A Style Block Is A Sentence You Reuse

The global style block is 1 paragraph of about 60 words and it goes into every still and every clip unedited, in the same order, every time. It names 7 things in sequence: a muted desaturated cinematic film still, golden-hour light, warm amber against deep navy, 35mm film grain, long soft shadows, gentle atmospheric haze and shallow depth of field. That is the whole block. The same 60 words render a 16:9 frame at 24 fps for a 3 second clip of 72 frames and a 30 second sequence of 720 frames, and nothing between those 2 jobs has to be edited, because the block never described the length. The reason it is a block rather than a habit is that 2 generation scripts read it: gen_stills_mmx.py and gen_video_mmx.py both pull from the same markdown file, so a change to the light lands in 1 place instead of being re-typed into 2 scripts and 40 prompts. The order is load-bearing. The 7 elements read as 1 sentence: subject first, then light, then colour, then texture, then shadow, then air, then depth. Move depth of field to the front and the renderer weights it first, and 10 frames come back with a shallow field and no atmosphere in them. 0 quality adjectives appear before the 7th element, which is exactly why the same block survives a model swap without a rewrite.

The tail of the block stops describing the picture and starts describing the material: quiet contemplative mood, editorial and slightly disquieting, photoreal, documentary film, natural skin texture, visible pores, unretouched, natural light, negative fill, slow deliberate movement, single subject. Then the last line is an instruction rather than a description, and it is the shortest line in the file: no text, no watermark, no caption, no logo. That line is the 1 that gets dropped by hand when someone types a prompt at speed, and it is the 1 that comes back 3 renders later as a caption burned into the lower third of a frame you have to re-generate anyway. So the exclusions were moved out of the block and into a second file: the negative block, Tier-1 canonical, pasted after every still and after every video. The whole source file is 392 words and it feeds prompts that run 40 to 120 words each, which means the reusable half of a prompt is longer than the specific half. 2 blocks, 2 jobs. 1 says what the frame is. The other says what the frame must never contain, and it never has to be remembered because it is never typed.

The character anchor is 1 sentence of 22 words: a young woman in her early twenties, sun-kissed tan, dark wavy hair, wearing a loose cream linen shirt, barefoot, calm quiet expression. It is repeated verbatim in every shot she appears in, and it is written down in the file so it is never retyped from memory at 1am. 22 words carried through a 24 shot set is 528 words of identity spread across the production, and the reason to spend them is arithmetic. A character who is not described re-rolls as a generic stranger, and the probability that a model invents the same stranger twice is not zero, but most people are pricing it at zero, which is why 2 shots generated a week apart come back as 2 different women standing in the same room. So the anchor is split in 2. The frontal half, the face and the expression, only appears when the face is actually in the frame. The angle-agnostic half goes into every shot she is in: hair colour, shirt, build, age, because those are the features that read from any camera position. And a from-behind view names the back-visible features explicitly, seen from behind, dark wavy hair, cream linen shirt, rather than assuming the model will carry them forward from the last prompt.

The lens table is 5 rows and it exists because a renderer does not know scale, only framing, so focal length is the cheapest way to force a decision before the frame is generated. The face, and specifically the hook, gets an 85 mm lens with a shallow depth of field, medium-framed, and never an extreme close-up on a clip of 3 seconds or less, because a 1.5 second cut into a pore-level macro reads as a mistake rather than as intimacy. Over-shoulder gets 32 to 35 mm. Mid and walking shots get 35 to 40 mm with the hands at the sides stated in the prompt, since an unstaged hand is the first thing that drifts. Wide and establishing gets a 24 mm lens with deep focus. And from-behind gets any lens at all, because the face is not visible, which makes it the safest composition a model can be asked for, and the 1 to reach for when a take has failed twice. 6 lenses, 6 shot types, 24 possible pairings, and 1 rule: the lens is chosen by what the shot has to prove, not by what looks expensive.

The style block is not a mood board. It is 1 sentence, and it goes into every render unedited.

The movement list is 7 fragments, and each 1 is written as a move plus its speed: locked-off tripod shot, static, composed. Slow dolly-in, pushing closer, intimate. Slow pull-back, revealing the scale. Tracking shot, moving alongside her. Smooth gimbal glide, floating, steadicam. Aerial drone establishing shot, drifting over. Whip pan, rapid blur transition. The instruction attached to the list is 1 line long and it is the whole point: name the move and its speed, or the model hands back a generic slow drift. That failure is not random. An unspecified camera reads as the lowest-energy interpretation of the frame, and the lowest-energy interpretation is a slow drift, which is why 18 clips generated from 24 well-written prompts can all move the same way and look like 1 clip. 7 moves and 7 speeds is a vocabulary, and a vocabulary of 7 is enough to cut a sequence of 30 beats that does not repeat itself. The difference between a dolly-in and a drift is 1 word in the prompt and about 15 frames of held intent.

Under the lens table and the move list sits 1 more fragment, the temperature anchor, and it is the 1 that decides whether the other 12 decisions look like they were made in the same hour: humid tropical golden hour, 2800 kelvin warm light diffused by milky humidity haze, sweat sheen on the skin, linen clothes, long shadows, atmospheric perspective. 23 words. 2800 kelvin is warm by measurement rather than by taste, and it is the number that ties 4 separate things to the same moment. Light, air, skin and wardrobe have to agree on the colour of the hour or the frame reads fake at 1 glance, and the eye catches the disagreement before it catches anything else in the picture. Skin lit at 2800 kelvin against a shirt graded cooler than 2800 kelvin is the single most common failure in a set that otherwise holds together, and it is not a colour-grading problem at the end, it is 1 missing line at the start. The whole system fits in 1 file: 7 style elements, 1 negative block of 392 words, 1 anchor of 22 words, 6 lenses, 7 moves and 1 temperature. 2 scripts read it. 0 prompts retype it.