Cinematic AI Video Prompts: A Step-by-Step Formula

TL;DR: Amateur AI clips come from underspecified prompts. Name the camera move, the lens, the light and the grade, and the model stops guessing. The order that works: subject, setting, action, camera, lens, lighting, mood, film-look, then the technical flags. Keep every prompt free of logos and readable screens — that one rule improves quality and makes the footage reusable.

This is post #3 of the AI Video Automation series. The prompts below are the ones that go in the prompt column of the shot list the batch runner walks.

Why most AI clips look amateur

Almost always the same cause: the prompt names a subject and stops. "A mechanic in a workshop." The model then has to invent the camera position, the movement, the lens, the light and the colour — and it invents all five badly, because there is no signal telling it what to do.

A cinematic prompt removes the guessing by stating what a cinematographer would decide before rolling. That is the entire technique. Everything below is detail on how to say it.

The prompt anatomy

Write the elements in this order. Models weight earlier tokens more heavily, so subject and setting lead, and technical flags trail:

[subject] + [setting] + [action] +
[camera movement] + [lens/framing] +
[lighting] + [mood] + [film-look] +
[technical: aspect ratio, duration, motion intensity, seed]

A complete example that renders cleanly:

A confident mechanic in clean overalls in a workshop, arms crossed,
slight smile, looking just off-camera — slow push-in, 50mm, shallow
depth of field — soft directional window light — calm, assured mood —
cinematic teal-and-amber grade, photorealistic — 9:16, 5s, motion: low

And a real establishing shot from a production b-roll bank, following the same order:

Wide interior of a busy modern auto workshop, two cars raised on
hydraulic lifts, tools neatly arranged, shafts of warm afternoon
sunlight cutting through a large roller door, faint dust in the air,
slow steady dolly-in, 35mm, shallow depth of field, cinematic
teal-and-amber grade, photorealistic, motion: low

Note how much of that prompt is not the subject. "Shafts of warm afternoon sunlight," "faint dust in the air," "slow steady dolly-in" — those clauses are doing the cinematic work.

The load-bearing lines

If you only add four things to a bare prompt, add these:

• Camera movement — "slow push-in", "gentle handheld drift", "steady dolly-in". This single line is the biggest quality jump available. A static prompt produces a static, lifeless clip.
• Lighting — "warm afternoon light through a window", "soft directional light", "golden hour". Light is what actually reads as cinematic; grade alone does not fake it.
• Lens and framing — "35mm, shallow depth of field", "macro close-up". Gives depth and tells the model where to put focus.
• Film-look — "cinematic teal-and-amber grade, photorealistic". Anchors both colour and realism, and stops the output drifting toward illustration.

Then one dial: motion intensity. Keep it low for people and subtle shots, medium for action. High motion on a human face warps it — this is the single most reliable way to ruin an otherwise good clip.

Model selection is part of the prompt

Which model you send a shot to matters as much as the wording, and the two decisions are linked.

• Kling 3.0 — strongest for realistic people and controlled camera moves. Use it for hero shots. Keep the action to one clear beat; multi-action prompts wobble in a five-second clip.
• Seedance 2.0 — faster and cheaper, excellent for abstract motion, gradients and background plates. Reserve it for non-hero shots.
• Duration — 5–6 second clips render more reliably and cost fewer credits than long ones. You cut them down in the edit anyway.

That advice is not theoretical. Across a real 57-shot library the split came out as:

Roughly a 3:1 ratio of hero to background. Sending those 12 background plates to Kling instead would have changed nothing visible and cost noticeably more — the kind of decision that only shows up when you price the library before running it.

The brand-neutral rule

Keep every prompt free of logos, product names and readable screens. This is worth stating as a rule because it wins twice.

It improves quality. Current video models render text as garbled pseudo-lettering. Ask for a screen with a UI and you get something that looks broken. So write the screen out of the shot instead:

Close-up of a technician's hands holding a rugged tablet in a workshop,
tapping and swiping, screen deliberately out of focus / glare so no UI
is readable, grease-flecked fingers, soft window light, macro, shallow
depth of field, cinematic, photorealistic, motion: low

"Screen deliberately out of focus / glare so no UI is readable" is doing real work there. You composite the actual interface in the edit, where it renders perfectly because it is a real screenshot.

It makes the footage reusable. A clip with no branding in it can serve any project. A generated library built this way keeps its value after the campaign it was made for is finished — which changes the economics of generating it at all.

Common mistakes

• No camera or lighting direction. The single biggest cause of flat, amateur output.
• Readable text or logos. Models render text as garbled shapes. Composite it later.
• Too much action in five seconds. One clear beat per short clip; chain beats in the edit.
• Generating once. Run 3–5 seeds, keep the best, discard the rest. First takes are rarely the best takes.
• High motion on faces. Warping is almost guaranteed. Drop to low.
• Requesting on-screen UI. Ask for glare or defocus instead.

Seeds are how you get consistency

A seed makes a generation reproducible — same prompt plus same seed gives you the same clip. Two consequences worth knowing.

First, running 3–5 seeds per hero shot is not waste; it is how you get a choice. The model's output varies enough between seeds that the third attempt is often materially better than the first, and you cannot tell which without generating them.

Second, once you find a seed that works for a subject, reuse it across related shots. That is the closest thing to character consistency these models offer, and it costs nothing extra.

Worth being deliberate about, though: seeds multiply spend. In a batch, video shots get multiple seeds and stills usually get one — a still that is nearly right can be fixed in an image editor, where a video clip cannot.

Putting it together

Apply the anatomy across a whole shot list and the output stops being a lucky dip. Every row carries the same structure, the same grade language and the same motion discipline, so the clips cut together as if they were shot by one person on one day — which is the actual goal.

The next question is what all that generation costs, and it is not what the pricing page suggests.

Related posts

• How to Automate AI Video Creation with Claude & Higgsfield — the pipeline these prompts feed
• Higgsfield Credits vs Unlimited: Real Per-Clip Cost Breakdown — what each of those seeds actually costs
• Automate Higgsfield with a Python Batch Queue Runner — running the whole list hands-off
• Browse the AI Video Automation series