Most AI video prompts fail for the same reason most recipes fail when someone just lists ingredients with no quantities or steps: they’re a pile of descriptive words with no structure. “Cinematic, ultra detailed, 8k, dramatic lighting, man running” tells the model what vibes you like. It doesn’t tell it what to actually put on screen.
This guide gives you a prompting system — a repeatable structure for writing prompts that produce directable, consistent results. It works across tools; only the syntax details change.
The core idea: prompts are shot descriptions
Stop writing prompts like search queries. Start writing them like you’d describe a shot to a cinematographer. A working video prompt answers six questions, in this order:
- Subject — who or what is on screen?
- Action — what is happening?
- Camera — angle and movement?
- Setting — where, and what does it look like?
- Lighting — what kind of light, from where?
- Mood/style — what should it feel like?
The template
Here’s the structure as a fill-in template:
[Camera angle + movement] of [subject with key visual details], [action in present tense], [setting with 2–3 concrete details], [lighting description], [mood/style keywords].
Example — weak version:
“cinematic man running city night cool”
Example — structured version:
“Low tracking shot of a courier in a yellow rain jacket sprinting down a rain-slicked alley at night, neon signs reflecting in puddles, harsh magenta and cyan practical lighting, tense urgent mood, photorealistic.”
The second prompt isn’t longer for the sake of it. Every clause answers one of the six questions, so the model has no gaps to fill with random choices.
Breaking down each element
1. Subject: be specific about what persists
Name the visual facts that must stay constant: “a courier in a yellow rain jacket” beats “a man” because it gives the model anchor details. If your tool supports reference images, tag them here — this is where consistency lives. (See character consistency techniques.)
2. Action: one clear verb, present tense
“Sprinting,” “pouring,” “turning to face the camera.” One action per prompt. AI video handles a single clear action far better than a sequence (“runs, then jumps, then waves”) — multi-action prompts are the top cause of morphing, melting messes. If the story needs three actions, that’s three shots.
3. Camera: the highest-leverage element
Camera language is the fastest way to make AI footage look intentional. Learn and use real terms: Angles: eye-level, low angle, high angle, bird’s-eye, Dutch angle. Movement: static, pan, tilt, dolly in/out, tracking, crane up, handheld. Framing: extreme close-up, close-up, medium shot, wide shot, aerial.
“Static medium shot” vs “slow dolly-in close-up” produces completely different footage from the same subject. Most beginners leave camera out entirely and get generic results. Don’t.
4. Setting: concrete details, not adjectives
“Rain-slicked alley, neon signs, puddles” beats “cool city street.” Models render nouns better than adjectives. Give it things to draw.
5. Lighting: describe the light, not just “good lighting”
“Harsh magenta and cyan practical lighting” beats “dramatic lighting.” Name the source and quality: golden-hour sunlight, flickering fluorescent, soft window light, moonlight. Lighting is half of cinematography; it deserves a full clause.
6. Mood/style: short and last
Two or three keywords max: “tense, urgent, photorealistic.” Style words work best as seasoning at the end, not as the whole prompt.
What to leave out
- Resolution buzzwords (“8k, ultra HD, 16k”) — the model renders at its own resolution; these words add noise, not pixels.
- Contradictory directions (“dark but bright, realistic cartoon”) — pick one.
- Multiple actions or scenes — one shot, one action.
- Dialogue or text — most video models still mangle on-screen text and lip-sync; keep text out of prompts unless your tool specifically handles it.
Negative prompts: saying what you don’t want
Many tools support negative prompts — things to exclude. Use them for the model’s most common failure modes:
Negative: morphing faces, extra limbs, warping background, flickering, watermark, text overlay.
Keep negatives short and focused on artifacts, not aesthetics. For more, see the negative prompt guide.
Iterating: change one thing at a time
When a generation misses, resist the urge to rewrite the whole prompt. Change one element — the camera move, the lighting, one adjective — and regenerate. This is debugging, and like all debugging, it works best with one variable at a time. Keep a simple log: prompt version, what changed, what improved. After a few projects, you’ll have a personal library of what works.
Building your reusable template
Once the structure clicks, save it as a template you reuse:
[CAMERA] of [SUBJECT + anchor details], [ONE ACTION], [SETTING with concrete nouns], [LIGHTING], [2–3 style words]. Negative: [artifacts to avoid].
Fill it per shot from your storyboard, and prompting stops being a creative gamble. It becomes what it should be: the mechanical step between planning and generating.
The honest truth about prompting
No prompt system guarantees a perfect clip on the first try. Models have quirks, and part of the craft is learning your chosen tool’s personality — what it renders well, where it struggles. But structure beats vibes every time: a systematic prompt that’s 80% right on generation one, fixed with one targeted tweak, will always beat fifty rolls of keyword dice.