Mira made 5 takes, and each has a different city and a different girl. What should she fix first?

Describe the character and location specifically. The model fills in whatever isn’t written. A specific character and location give similar takes you can actually choose from.

In one clip the model runs in, jumps, turns and waves, and it all comes out smeared. What should Mira do?

Split it into shots: one main action each. Four actions in one clip is overload. One action per shot comes out cleaner, and each shot can be redone separately.

Mira needs the hoodie to look exactly like the real product. How should she start?

With a product photo as the first frame. A first-frame photo locks the product’s look, leaving motion and camera to the prompt. Using your own photo also means no rights issues.

In different takes the model runs left, then right, and the shots don’t match. What’s missing?

A “first frame” block with positions and direction. Without blocking the model picks the direction itself every time. “Right to left, shop windows on the left” removes the mismatch.

Anatomy of a prompt

Lesson 4 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.

Why “a cool video” doesn’t work

Whatever you don’t write, the model fills in by itself, and it fills it in like “the internet average”. So the prompt “a cool video of a girl in a hoodie” gives something new every time: a different city, a different girl, different light. A specific prompt answers a director’s questions: who, where, doing what, and how it’s shot. Then takes resemble each other, and mistakes are visible and fixable with words.

10 blocks of a strong prompt

Pros write a prompt like a short script: one block with labelled parts. The order: style, a one-line scene, characters, the location separate from the people, the first frame, shots step by step, camera, physics, light, sound. This order is recommended, for example, in the Seedance 2.5 guide. You don’t have to fill all ten every time, but if a clip comes out weird, one of the blocks is usually missing.

The first line sets everything

The style line is the rule for the whole video. Genre: a clothing ad with a documentary feel. Film or digital: “35mm, light grain” instantly removes the plastic AI sheen. Colour: cool shadows, warm skin. Format: 9:16 for Reels and TikTok, 16:9 for YouTube. And what must not appear: text, logos, extra people. If the style is the same in every shot, the video cuts together without jumps.

First frame and blocking

Directors call it “blocking”: where each person stands and which way they face at the start of a scene. Write it out: “model on the right, facing left, camera low on the left, shop windows along the left side”. When the starting point is clear, the model doesn’t mix up left and right between takes, and shots match later. A vague location, on the other hand, is the main reason a scene “drifts” from shot to shot.

Breaking it into shots

One shot, one main action. If in a single clip the model runs in, jumps, turns and smiles, the model will mash it all together. Mira splits the video into three shots: a wide shot of the street, a detail of sneakers in a puddle, a close-up of the face. Each shot can be generated and redone separately. In models that handle several shots in one clip, you write “Hard cut” between them, meaning a straight cut.

Text or a photo to start

There are two ways to start. Text to video: fast for trying ideas, but the look is different every time. Image to video: first you make or take a picture for the first frame, and the model “brings it to life”. Then the face, the hoodie and the light are already locked, and the prompt only needs to describe motion and camera. For ads where the product matters, Mira almost always starts from a photo. Just make sure the photo is yours or you have the rights to use it.

Building Mira’s prompt

Here’s shot 1 in full. Style: clothing ad, 35mm, light grain, 9:16, no text. Character: a 20-year-old woman, short bob, black oversized hoodie. Location: a wet street at night, shop windows on the left. Action: runs in from the right, torso leaning forward, arms pumping. Camera: low, 24mm, smoothly tracking her. Audio: footsteps in water, distant city, no music. Compare it with “a cool video of a girl”: the difference shows in every take.

Check yourself

Mira made 5 takes, and each has a different city and a different girl. What should she fix first?
Correct answer: Describe the character and location specifically. The model fills in whatever isn’t written. A specific character and location give similar takes you can actually choose from.
In one clip the model runs in, jumps, turns and waves, and it all comes out smeared. What should Mira do?
Correct answer: Split it into shots: one main action each. Four actions in one clip is overload. One action per shot comes out cleaner, and each shot can be redone separately.
Mira needs the hoodie to look exactly like the real product. How should she start?
Correct answer: With a product photo as the first frame. A first-frame photo locks the product’s look, leaving motion and camera to the prompt. Using your own photo also means no rights issues.
In different takes the model runs left, then right, and the shots don’t match. What’s missing?
Correct answer: A “first frame” block with positions and direction. Without blocking the model picks the direction itself every time. “Right to left, shop windows on the left” removes the mismatch.

← The 2026 model map Camera and lens →