Most prompts fail before the second sentence. The model has already decided what the shot is about, who matters, and what to ignore — and if your subject shows up late, buried under style words and mood adjectives, the generation drifts.
Subject-first writing isn’t a trick. It’s just putting the who and the what where the model can use them, then letting everything else support that core.
If you’ve worked through the anatomy of a strong AI video prompt, you already know a prompt has to answer subject, action, camera, setting, light, and style. This article is about order. The same information, rearranged, can produce a clean shot or a confused one. In a storyboard-first workflow, this is natural — every panel is drawn subject-first. The prompt should inherit that order instead of scrambling it.
Models Read Prompts Like Headlines
Think of the first line of your prompt as a headline. A headline doesn’t open with adjectives. It names the actor and the act: “Vendor flips skewers over charcoal.” Everything after that is context.
Video models behave in a similar way in practice. Early tokens set the frame’s priorities — what object the model builds first, what motion it commits to, what it treats as background. Words placed later still matter, but they compete with an already-formed interpretation.
That’s why prompts that open with “cinematic, moody, atmospheric, beautiful…” often produce pretty fog and no clear subject. The model spent its strongest attention on the vibe and only learned later who was supposed to be in the fog. The result isn’t always bad. But it’s rarely directed. You get a mood in search of a scene.
Subject-first writing fixes this without adding length. You’re not writing more. You’re writing in the order the model can execute: subject and action first, environment second, camera and light third, style last. It’s the cheapest improvement most creators can make — costs zero credits to reorder sentences you were already going to write.
There’s a second benefit. When the subject is named clearly at the top, your own thinking gets clearer. You notice when a shot has two competing subjects. You notice when the action is vague. You notice when you’re asking for a close-up and a wide landscape in the same breath. The prompt becomes a small shot list. Not a wish.
The Subject Sentence: Who, Doing What, Where
Start with one sentence that could stand alone as a caption: who is on screen, what they’re doing right now, and where they are. That’s the whole job of the opening.
A weak opening: “A cinematic shot with warm lighting and a nostalgic feel of someone cooking at night.”
A subject-first opening: “A street vendor in a white apron flips skewers over a charcoal grill at a night market.”
The second version gives the model a person, a costume cue, an action, a tool, and a place in eleven content words. The first version gives it feelings. Feelings are hard to render. Skewers are easy.
Build the subject sentence with concrete nouns and one present-tense verb. “A woman in a red coat pauses mid-step” is stronger than “a beautiful woman elegantly walking.” The first describes a frame a storyboard artist could draw. The second describes a compliment. Models draw frames better than they take compliments.
Keep the “who” specific but not overloaded. Two to four defining details is usually right: age range, one clothing item, one distinguishing feature. “A man in his 60s, grey beard, faded blue work shirt” is a castable person. “A tall handsome charismatic mysterious man with piercing eyes and a charming smile” is five opinions fighting for one face. The model will pick two and drop the rest unpredictably.
The “where” belongs in the same sentence or the next one — still early. Location isn’t decoration. It tells the model what background to build and what light makes sense. “In a crowded night market” and “in an empty concrete parking garage” will change the same subject’s performance completely, even with identical action words.
Details In Descending Order Of Importance
After the subject sentence, add details in the order you’d defend them if credits were on the line. If the model could only keep three things, which three must survive?
That question produces a natural hierarchy:
- Identity details first: face, hair, wardrobe items that must not change.
- Action details second: direction, speed, start and end of the movement.
- Setting details third: the three concrete objects that sell the place.
- Camera and light fourth: framing, movement, named light source.
- Style last: one or two words that set the finish, not the content.
This is the same priority a storyboard panel uses. The panel draws the person large and clear, places them in a readable environment, and only then worries about polish. When you write in that order, the prompt and the board agree. When you reverse it — style, light, setting, person — the prompt argues with itself.
Descending order also helps when a model truncates attention. Not every token gets equal weight in a long prompt. If your critical wardrobe detail is word 58, it may lose to a throwaway adjective at word 12. Put the non-negotiables early enough that they can’t be crowded out. This matters even more when you care about character consistency across a series of shots: the identity line should appear in the same early position every time, worded the same way, so the model meets the same person at the start of every prompt.
A practical habit: read only the first sentence of your prompt out loud. If a stranger couldn’t picture the shot from that sentence alone, the sentence isn’t finished. Then read the first two sentences. By sentence two, the viewer should know where we are and what’s moving. If that takes until sentence four, move the sentences — don’t add more of them.
What Belongs At The End (And What Gets Ignored There)
The end of the prompt is for seasoning and constraints. Things that refine a shot that’s already described. Style anchors live here. So do finish notes: “photorealistic, natural skin texture” or “muted colour palette.” So do simple avoidances, phrased positively when the tool has no dedicated negative field: “clean background, no extra people.”
What doesn’t belong at the end is anything the shot can’t live without. If the character must hold a red umbrella, that umbrella can’t be a final afterthought after 40 words of atmosphere. If the camera must stay locked off, that instruction can’t trail behind three sentences of subject movement that imply a tracking shot. Late instructions are the first to be negotiated away.
The end is also a risky place for contradictions. Models resolve contradictions by compromise, and compromise in video looks like morphing. “Wide shot… tight close-up on her face” at the end of a long prompt doesn’t produce a clever transition. It produces a face that grows. If two instructions conflict, the earlier one usually wins the opening frames and the later one hijacks the ending frames — which is why so many clips start correctly and fall apart at second four.
Subject-first writing prevents most of this by forcing you to choose the shot before you describe it.
One more end-of-prompt trap: stacking quality bait. “8K, ultra-detailed, masterpiece, award-winning” at the end doesn’t raise resolution and rarely raises quality. It spends tokens that could’ve described the subject’s hands, the direction of light, or the end point of the action. Replace bait with facts. The model can use facts.
If you’re unsure how much total description is useful before returns diminish, the companion guide on how long an AI video prompt should be covers the length side of the same problem. Order first, then length.
Rebuilding A Real Prompt Subject-First
Here’s a prompt in the order many creators naturally write it. Mood first, subject last:
“Moody cinematic night scene with neon reflections and atmospheric rain, very emotional and beautiful, a girl standing in Tokyo feeling lonely and nostalgic, photorealistic 8K.”
Nothing in that prompt is unusable. It’s just arranged so the model has to guess the shot. Now rebuild it subject-first, keeping almost the same ingredients:
“A young woman in a red coat stands still on a rain-soaked Tokyo side street at night, facing camera, neon signs reflecting in puddles around her shoes, slow push-in, photorealistic.”
The rebuild did four things. It named the subject in the first five words. It replaced “feeling lonely” with a visible behaviour — standing still while the city moves. It turned “atmospheric rain” into a concrete surface the model can render: wet pavement and puddles. And it moved the camera instruction next to the action, where it can shape the movement instead of trailing behind it.
Notice what disappeared: “very emotional,” “beautiful,” “nostalgic,” “8K.” The shot didn’t get poorer. It got more filmable. Loneliness is now carried by stillness, scale, and a slow push-in — choices a viewer can actually see. That translation, from feeling word to visible choice, is most of prompt craft.
Try the same rebuild on one of your own recent prompts. Keep every fact, change only the order, and generate both versions with the same settings. In our testing across Seedance and Veo shots, the subject-first version usually wins on the first or second generation — not because the model got smarter, but because it stopped spending its best attention on adjectives.
A Subject-First Template For Any Shot
You don’t need a long template. You need a repeatable sentence order you can fill in under pressure. This one works for portraits, product shots, landscapes, and story beats:
“[Subject with 2–3 defining details] [one present-tense action] [in/at concrete location]. [Camera framing and movement]. [Named light source and direction]. [One or two style words].”
Filled in, that becomes: “A barista in a black apron wipes a ceramic cup behind a small café counter. Medium close-up, locked-off camera. Warm window light from the left, soft shadows. Photorealistic, natural colour.”
Four short sentences. Subject first, action attached to the subject, world attached to the action, and every later sentence in service of the first. If you want a fuller reusable system around this shape, the reusable prompt template on the site extends the same skeleton with slots for continuity lines and negative prompts.
Use the template until the order becomes automatic, then loosen it. The goal isn’t rigid grammar. The goal is a habit: when you open the prompt box, your hands type the person and the action before they type the vibe. Vibes are easy to add at the end. A missing subject is expensive to discover after generation.
The working rule is simple enough to tape above a monitor: who, doing what, where — then how we see it, then how it feels. Write in that order for your next ten shots and compare the discard rate with your last ten. Subject-first won’t rescue a bad idea or an impossible action, but it’ll stop good ideas from dissolving into mood soup before the model ever gets to them. In a medium where every generation costs time or credits, that’s the closest thing to a free upgrade you’ll find.