Prompting for video editors: what works and what doesn't

TL;DR

Write prompts like shot descriptions: subject + action + environment + style + camera, 15–40 words max. Use camera terms you already know (dolly in, shallow DOF, tracking shot) — they work. Specific and short beats long and vague every time.

A prompt is a shot description written for an AI instead of a camera operator. The skills you use in shot lists and creative briefs translate directly.

When to use this in your workflow

• Every text-to-video and text-to-image generation: The prompt is the primary input. Quality directly affects how many generations you need — and each one costs money.
• Image-to-video animations: Most image-to-video models accept an optional prompt to guide motion direction (“camera slowly pans right,” “the subject turns to face the camera”).
• Reducing wasted generations: A precise prompt means fewer retries. Vague prompts force you to pay for the AI to guess.

How it works in modelBridge

The prompt field appears at the top of every text-based model’s parameter panel. modelBridge sends your prompt directly to the model — no preprocessing or rewriting.

The prompt recipe

Structure every prompt as: subject + action + environment + style + camera.

A woman in a red coat walks through a rainy Tokyo street at night, neon reflections on wet pavement, shallow depth of field, handheld camera.

Breaking it down:

• Subject: A woman in a red coat
• Action: walks through
• Environment: rainy Tokyo street at night, neon reflections on wet pavement
• Style: (implied by the scene details)
• Camera: shallow depth of field, handheld camera

More examples

Slow-motion macro shot of rain drops hitting a green leaf, backlit by morning sun, shallow focus.

Aerial drone shot of a coastal city at sunset, golden light reflecting off glass buildings, smooth orbit.

Extreme close-up of coffee being poured into a white ceramic mug, steam rising, warm studio lighting, locked-off static camera.

Camera terms that work

These terms reliably produce the expected result in most video models:

• Camera movement: dolly in, dolly out, push in, pull back, pan left, tracking shot, crane up, orbit, static locked-off
• Shot size: extreme wide, wide, medium, close-up, extreme close-up, over-the-shoulder
• Lens: shallow depth of field, rack focus, anamorphic, fisheye, telephoto compression
• Lighting: golden hour, overcast flat light, harsh top light, neon, backlit silhouette, Rembrandt lighting

Key principles

• Describe the shot, not the story. “Wide aerial shot of a coastal city at golden hour” works. “A beautiful city where people live happy lives” does not.
• Keep it short. 15–40 words. Long prompts often confuse the model. If yours is longer than two sentences, cut it in half.
• Be specific. “A cool video of nature” gives the AI nothing. “Slow-motion macro shot of rain drops hitting a green leaf, backlit by morning sun” gives it everything.

Negative prompts

Some models support a negative prompt field in advanced settings. This tells the AI what to avoid — a blocklist of terms that reduces common artifacts. For a starter negative prompt and detailed guidance, see Negative Prompts.

Tips

• The prompt doesn’t affect cost — long and short prompts cost the same.
• When you get a result you like, save the prompt alongside the seed number for reproduction.
• Comma-separated phrases work just as well as complete sentences.

Common mistakes

• Being too vague. “A cool video” gives the AI nothing. Always include subject, environment, and camera.
• Writing an essay. Prompts over 40 words often confuse the model. Cut to the essentials.
• Describing a story instead of a shot. AI generates a single moment, not a narrative.
• Ignoring the negative prompt. A simple blocklist cuts common artifacts for free. See Negative Prompts for a starter list.
• Rewriting the prompt every time instead of using seeds. Change the seed for variations, not the prompt.

Quick answers

How long should my prompt be? 15–40 words for most models. Specific and concise consistently outperforms long and detailed.

Should I use complete sentences? Either works. Comma-separated phrases (“coastal city, golden hour, aerial drone shot, cinematic”) are fine. The model reads keywords, not grammar.

Do camera terms actually work? Yes. Most video models were trained on footage with shot descriptions. “Dolly in,” “tracking shot,” and “shallow depth of field” reliably produce the expected result.

What if the model ignores part of my prompt? Try increasing the guidance scale (CFG) slightly — it controls how closely the AI follows your prompt. Or simplify the prompt to focus on fewer elements.

Can I use the same prompt across different models? Yes, but results vary. Each model interprets prompts differently. A prompt that works perfectly on one model may need adjustments on another.