Writing Prompts That Produce Usable Video

The instinct when a generated shot comes out wrong is to add adjectives. Cinematic, beautiful, highly detailed, professional. The next attempt is more polished and still not the shot you needed, because none of those words told the model anything about what should be in the frame.

Prompts fail in a consistent way: they specify the subject and leave everything else — the camera, the movement, the light, the duration — for the model to decide. It does decide, differently every time, which is why results feel random. This guide covers a five-part structure that closes those gaps, what each part actually changes, and a set of before-and-after rewrites.

Five Things a Prompt Should Settle

Camera is the one people leave out most and the one that changes the result most. An unspecified prompt returns a medium shot at eye level — the most neutral option and therefore the least useful, because an edit needs wides and close details rather than more mid-shots.

Describe the Camera Like a Shot List

The vocabulary that works is the vocabulary of a shot list, not of a mood board.

Shot size: wide, medium, close, extreme close-up. Angle: eye level, low, high, overhead. Movement: static, slow push in, slow pull back, pan left, handheld. Lens character: shallow depth of field, wide angle, telephoto compression.

These terms work because they correspond to things that visibly differ in the training material. "Cinematic" does not correspond to anything specific; "shallow depth of field, slow push in, low angle" does.

Movement deserves a caution. Complex camera moves are among the least reliable things to request, and a slow push or a static frame will come out cleaner than an elaborate crane move. Where the shot needs motion, asking for less usually produces more usable footage.

One Action, Not a Sequence

A prompt describing two events asks the model to produce a shot that contains a change, and change across time is where generation is weakest.

Weak:
A woman walks into a workshop, picks up a tool, and begins working
at a bench, cinematic lighting, highly detailed.

Better:
Medium shot, static camera. A woman in a canvas apron works at a
wooden bench, hands moving steadily. Late afternoon light from a
window on the left. Shallow depth of field. Four seconds.

The rewrite removes two of the three actions and specifies the frame instead. What comes back is a shot you can cut into a sequence — which is what you wanted from the original prompt, since walking in and picking up the tool were separate shots anyway.

This is the most common single improvement available. If a prompt contains the word "then", or two verbs joined by "and", it is describing an edit rather than a shot.

Light Is Nouns, Not Adjectives

Mood words are the weakest part of most prompts because they compress a physical description into a judgement.

Weak:   moody, dramatic, cinematic atmosphere
Better: single light source from a low window, deep shadows on the
        right side of the face, dust visible in the beam

The second version describes something a camera could record. It also gives you control: if the result is too dark you know which clause to change, whereas "moody" offers nothing to adjust.

The same applies to time of day and weather, which are among the most reliable prompt elements. "Overcast midday" and "low sun, long shadows" produce genuinely different and predictable images.

Before and After

Before: A beautiful mountain landscape, breathtaking, 4K, cinematic

After:  Wide static shot of a ridge line at dawn, mist filling the
        valley below, sun just clearing the far peaks. Slow drift
        of cloud. Six seconds.

Before: Product shot of a coffee machine, professional, studio quality

After:  Close-up, slow push in on a matte black espresso machine on
        a concrete counter. Steam rising from the group head. Soft
        light from above and behind, dark background. Three seconds.

Before: Someone typing on a laptop in a modern office, dynamic

After:  Extreme close-up of hands on a keyboard, shallow depth of
        field, screen glow from the front, office window blurred
        behind. Static camera. Two seconds.

Each rewrite does the same three things: names a specific shot size, replaces mood words with light description, and reduces the action to one continuous thing. None of them is longer than the original by much, which is the point — this is about specificity rather than volume.

Note the durations. Asking for short clips is a deliberate choice, because generated footage tends to lose coherence as it runs, and a two-second insert is what most b-roll actually needs.

Iterate on One Variable

When a result is close but wrong, the temptation is to rewrite the whole prompt. That makes it impossible to learn anything, because the next result differs in every respect.

The last step matters because prompt iteration has no natural end. Knowing which shot types tend to resist regardless of prompting saves the most time, and the reliability map by shot type is where to check before spending an afternoon on one frame.

What Prompts Cannot Reach

Some qualities do not respond to prompting at all, and recognising them saves the most time of anything here.

Exact likeness. No description produces a specific person or a specific product. Words can narrow the space and cannot pin it down, which is why anything that has to look like a particular thing starts from an image instead.

Legible text. Asking for a sign that reads something specific will produce something sign-shaped with letter-like marks. This has improved and remains unreliable, and the practical answer is to add text in the edit rather than to generate it.

Precise counts. Three people, five windows, two hands holding one object. Numbers are weakly enforced, and repeating the number in the prompt does less than expected. Where a count matters, frame so that it cannot be checked.

Long sequences of events. Already covered above, and worth restating as a hard limit rather than a preference: a prompt is a shot, not a scene.

For each of these, the answer is not a better prompt. It is a different approach — a reference image, a graphic added afterwards, a framing that avoids the problem. Recognising the category early is the difference between five minutes and an afternoon.

Prompt Vocabulary Is Tool-Specific

One warning about advice, including this. Every model responds differently to the same words, because each was trained on different material with different labelling.

That means prompt guidance transfers in structure and not in detail. The five-part shape works everywhere, because it is about specifying things that must be specified. The particular phrasing that produces a good result — whether a tool responds to "35mm", "wide angle", or neither — is a property of the tool and has to be learned on it.

The practical consequence is to distrust prompt lists found elsewhere, including collections of phrases that supposedly improve any output. Most are either tool-specific or superstition carried over from image generation. Your own file of prompts that worked, on your own tools, is worth more than all of them, and it is why building that file is the recommendation below.

It also means switching tools resets part of what you know. The structure survives; the vocabulary does not, and expecting otherwise is why a familiar prompt can produce unfamiliar results on a new model.

Prompts Are Not a Substitute for a Plan

A well-written prompt produces a good shot. It does not produce a sequence, because each generation is independent and nothing carries between them.

This is also why some formats sidestep prompting almost entirely. A structured explainer built from drawn elements, for instance, is authored rather than generated, and the sequencing that makes it work has nothing to do with prompt quality — the structure is set out in how a whiteboard explainer holds attention.

That is why prompting works best downstream of a shot list rather than instead of one. Knowing what each shot has to do — and which ones need to match each other — tells you what to prompt for and, more usefully, which shots not to generate at all. The case for doing that first is in storyboarding before you generate anything, and the techniques for holding a look together across shots are in consistency across generated shots.

Next Steps

Take a prompt that has been giving you inconsistent results and rewrite it as five clauses: subject, action, camera, light, duration. Generate a batch from each version and compare the spread rather than the best result.

Then start the prompt file. Every prompt that produced something usable, saved with the tool it was run on. Within a few projects that file will teach you more about your own tools than any general guidance, including this one. Related articles are gathered on the generation hub.