To write a prompt for AI video generation, brief the tool the way you would brief a camera operator: name the subject, the setting, the action, how the camera frames and moves, the visual style and the light. Plan one prompt per clip, ask only for as much action as the clip has time to show, set length and shape in the tool’s settings rather than in the text, and change one thing at a time when you revise. The AI generates the clip; a person on your team decides whether it goes out under your brand.
Why do most AI video prompts produce clips you cannot use?
Most unusable clips come from prompts that describe an idea instead of a shot. When the prompt leaves out the setting, the movement or the framing, the tool fills those gaps with its own generic choices.
A common first attempt in a marketing team or an SME is a single line, such as “promotional video of our new pandan cake for Instagram”. That names a campaign, not a picture. The tool has to guess the plate, the counter, the light, what moves and where the camera sits, and it guesses differently every time. The team generates again and again, spending credits and an afternoon on clips that are almost right. Many teams also try to fit a whole advertisement into one prompt (opening scene, product close-up, happy customer, logo), which asks one short clip to carry a whole sequence.
Decide what the clip is for before you open the tool
Write a two-line brief first: where the clip will run, the one thing the viewer should notice, and how long and what shape it must be. The prompt then translates that brief instead of replacing it.
A vertical clip for Instagram Reels or TikTok, a landscape clip for LinkedIn and a looping clip for an in-store screen each need different framing from the first frame. Decide the must-haves too: the product must be recognisable, the brand colours should appear, and any on-screen text will be added in editing rather than generated. This is the role, task, context and format brief from our earlier guide, How to Write Prompts for AI Video Generation That Give You Usable Clips, adapted for moving pictures.
Cover subject, setting, action, style, camera, framing and light
Google’s prompt guide for its Veo video model lists the elements a video prompt should include: subject, context, action and style, with camera motion, composition and ambiance as optional extras. One sentence per element removes most of the guessing.
Compare the one-line prompt above with this version:
A slice of pandan layer cake on a white ceramic plate, on a wooden café counter. Morning sunlight comes through a window on the left. A hand slowly slides the plate towards the camera and stops. Close-up, shallow depth of field, slow push-in. Warm natural tones, realistic food-advertising style.
Each sentence settles one decision the tool would otherwise make for you. Google’s guide shows the same effect with its own before-and-after example, noting that “a more detailed prompt results in a video that is more focused with a richer environment.”
The tools named in this guide are examples, not recommendations, and they change quickly. OpenAI, for instance, closed its Sora video app in April 2026 and its Sora API in September 2026. Keep your team’s prompt template tool-neutral, and test it in whichever tool you use today.
Plan one prompt per clip and match the action to the clip length
Each generation produces one short clip, so write a prompt for each clip and ask only for as much action as that clip has time to show. Build longer videos by joining clips in editing.
In Runway’s current Gen-4.5 model, for example, you choose a clip length between 2 and 10 seconds. Runway notes that “prompts that request multiple sequential actions may benefit from longer durations.” A 15-second Reel therefore needs at least two clips, and is usually easier to plan as three or four shots, each with its own prompt. Newer models handle more than earlier ones did; Runway says Gen-4.5 “excels at understanding and executing complex, sequenced instructions”. Even so, a five-second clip that has to show a hand, a plate, a camera push-in and a logo is likely to feel rushed. Start with one clear action per clip and add more only when the clip has time for it.
Set length and shape in the settings, not in the prompt
Duration and aspect ratio are usually controls in the tool, not words in the prompt. Choose them before you write.
In Runway Gen-4.5, you set the duration in the interface. Aspect ratio is set there too, and it depends on the mode: text-to-video produces 16:9 landscape only, while image-to-video also offers 9:16 vertical, 1:1 square and other ratios. That matters for social media. If your Reel must be vertical, check that the mode you are using can produce it before you write a word of prompt. Other tools have their own settings panels, and the options change as models update.
Start from an approved image and describe only the motion
When the product must look exactly right, start from an approved photo and use the prompt to describe what moves. The image sets the look; the text sets the movement.
Runway’s Gen-4.5 guidance says “Image to Video prompts should focus on describing the motion of the scene,” while text-to-video prompts “should describe both the visual elements and motion of the scene.” Its getting-started guide adds: “You do not need to describe the contents of the image.” In the pandan cake example, the prompt shrinks to the movement: a hand slides the plate slowly towards the camera and stops; slow push-in. For brand work this is often the safer route, because the product shot is already approved and the AI only adds movement.
Take care with images of people. If a reference photo shows a real customer or staff member, follow your organisation’s data policy and get that person’s permission first. Malaysia’s Personal Data Protection Department sets out the General Principle of the Personal Data Protection Act 2010: a person’s personal data may not be processed without their consent. Check the tool’s own rules on images of real people as well.
Describe what you want, and list what to leave out separately
Write the main prompt in positive terms. Where the tool offers a negative prompt field, list unwanted elements there as plain nouns, not instructions.
Google’s guide advises against “instructive language or words like no or don’t”, such as “No walls”, and suggests listing what you don’t want instead, for example “wall, frame”. For brand clips, generated text, logos and watermarks are common items to keep out, since those are better added properly in editing.
Revise one element at a time, then have a named person approve the clip
When a clip is close, keep it and change only what is wrong. Before anything is published, a named person checks it against the brief.
Adobe’s Firefly prompt guide notes that “small changes in your AI video prompts can significantly impact the final video output.” That is the reason to change one element per attempt. If you alter the light, the camera and the action together, you cannot tell which change fixed the clip or broke it. Keep the closest version, note what you changed, and adjust the next element.
The second half is a human job. The reviewer checks that the product’s shape, colour and packaging are accurate, that hands, faces and any background text look right, and that the clip does not suggest the product does something it cannot. AI drafts the footage. People decide what is published, and the approver, not the tool, is accountable for it.
How managers can turn video prompting into a team habit
Agree one prompt template, keep a shared log of prompts that worked, and review a few clips together each week.
The template can be the seven elements as blank lines, plus platform, duration and aspect ratio. The log records the tool, the prompt, the settings and the clip it produced, so the next person starts from a proven version instead of a blank box, and the team can see quickly what needs redoing when a tool changes. In the first month, review two or three clips per person against their brief. Practise it in the workshop, then reinforce it on the job.
What changes when your team prompts this way
Fewer generations go on clips that were never going to work, because every prompt starts from a brief. Clips fit the platform they were made for, product shots stay accurate because they begin from approved images, and revisions get smaller. Approvals move faster because the reviewer compares the clip with a written brief instead of a feeling.
Frequently asked questions
What should a prompt for AI video generation include?
Google’s Veo prompt guide lists subject, context, action and style as the core elements, with camera motion, composition and ambiance as optional additions. For a short marketing clip, one sentence on each is usually enough.
How long should an AI-generated video clip be?
It depends on the tool and on how much happens in the clip. In Runway’s Gen-4.5, for example, you choose between 2 and 10 seconds, and Runway notes that prompts asking for several actions in a row may work better with a longer clip. For a longer video, plan several clips and join them in editing.
Can we use photos of our staff or customers to generate AI video?
Only with each person’s permission and within your organisation’s data policy. Under the General Principle of Malaysia’s Personal Data Protection Act 2010, as set out by the Personal Data Protection Department, a person’s personal data may not be processed without their consent. Check the AI tool’s own rules on images of real people too.
Do we have to label AI-generated video in Malaysia?
The rules are still taking shape. On 2 September 2026, The Star reported that the Communications and Multimedia Content Forum of Malaysia had proposed influencer-marketing guidelines. Under them, material use of AI or digital alterations that affects how a product, result or human appearance is perceived should be disclosed, and public consultation was open until 30 September. Check the final guidelines and each platform’s own disclosure rules before you publish.
Planning in-house training for your team?
If your team has an in-house training request, you are welcome to enquire with us. ClimbX Academy is an HRDC/HRDF registered training provider in Kuala Lumpur. Email hello@climbxacademy.com or WhatsApp +60 10 872 5194.
Carmen Lee is the founder of ClimbX Academy and has over ten years of experience training frontline and commercial teams, including in hotels.
Sources
- OpenAI Help Center, What to know about the Sora discontinuation: the app and web version closed on 26 Apr 2026, and the API on 24 Sep 2026
- Google Cloud, Vertex AI video generation prompt guide (no date or model version shown; checked again 12 Oct 2026)
- Runway Help Center, Creating with Gen-4.5
- Runway Help Center, Getting Started with Generative Video
- Adobe Firefly, AI video prompts guide
- Jabatan Perlindungan Data Peribadi, Prinsip Perlindungan Data Peribadi (in Malay; General Principle)
- The Star, Low Co Sin, on the content forum’s proposed influencer guidelines, 2 Sep 2026
Tell us what your team is facing. We will come back with a plan built around it.
Contact Us