the age shifts by a decade.
Keep anchors to three to five visual facts. More than that and they start competing for attention.
If your tool supports image-to-video or subject referencing, generate one approved still of each main character and each key location, then use those stills as the starting frame for every shot in that setup. The still is now your source of truth. Prompts refine; the reference defines.
Props behave like characters. If a red envelope appears in shot two, it must be the same red envelope in shot nine. Lock the phrasing: a creased red paper envelope with a broken wax seal. Never shorten it to the envelope.
Run this before generating a batch:
The vocabulary of cinematography mostly translates, with a few traps.
Reliable terms include locked off, slow push in, dolly left and right, tracking shot, handheld follow, crane up, rack focus, close-up, wide shot, over-the-shoulder, low angle, high angle, and dutch tilt.
Traps worth knowing:
A practical rule: one camera instruction per shot. If you need two, split the shot.
Storyboarding with AI is faster when you separate stills from motion. For each shot card, generate the strongest single moment as a still. Approve the stills as a set before animating anything, because stills are cheap and clips are expensive in time.
Lay the approved stills in a row and look at them as a sequence. You will spot problems immediately: two shots that are visually identical, a jump in wardrobe, a light direction that flips between frames.
When you animate a keyframe, keep the motion prompt modest. The most common mistake is adding a dramatic camera move to a shot that already has a strong composition. Small motions read as cinematic. Big motions read as unstable.
Cut length is a creative decision, not an afterthought. Rough guides:
If your generated clips are all the same length, vary them in the edit. Uniform shot length is the fastest way to make AI footage feel like AI footage.
Generate three test shots of the same subject in the same location. Do not move on until all three match in light, wardrobe, and lens. This is the cheapest quality control you will ever do.
Write shot cards for the whole scene before generating a single clip. Ordering the list forces you to notice missing coverage: no establishing shot, no reaction shot, no closing image.
Produce stills for every card. Reject hard. A weak still almost never becomes a strong clip.
Animate two or three shots at a time and review them together. Reviewing in batches makes continuity errors obvious; reviewing one clip at a time hides them.
Edit the sequence first, then re-generate only the shots that visibly fail. Editing tells you which failures actually matter, and it is common for a shot that looked weak in isolation to work perfectly in context.
Save the exact prompt, the model, the settings, and a one-line verdict for every approved shot. Your log becomes the most valuable asset in your workflow, because it lets you reproduce a look months later instead of reinventing it.
| Mistake | Symptom | Fix |
|---|---|---|
| Vague subject | Generic face, drifting identity | Add three to five concrete descriptors |
| Contradictory camera | Muddy movement | One camera instruction per shot |
| No duration | Clip ends mid-action | State seconds and the beat that lands |
| Paraphrased anchors | Wardrobe and hair change | Paste the anchor block verbatim |
| Style by name | Inconsistent results | Describe the look instead |
| Too many negatives | Stiff, flat output | Cut to the five that matter |
| No still approval | Expensive re-renders | Approve keyframes as a set first |
One more subtle mistake: generating dozens of variants without criteria. Decide in advance what makes a take usable, such as stable face, correct wardrobe, usable motion, no artifacts. Without criteria you will keep the last one you watched rather than the best one.
Establishing shot. Wide shot, high angle, 24mm look, locked off. Location description, time of day, light direction. No people, or one small figure in the frame. Four to five seconds. Negative list: no text, no camera shake.
Dialogue coverage. Medium close-up, eye level, 50mm look, locked off. Subject anchor block, single facial action, background soft. Two to three seconds. Keep framing identical across both sides of the exchange so the cut feels clean.
Movement beat. Tracking shot, chest height, 35mm look, steady follow. Subject anchor, one continuous action with a clear endpoint, environment consistent with the preceding shot. Three to four seconds.
Insert shot. Extreme close-up, slight top-down angle, 85mm look, shallow depth of field. Object description with exact colour and wear, hands entering frame if relevant. One to two seconds. Insert shots are the cheapest way to repair continuity problems in an edit.
How long should a video prompt be?
Long enough to cover the six layers, short enough that nothing competes. For most shots, seventy to one hundred twenty words works well. If you pass two hundred words, you are probably writing a scene description rather than a shot.
Should I write prompts in one language and use them in another?
Pick one language and stay consistent, because the model weights phrasing and vocabulary. Mixing languages within one prompt often produces inconsistent style between shots.
Why does my character change between shots?
Almost always because the subject descriptors were paraphrased. Build a fixed anchor block and paste it without edits. If your tool supports reference images, use an approved still as the starting frame for every shot in that setup.
Do negatives actually help?
Yes, but only when they are specific. No on-screen text, no logos, no camera shake, and no scene cuts all do measurable work. Vague negatives such as no bad quality rarely change the output.
How do I avoid the AI look?
Vary shot length, keep camera moves small, add one imperfection per scene such as a flickering practical light or a slight handheld sway, and avoid the same shallow depth-of-field close-up for every beat.
Should I storyboard before or after writing prompts?
Write the shot list first, then generate keyframes, then animate. Skipping the shot list is the single biggest cause of unusable sequences, because you end up with attractive clips that do not cut together.
How many takes should I generate per shot?
Three is a reasonable default: one safe, one with a slightly different camera angle, and one experiment. More than that and you are searching rather than directing.

Alexander
Build a reliable AI transcription pipeline for video: model choice, speaker labels, timestamps, quality checks, and workflow automation.
Learn a practical workflow for replacing video backgrounds in After Effects using AI mattes, cleanup, compositing, and motion matching for believable shots.
A practical guide for Luxembourg video agencies building cross-border AI workflows: planning, localization, visual consistency, review, and delivery.