How to Write AI Video Prompts That Look Real

The difference between AI video that looks like a glitchy dream and AI video that passes for real footage is rarely the tool. It's the prompt. Two creators can use the same generator with the same subject and get wildly different results — the one who understands prompting craft gets usable footage; the other gets melting faces and impossible physics.

This guide teaches that craft. It applies to every major text-to-video tool (Veo, Kling, Runway, and the rest), because they all respond to the same fundamentals: a clear subject, deliberate motion, camera language, lighting, and knowing what to exclude.

Why Most AI Video Prompts Fail

Weak prompts fail in predictable ways:

• Too vague. "A woman walking in a city" gives the model a thousand interpretations, and it picks the most generic one.
• Too crowded. Five subjects, three actions, two camera moves, and a lighting change in one prompt — the model drops half of it and mangles the rest.
• No camera language. Real footage has a camera. If you don't specify one, you get the model's default "drifty" look.
• No motion thinking. A prompt that describes a still image produces video where nothing meaningful happens — or worse, where the model invents motion you didn't want.

The fix is a simple mental model: describe a shot, not a scene. Think like a cinematographer framing 3–5 seconds of film.

The 5 Building Blocks of a Realistic Prompt

Every strong prompt covers these five elements, in roughly this order:

1. Subject — Be Absurdly Specific

Generic subjects get generic results. Replace every vague noun with a concrete one:

• Weak: "a woman drinking coffee"
• Strong: "a woman in her early 30s with curly dark hair, wearing a cream knit sweater, holding a ceramic mug with both hands"

Include age cues, clothing, and one distinguishing detail. You don't need a novel — one precise sentence beats three vague ones.

2. Motion — Describe Exactly What Moves

This is where most prompts go wrong. The model needs to know what moves, how, and how much:

• Weak: "a dog running"
• Strong: "a golden retriever running toward camera across a grassy field, ears bouncing, tongue out, kicking up small clumps of grass"

Rules of thumb:

• One primary motion per prompt. A subject can walk or wave, not walk while waving while the camera orbits.
• Keep motion moderate. Extreme action (backflips, high-speed chases) is where physics breaks most often. Start with walking, turning, reaching, pouring — ordinary motion renders most reliably.
• Specify direction relative to camera: "toward camera," "left to right," "away from camera." This alone fixes half of all weird outputs.

3. Camera — Speak Like a Cinematographer

Camera language is the single highest-leverage addition to a prompt. Learn these terms and use them:

• Shot size: extreme close-up, close-up, medium shot, wide shot, aerial
• Angle: eye-level, low angle, high angle, overhead
• Movement: static, slow push-in, slow pull-back, lateral tracking, gentle pan
• Lens feel: "shot on 35mm," "shallow depth of field," "slight handheld shake"

Example upgrade:

• Before: "a bakery interior with fresh bread"
• After: "slow push-in on a rustic bakery interior at dawn, warm light through windows, loaves of bread on wooden shelves, shallow depth of field, slight handheld feel, 35mm film look"

Static or slow-moving cameras produce the most reliable results. Save the dramatic crane shots for after you've mastered the basics.

4. Lighting — The Realism Cheat Code

Lighting does more for perceived realism than almost anything else. Always include it:

• Time of day: golden hour, blue hour, midday sun, overcast
• Light quality: soft diffused light, harsh shadows, warm practical lamps, neon glow
• Mood words: "cinematic," "natural," "documentary-style" — these steer the model toward photographic rather than illustrated looks

"Golden hour sunlight raking across the scene, long soft shadows" will improve nearly any outdoor prompt. "Soft window light, gentle fill" does the same indoors.

5. Style Anchors — photographic, Not "Cinematic 8K"

End prompts with a style anchor, but choose honest ones:

• Good: "photorealistic," "documentary footage," "shot on 35mm film," "natural colors"
• Avoid stacking: "ultra hyper mega realistic 8k unreal engine cinematic masterpiece" — this word salad doesn't help and can actually push outputs toward a glossy, artificial look.

One or two anchors. "Photorealistic, natural lighting" is enough.

Negative Prompts: What to Exclude

Many tools support negative prompts — things you explicitly don't want. Use them to preempt the most common AI artifacts:

deformed hands, extra fingers, morphing faces, warping background,
flickering, oversaturated colors, cartoon look, watermark, text, blurry

A reusable negative prompt you can paste into any tool that supports it:

extra limbs, deformed hands, morphing face, unstable background,
flickering lights, oversaturated, plastic skin, cartoon, illustration,
watermark, logo, text overlay, blurry, low quality

Not every tool exposes negative prompts, but when available, they're free quality.

Putting It Together: 3 Worked Examples

Example 1 — Product shot (skincare serum):

Extreme close-up of a frosted glass serum bottle on a marble bathroom
counter, morning window light creating soft highlights on the glass,
a single water droplet sliding slowly down the bottle, shallow depth
of field, background softly blurred towels, photorealistic, natural colors.
Static camera.

Why it works: one subject, one small motion (the droplet), specified light, static camera, photographic anchor.

Example 2 — Lifestyle b-roll (coffee shop):

Medium shot, eye-level, of a barista pouring latte art in a busy
specialty coffee shop, slow lateral camera drift, warm pendant lights
glowing in the background, shallow depth of field keeping the cup sharp,
documentary footage feel, natural motion.

Why it works: single action, gentle camera move, lighting and depth specified, "documentary" steers away from glossy artificiality.

Example 3 — UGC-style clip (fitness product):

Handheld-style medium shot of a woman in athletic wear doing a kettlebell
swing in a bright home gym, natural window light, slight camera shake for
authenticity, realistic sweat and effort on her face, vertical 9:16 framing,
looks like phone footage, photorealistic.

Why it works: it deliberately asks for imperfection — slight shake, phone-footage feel — which reads as authentic for UGC-style content.

Iterating: The 3-Pass Method

Don't expect the perfect clip on the first generation. Professionals iterate:

1. Pass 1 — Composition. Generate with your full prompt. Judge only the framing, subject, and lighting. Ignore small glitches.
2. Pass 2 — Motion fix. If the composition is right but motion is off, simplify the motion description and regenerate. Often removing one motion verb fixes everything.
3. Pass 3 — Polish. If it's 90% there, try small tweaks: adjust the camera move, strengthen the negative prompt, or change the style anchor.

Budget 3–5 generations per final clip. Anyone promising one-shot perfection is selling something.

Mistakes That Scream "AI"

Watch for these in your outputs and prompt against them:

• Morphing backgrounds. Walls, crowds, or foliage that subtly shift and breathe. Fix: "stable background" in the prompt, simpler scenes.
• Plastic skin. Over-smooth faces. Fix: "natural skin texture" and avoid stacking "beautiful/perfect/flawless."
• Impossible hands. Still the classic failure. Fix: keep hands simple, small in frame, or out of focus; use negative prompts.
• The AI glide. Subjects that move with eerie smoothness. Fix: "natural motion," moderate speeds, slight handheld camera feel.
• Over-polish. Everything too perfect, too golden, too symmetrical. Fix: ask for ordinary details — "slightly worn," "lived-in," "natural imperfections."