What You Will Do
Before the first generation, you will record five decisions and prepare one small test. After that, you will know what to run later and what visible result counts as acceptable.
This is not a huge preparation document or a place to move the whole script, every reference, and future prompts. It is one page that stops the prompt from becoming a container for chaos. A prompt is a text instruction for an AI tool; you do not need to write it in full yet.
“Where an AI Video Begins” may help if the idea is still vague. It is not a compulsory gate: everything needed for this check is explained below.
Five Decisions Before You Start
Write the answers in ordinary language. The card is not a decoration; it makes the first attempt answer one clear question.
1. What are we making, and for whom?
2. Which moment or result are we testing first?
3. Which real materials already exist?
4. What must not be lost or changed?
5. What visible sign makes the test acceptable?The full script, camera map, references, and prompts remain separate materials. Do not copy them here in full: this card has a different job.
1. What Are You Making, and for Whom?
Avoid “a beautiful AI video.” Name a concrete result for a concrete viewer.
Weak:
Make a cinematic video about space.
Better:
Make a short comedy video: a crew mistakes a dangerous water planet for a resort because the robot gets it wrong. The viewer should first believe in the serious expedition, then understand the joke.
If the idea does not fit in two or three sentences, clarify it first. A beautiful frame without a clear task is a beautifully unemployed frame.
2. What Are You Testing First?
Do not ask the first generation to prove the whole video. Choose one moment that the next work depends on: a character’s appearance, location, composition, key action, object scale, or readability of an important element.
3. Which Materials Already Exist?
Record only what exists in reality: an approved script or timing, an image of a character, object, or location, a light, colour, composition, or motion reference, a logo, package, exact text, or an accepted frame.
Write the job of each material beside it:
Robot image — preserves the silhouette and body material.
Water-planet frame — preserves the environment and cold light.
Exact ending line — must not be paraphrased or reinvented.
4. What Must Not Be Lost?
Choose only the conditions that genuinely need protection: the character’s identity or appearance, a product, logo, or object shape, scene geography, the meaning of the joke or action, exact text, or an important end state.
Do not put everything on this list. Twenty-seven non-negotiable conditions are no longer protection; they are a panic attack formatted as bullets.
5. What Does an Accepted Result Look Like?
The criterion must be visible.
Weak:
It feels atmospheric and professional.
Better:
The same practical robot is visible; there is only water around it; the cold light remains; the robot does not look like a toy; the silhouette reads at a small size.
Filled Example
This card is for one small test. It does not promise to make the whole video at once; it fixes the first question worth checking.
What are we making, and for whom?
A short comedic sci-fi video. The viewer first sees a serious expedition, then realises that the robot selected a dangerous planet as a resort.
What are we testing first?
The first wide shot of the water planet with the practical robot.
Which materials already exist?
- robot image — preserves the silhouette and matte metal;
- environment reference — preserves endless water and cold light;
- scene idea — preserves the serious delivery before the joke.
What must not be lost?
- there is no land or random architecture around the robot;
- the robot does not become a cute toy;
- the delivery stays serious and the joke does not move into the props.
Accepted result:
The robot is recognisable, the environment reads as a landless water planet, the light is cold, and no extra comedy props appear.
Plan One Small Test
Plan one version that checks the chosen moment. Do not test the character, light, camera, composition, and background at the same time.
If you are testing appearance, composition, or light, start with one still image. If you are testing an action, make one short clip. Attach only a material with a defined role, describe the observable result in two to four lines, and after the future run compare the actual result with the criterion.
Test plan: one still image of the robot on the water planet.
Check: robot silhouette, no land, and cold light.
If land appears: change only the environment description in the next attempt.
Readiness Check
You can move on when you have:
- one real material that supports the work;
- one moment being tested now;
- one condition that must not be broken;
- one visible criterion for an accepted result;
- a small test plan and exactly one change for the expected failure.
If the test depends on motion, a spoken line, or exact duration, identify where the timing comes from: finished audio, source video, or a chosen duration. If it is still unclear what belongs in the frame, collect visual rules and references. If the action is too large for one generation, split it into controlled segments. These are next routes, not a task list you need to open all at once.
Takeaway
You do not need a giant document before the first generation. You need a short agreement with yourself:
what we test → what supports it → what we protect → how we know it worked
Then continue with the material you actually need: story and timing, references, image preparation, route selection, or a video segment. That is where the plan becomes a real run and an observed result. The generator can wait. It usually stays put—unlike good decisions you decided not to write down.