Learn

How to write a scene the engine draws exactly

Dmytro Klymentiev · 7 September 2026 · 8 min read

Short answer: Each frame is drawn on its own, from the words of that scene alone, so anything left vague is invented afresh every time. Name the object once and repeat that sentence word for word in every scene it appears in, write the printed strings yourself in full, give the count of anything countable, end each description with what is not in the frame, and never name a colour: colour belongs to the style.

Every frame of a video is drawn on its own, from the words of that one scene and nothing else. It cannot see the frame before it and it does not remember what you meant. So anything you leave vague is invented afresh, and invented differently, every single time.

That single fact is behind almost every video that comes back not quite right, and it is also the whole method for getting one back exactly.

What the engine actually reads

A scene carries three fields, and they do different jobs.

FieldWhat it decides
The spoken lineWhat a person hears, and therefore how long the scene lasts
The captionThe only sentence a viewer with the sound off reads
The scene descriptionEverything that appears in the picture, down to the words printed inside it

The style is chosen once for the whole video and is never mentioned inside a scene. Palette, line weight, lettering and surface all come from there. A scene that names a colour is arguing with the style, and the scene wins.

Rule 1. Name the object once, then copy the sentence

If a person, a tool or a document appears in more than one scene, write one sentence that describes it and paste that same sentence into every scene it appears in. Not a paraphrase. The same sentence.

In a receiving procedure we shot, the operator, the pallet and the clipboard each carried one fixed description, repeated word for word across six scenes. The pallet in scene one is recognisably the pallet in scene three, and the operator keeps the same vest, the same gloves and the same clipboard all the way through.

The same video written loosely, "a sheet of paper, seen straight on", came back with a different object in every frame: a spiral notepad in one, a ring binder in the next, a clipboard in the third. Nothing was wrong with the engine. Five separate frames were told five slightly different things.

Rule 2. Write the printed strings yourself, in full

Every string given in full comes back character for character. Every string merely described comes back invented.

Weak:   the clipboard shows the order number and the pallet count
Strong: printed on the clipboard, exactly two lines: PO 4471 and 6 PALLETS

Two things follow from this. Keep the strings short, three to five lines per frame at most, because a paragraph drawn on a clipboard degrades into texture. And write them bare: we once wrapped labels in square brackets to mean "this is the printed text", and the brackets were dutifully drawn onto the clipboard.

Rule 3. Count everything countable

"A grid of blocks" produces a different number of blocks in every frame. "A grid of exactly four blocks, two above two" produces four.

The same goes for lines, panels, steps and columns. If the number matters to the meaning, and it usually does, put the number in the sentence.

Rule 4. End with what is not in the frame

An engine given empty space fills it with props, and props are what make a frame look generated. So close every description with the absences that matter:

Nothing else in the frame. No people in this frame. No desk, no shelving, no forklift.

Be specific about it. In one run the blocks we asked for came back as spiral bound notepads, complete with rings, because nothing had said they were plain rectangles with square corners and no binding.

Rule 5. Nothing can be held across frames

This is the rule people fight hardest, and it always wins.

We once put a block at the top of a script asking every scene's header to match scene one in weight, height and position. The headers came back light and left aligned in the first two frames, bold and centred in the next two, and higher up in the last. Prose at the top of a script does not reach the renderer frame by frame.

Repeating the header sentence inside every scene helps with the wording, and it still does not fix the typography: size and position are decided anew each time. Layout that must be identical everywhere belongs to the style, not to the script.

The practical version of this rule: design each frame to stand alone. A comparison does not have to live inside one frame. Show the policy in one scene and the procedure in the next, and the viewer does the comparing. That version works. A two column table that has to be reproduced five times does not.

Rule 6. One action per scene

A scene that says "check, count, photograph and sign" is four scenes wearing one hat. Twenty to thirty-five spoken words is about ten seconds, which is roughly how long a viewer needs to take in a picture.

If a script runs long, split the scene rather than speeding up the narration.

What to ask for, and what to stop asking for

Stop asking forAsk for this instead
The same document reproduced in five framesOne object per frame, each described in full
A layout rule written once at the topThe same sentence inside every scene that needs it
"A red warning line""The line the scanner beam makes", colour left to the style
"Show the process being checked""A scanner pointed at a label reading BIN A-12"
An empty frame, "the table is blank"Something concrete to draw, however little

That last one is worth a word. Asked for an empty two column table, one run filled both columns with enormous check marks that meant nothing. Emptiness is not an instruction.

A scene, before and after

Before
A sheet of paper with the policy on it, seen straight on.

After
Centred and large, filling most of the frame: a single plain sheet of paper,
portrait, seen straight on, with square corners, a plain flat surface and a wide
empty margin down its left side, lying flat with nothing behind it. Printed on the
sheet, exactly: the heading POLICY 4.2 in the upper third, and under it two lines,
SPEND OVER $500 IS APPROVED FIRST and RECEIPTS ARE KEPT SEVEN YEARS. In the left
margin, level with those two lines, one short line in a second hand: WHAT MUST BE
TRUE. Nothing else in the frame. No second sheet, no spiral, no rings, no binder,
no clip, no desk, no hands.

The second one is longer to write and it is written once. The first one is shorter and gets rewritten four times.

The checklist

Before you spend a render, read each scene and ask:

  • Is there exactly one action in it?
  • Does the caption state the rule, not the topic?
  • Is every printed string written out in full, and short?
  • Is every recurring object described in the same words as in the other scenes?
  • Is anything countable given a number?
  • Does the description say what is not in the frame?
  • Is any colour named? Remove it.
  • Would this frame make sense if it were the only one anybody saw?

Planning and editing cost nothing, so all of that is fixed while the video is still text. Fixing it afterwards costs a render.

Keep reading

Make your first video