Generating an image or a short video has become more accessible. Creating content with intention, continuity and a visual language of its own still requires choices. It is in that transition — from generation to direction — that artificial intelligence starts to become part of a real production process.
For me, the conversation about AI video needs to move beyond amazement at the first result. A beautiful shot can open up a possibility. The work is figuring out what it communicates, how it connects to the next shot and why someone should keep watching.
The evolution: from an isolated response to a creative process
Generative creation first gained ground as a quick way to experiment: request an image, try an aesthetic, visualise an idea. In video, the ambition grows. Beyond the appearance of a frame, movement, duration, continuity and sound come into play. The image needs to work over time.
Tools have introduced visual references and controls for directing scenes. Google's guide to Veo, for example, describes reference and audiovisual direction features. That expands the possibilities, but it does not remove the need to select and review. Having more control available does not mean every generation will respect every detail.
Video generation: think in shots before thinking about the film
A film does not have to come from a single generation. Breaking an idea into shots helps give each segment a purpose: establish a setting, show an action, reveal information or change the pace. A shot can work on its own and still fail to serve the sequence. Editing exposes that difference.
Before generating, I would prioritise concrete questions: what is the main action? Where is the camera? What information needs to stay the same between shots? How does this scene end to allow the next cut? Then I would review the movement frame by frame, looking for unintended changes in objects, faces, scale and lighting.
A prompt is written direction, not a collection of adjectives
A useful prompt turns intention into observable decisions. ‘An incredible, cinematic video’ leaves almost everything open. Describing the subject, action, framing, camera movement, light and atmosphere gives the model a clearer starting point. OpenAI's prompting guide for Sora also suggests thinking of a shot as a storyboard and working with understandable actions.
An example prompt to experiment with: ‘Medium shot of a ceramicist in a workshop. She places a clay piece on the table and moves her hands away. The camera slowly moves closer, without cuts. Soft side light from a window; terracotta and wood tones. Keep the piece's shape unchanged throughout the action. Subtle ambient sound, no speech.’ This is a direction hypothesis, not a guaranteed result. If it fails, simplify the action and test one change at a time.
Skill: making a good procedure reusable
A prompt and a skill are not the same thing. A prompt guides a task or a generation. An agent skill brings together instructions and, when needed, resources and scripts to carry out a type of work more consistently. Anthropic's documentation presents this format as a way of organising specialised knowledge that an agent can load when it is relevant.
In a content workflow, a skill could help break a brief into shots, check script fields or organise file names. It does not install talent in a model or guarantee quality by itself. It needs a clear scope, examples, limits and evaluation. Here, skill means a reusable procedure for an agent, rather than a command every video generator understands.
A workflow that leaves room for creation and review
I would organise production into six stages. First, a brief covering audience, message, format and objective. Then a script and shot list. Next, visual references to establish character, objects, lighting and palette. Only then would generation begin, with labelled versions and notes on what worked. The fifth step would be selection and editing. The sixth, finishing and delivery review.
In editing and finishing, rhythm, colour grading, motion design and sound design help bring the material together. AI can contribute to several of those stages, but the criteria need to remain explicit. A final review should check legibility, continuity, audio, framing in the target format and brand fit. To reduce aimless attempts, it helps to define in advance what makes a shot approved and when it is better to rebuild it through another approach.
AI slop: when volume replaces judgement
AI slop is an expression used for low-quality content produced with AI, often in large quantities. The definition recorded by Merriam-Webster points to that combination of artificial production, low quality and volume. That does not make every piece of AI-generated content slop. The discussion is about the result and the editorial care behind it.
In practice, the signs can include a striking image without a purpose, unchecked information, characters changing between shots or a repeated aesthetic unrelated to the message. A simple, well-resolved piece can be stronger than a sequence full of effects. To avoid working on autopilot, I would ask four questions: does this meet the brief? Is there a recognisable idea? Do the details hold up under review? Would I publish this result without relying on the argument that ‘it was made with AI’?
The difference lies in the decisions that reach the screen
The evolution of AI opens up ways to experiment with visual language and make scenes possible. It also increases the responsibility of those publishing the work. Not presenting synthetic images as documentary records, checking information and considering the origin of references are parts of the process. Transparency helps put creation in context, especially when an image could be mistaken for a real event.
What interests me is turning these tools into films that make sense: a clear idea, a coherent visual language and a finished result. A prompt opens a possibility; a workflow organises the work; direction decides what deserves to reach the screen. If you want to develop an AI video for your brand, let's talk about the idea before choosing the effect.
References and further reading
Google Cloud — The ultimate prompting guide for Veo 3.1. Reference on scene direction and generation controls; features vary by tool and version.
https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1
OpenAI — Sora 2 Prompting Guide. Reference on describing shots, actions, camera and iteration; this is not an indication of the product's current commercial availability.
https://developers.openai.com/cookbook/examples/sora/sora2_prompting_guide
Anthropic — Equipping agents for the real world with Agent Skills. Reference for the definition of agent skills.
https://www.anthropic.com/research/equipping-agents-for-the-real-world-with-agent-skills
Merriam-Webster — 2025 Word of the Year: Slop. Reference for the use of the term.
https://www.merriam-webster.com/wordplay/word-of-the-year
Editorial text supported by AI and an AI-generated cover for this publication. References consulted on October 22, 2026. The prompt and workflow examples are proposed processes, with no promise of automatic results.
