AI video prompts: Runway, Kling and Gemini alternatives to Sora

Write an AI prompt to create video is different from the prompt for text or images: you are describing a moving scene, not a snapshot. The generator needs to know what moves, how the chamber moves, what atmosphere you want and for how long. If you give a vague statement you get a generic clip; if you give a structured statement you get something usable.

The main tools today are Runway, Kling AI, Pika, the video models of Google (Gemini Omni in the Gemini app, Veo for developers) and HeyGen for avatars. Each platform interprets prompts slightly differently, but the basic logic is the same: the more visual context you provide, the less the model "invents" details at random. This guide shows how to build that context, with practical examples for those who work in the company.

If you get here looking for Sora: OpenAI closed the app and site on April 26, 2026 and fixed the API shutdown on September 24, 2026, as explained by the OpenAI Support Page. The prompts you wrote should not be thrown away: the five-block structure you find here also works on the tools of this guide.

To deepen the basis of prompt engineering, read first What is prompt engineering and how to get started and the general guide on AI prompt.

What is an AI prompt for video and why is it different from the image prompt

A AI Image Prompt describes a frame. A video prompt describes a scene with a time frame: how it starts, what moves, how it closes. The model must generate dozens of frames consistent with each other, which makes each ambiguity much more expensive in terms of result.

The practical difference: if you write "a cup of coffee on the table" for an image, you get exactly that. If you write it as a video prompt, the model decides for itself if the cup fills, if the camera approaches, if there is sunlight or artificial light, if it lasts 4 or 8 seconds. Every unspecified variable becomes a random choice.

The Generative AI for video is even less "reliable" than text AI: the models are younger, the parameters more complex, the results less predictable. The prompt is not just a creative tool, it's the only control you have.

The structure of an effective video prompt

The framework that works on all major tools follows this pattern in five blocks:

1. Scene. Main subject + action + environment. Example: "A bartender prepares an espresso in a modern bar, soft lighting in the morning."
2. Camera movement. Shooting type: static, dolly in/out, pan horizontal, tracking shot, aerial. Example: "Slow dolly in towards the cup, starting from half a torso."
3. Visual style. Cinematic, documentary, corporate, hyper-realistic, 3D rendered, vintage, drone footage. Example: "Cinematic style, shallow depth of field, warm colors."
4. Duration. Indicates the target length in seconds (tools have different limits; verification). Example: "6-second clips."
5. Mood. Rhythm, energy, light. Example: "Quiet atmosphere, slow pace, early morning light."

Full prompt example: "A bartender prepares an espresso in a modern Milanese bar. Slow dolly in towards the cup as the steam comes out. kinematic style, shallow depth of field, warm colors. 6-second clip. Quiet atmosphere, early morning."

This prompt is workable on Runway, Kling or Gemini. The result will not be perfect on the first attempt, but will be much closer to the goal of a generic "barist who makes coffee".

Comparison of major AI video tools

The tools evolve quickly: the data below was checked in September 2026 on the official pages, but always checks prices, limits and availability before choosing.

Prompt for Gemini and Veo: what works and what doesn't

In the Gemini app, video generation today switches from Gemini Omni, the model that Google has put instead of Veo in the app. Create 10-second clips with native audio, start with any combination of text, photos, and videos, and accept chat fixes:change the background, light, or character without rewriting the prompt from scratch. It serves a Google AI subscription, features vary by plan and country, and use is reserved for adults.

He who develops finds Veo in the Gemini API, with clips of 4, 6 or 8 seconds in horizontal or vertical format and the ability to extend an already generated video. The same page hosts the official Veo prompt guide, also useful to those who use the app: it asks for subject, action and style, plus some optional elements such as position and movement of the camera, composition, focus and atmosphere. They are almost the same five blocks as this guide.

What works well: concrete descriptions, with adjectives and adverbs that clarify the scene, and indications on the audio. Google's guide suggests putting the dialogues in quotes and describing in words the sound effects and background noise.

What is best to avoid: ask for too many actions in a clip of a few seconds or mix incompatible styles. If the subject is to remain the same from scene to scene, start from an image: Veo accepts up to three reference images of a person, character or product and maintains its appearance in the video.

Example prompt for the Gemini app: "Aerial view of a Tuscan vineyard at dawn, light fog among the vineyards, drone shot slowly lowering towards the leaves. Cinematic style, golden colors. 10 seconds. Silent and contemplative atmosphere. Audio: light wind in the rows, a rooster in the distance."

AI Runway Prompt: Taking advantage of camera control

With Runway's newest model, he explains the Official Guide, you can describe in a single prompt the camera movements, the composition of the scene, the timing of events and the changes in atmosphere. Clips last 2 to 10 seconds and the model requires at least the Standard plan.

Recommended structure for Runway:

1. Describe the scene in a clear sentence
2. Specifies the type of shot (close-up, wide, medium)
3. Write in the prompt also the movement of the camera: if you start from an image, Runway advises to describe the movement above all
4. Set the tone: "cinematic", "documentary", "commercial"

Example prompt for Runway: "Cosmetic product on white marble surface. Close-up, light dolly in towards the bottle. Soft side light, defined shadows. Commercial photography style. 5 seconds."

Kling AI Prompt: Realism and Natural Movement

Kling AI (developed by Kuaishou) is known for the realistic quality of human movement and product videos. It handles scenes well with water, textiles and organic materials.

Example prompt for Kling: "A runner walks through a park at dawn, side-framed tracking. Light mist, grazing light. Documentary style, natural colors. 8 seconds. Energizing but relaxed."

For product videos Kling often produces ready-to-use results with simple prompts: just specify the material, light and environment.

Practical examples for companies

Video produced for e-commerce

A company selling extra virgin olive oil wants a short clip for the product card.

Prompt: "Dark glass bottle with an artisanal label on a rough wooden table. Close-up following the stream of oil slowly dripping from a spoon. Warm natural light, food photography style. 5 seconds. Authentic, artisanal atmosphere."

Expected result: clip usable on e-commerce, social and digital packaging, produced in minutes instead of with a photo set.

Short spot for social media

A physical therapy studio wants a reel for Instagram with a professional but accessible atmosphere.

Prompt: "A physiotherapist in a lab coat works on a patient's back in a modern practice, with large windows and natural light. A slow tracking shot entering the room. Warm corporate style, neutral colours. 6 seconds. An atmosphere of care and professionalism."

The clip can be used as an intro reel, a story cover or b-roll for a longer video.

Short formative video

A consulting firm wants an animated intro to an online course.

Here HeyGen with talking avatar is often more effective than pure generators: you load a script, choose an avatar and get a training video with virtual speaker. The prompt becomes the speech script, not the visual description.

Rights, commercial use and ethical issues

Commercial use: rights change from tool to tool and sometimes from plan to plan. Runway declare that the content created is yours, with no non-commercial restrictions on its part; Pika includes the commercial license only from the Creator up plan; Kling and Google have their own terms. Read the section on commercial use in the terms of service before posting.

Deepfake and real people: Generating realistic videos of real people without consent is illegal in many jurisdictions and violates the terms of all major tools. This also applies to satirical or parodic content in certain contexts.

Transparency: in Italy and the EU there are increasing indications (AI Act) towards the declaration of AI origin in content. For B2B marketing it is also a choice of credibility.

Mistakes to Avoid

• Vagus prompt: "a nice video of a product" is not a prompt, it's a thought. The model fills in the blanks randomly.
• Too many contradictory instructions: "minimalist style, very colorful, crowded, with so many details" produces visual artifacts.
• Ignoring camera movement: if you do not specify it, the camera does something by default (often a slight random zoom). Always specify.
• Expecting consistency between separate clips: each generation is independent. If you need the same subject in multiple scenes, use the "image-to-video" function from a fixed frame, or tools that support character consistency.
• Skipping the test on short clips: always start from 4-5 seconds to verify that the direction is right, then extend or regenerate.

A reference video can drive multiple variants of the same message. With the new interface of Higgsfield Ad Multiplier for Advertising Videos, the work on the prompt must be accompanied by control of rights, brand consistency, cost before generation and testing of creativity.

How to apply it in the company

AI video makes sense in the company when the cost of traditional production is disproportionate to the goal. Concrete cases:

• E-commerce: product clips for each variant, without photo set each time.
• Social media: b-roll for reel and stories without turning hours of material.
• Presentations and Pitches: animations and visual scenarios for abstract concepts.
• Prototyping: show the customer what a spot might look like before investing in the actual shot.
• Internal training: intro and transitions for e-learning courses.

The Practical Flow: Define the Video's Goal → write the prompt following the structure of the 5 blocks → generates 3-5 variants → select the best one → refine with a second prompt or interface controls → exports.

To understand which tools to use in your specific situation, the guide on Best AI Tools and the one about Best AI for Video offer an updated comparison. For use in marketing, see also the best AI for marketing.

Conclusion

Write a good one AI prompt for video is not complicated, but requires discipline: describing the scene, camera movement, style, duration and mood. Every element you omit is decided by the model, not by you. More mature tools today, such as Runway, Kling and Google's video models, produce professional-quality clips for specific uses (product, reel, short spots), but they still do not replace complex video production.

The starting point for any company is simple: identify a video you often make with high costs (product photo, b-roll for social, formative intro) and test AI on that. Results are seen in hours, not weeks.

Do you want to understand what processes in your company can be automated with AI? Giallo Studio can help you turn the problem into a real workflow.