The Prompt Pack Cuts Like A Film

The difference between the two is 1 word, and it changes the whole job. A wish list asks the model for a good picture. A shot list tells it which frame of the film this is. 5 frames carry 1 night, in order. Frame 1 is the invitation. Frame 5 is a door left closed with a line of light under it. The sequence is written as an edit for 1 35 mm frame size, rather than as 5 separate renders that hope to match later. The probability that 5 independently prompted frames cut together as 1 night is not 0. It is just low enough that the pack stops being a storyboard and becomes a folder of pictures. So the effort goes where the joining happens: 1 line of style, 1 block of negatives, 1 identity rule. The wording of any single frame matters far less than the 3 things every frame shares, which is why the next 3 sections are about sharing rather than about writing.

Every prompt opens with the same style line, word for word, every time. A film still, anamorphic on 35 mm. Fine film grain. Thin atmospheric haze. Shallow depth of field. Unhurried. No text, no watermark. 26 words, written once at 24 fps, pasted under frame 1 and frame 5 alike. Not 1 of those 35 mm decisions improves any single frame on its own. Together they are what stops 5 generations drifting into 5 different films. Grain does the joining. Haze does the joining. The anamorphic 35 mm frame does the joining, and the 24 fps register holds the motion language the same way. Change 1 descriptor and the pack splits along that seam. Frame 3 reads as a different film from frame 4, and the cut shows it in under 1 second. Copy and paste is the cheapest consistency tool that exists, and it is also the 1 most often skipped. Write the anchor once at 24 fps, keep it in a text file, and never retype it from memory.

Every prompt also carries the same negative list, and the list is written as failures rather than wishes. Cartoon. 3D render. CGI look. Bright daylight. Green tint. Oversaturation. Extra limbs. Deformed hands. Text. Logo. Watermark. 11 items, 17 words, and each 1 names a specific failure that was seen once and banned after that. This is where most packs are thin. The prompt states what is wanted and leaves everything else to the tool. The tool will not police it. If a watermark sits in the training distribution, a watermark is what comes back. Naming the trap costs 1 line and saves 1 re-render, which is the cheapest trade in the file. 2 items in that list do most of the work: deformed hands and text. Both are the standard failure of every image model, and both are the 2 that ruin an otherwise finished frame.

Across all 5 frames, the same woman, the same face, the same silk slip. Every frame. No exceptions. This is the load-bearing rule, and dropping it is what returns the pack to a folder. A viewer does not read 5 images. A viewer reads 1 woman across 5 moments, and the moment only holds if the woman is identical. So the description of her does not move - not 1 word, not 1 adjective. Same hair. Same slip. Same build. Same age. What changes is everything around her. Frame 2 holds no face at all. It is a macro close-up of a hand, and the body inside that frame is the same body. Frame 5 holds no face either. It is a black frame, 1 line of amber light under a closed door, and a silhouette behind it. 2 of the 5 frames never show her face, and the sequence still reads as 1 woman, because the other 3 frames fixed her.

A wish list asks for a good picture. A shot list tells the model which frame of the film this is.

Every frame is written to be cut against the frame before it and the frame after it. That is why the sequence carries 3 light temperatures and 2 camera distances, and never repeats the same beat twice. Frame 1 is an invitation. Frame 2 is a tracing at a centimetre a second. Frame 3 is a pause with a hand held 1 inch from the lens. Frame 4 is the face, held. Frame 5 is the closed door. Read that as rhythm and it is obvious: it opens close, opens closer, pulls back, holds, and ends on nothing but a line of light. 1 change of distance per beat. That is the edit, written before anything is rendered. A frame built to stand alone fights every frame it is cut against, because it was composed to be looked at rather than left. Write to cut and the edit becomes a decision already made instead of a rescue in the timeline.

The pack is small. 5 frames. That is the point of it. A pack is not a shortcut around craft. It is craft stored so that it can be reused: 1 anchor at 24 fps, 1 negative block, 1 identity rule, 1 rhythm. Once those 4 things are written down, frame 6 costs almost nothing to write. The first frame is expensive because it carries every decision, and the 5th is cheap because it inherits them. The same arithmetic ran the physical builds. Companies built across 12 countries. 210 energy systems across Africa and Asia. 200 wood-gasification machines across the UK and Europe. Not 1 of those rows was improvised on the day. The specification came first. The 5 frames came out of 1 specification, and 1 specification is what makes 20 frames possible rather than 5. The probability that a blank prompt returns the exact frame in your head is not 0. But it is 20 attempts priced as 1 frame, and a pack is 1 attempt priced as 20 frames. So write the anchor once. Write the negatives once. Write the identity rule once. Everything after the first frame is a variation on a decision already made.