Let’s Connect
Chapter 02 · Foundation

Prompt Physics

The core engineering idea of this entire brain: an aesthetic is a set of independent control axes, and a prompt is a weighted parameter stack, not a description. Name the axis, choose its value, and the look becomes deliberate and transferable.

"Moody" is an output. 2900K single practical, underexposed one stop, warm lifted shadow is the cause. Models render causes reliably and adjectives vaguely. Name physics, get the look.PROMPT THE CAUSE, NOT THE ADJECTIVE

The six-axis budget

Treat every prompt as a specificity budget. Light gets the most words because it drives the most result. When output looks wrong, diagnose which axis is starved. It is almost always LIGHT or ATMOSPHERE.

LIGHT40 to 50%
Direction, quality, time, colour temperature, number of sources, contrast ratio. Starved light = flat midday phone photo.
ATMOSPHERE20 to 25%
Haze, bloom, glow, flare, diffused air, dust. This is the layer that separates "cinematic" from "clinical AI-sharp".
DEPTH15%
Foreground element, midground subject, blurred background, shallow depth of field. Without it: sticker-on-background.
GRADE10%
Palette, black level, saturation, contrast curve. Cohesion and elegance. Starved grade = garish or generic.
COMPOSITION5%
Rule of thirds, negative space, subject scale, eyeline. Calm or tension is decided here.
TEXTURE5%
True skin and material texture, organic film grain around 10%, never plastic, never fake glow. The realism anchor.

The universal prompt skeleton

One skeleton, filled in order. For diffusion image models lead with light and atmosphere; lead with subject only when identity must lock.

SKELETON · UNIVERSALALL MODES
[SUBJECT / SPACE]
+ [LIGHT: direction · quality · time · colour temp · sources]
+ [ATMOSPHERE: haze · bloom · flare · air]
+ [DEPTH: foreground / midground / background · aperture]
+ [GRADE: palette · black level · saturation · contrast]
+ [CAMERA: body · lens · stop · ISO]
+ [MOTION: video only, one clear verb]
+ [TEXTURE: true texture · organic grain · soft not plastic]
+ [NEGATIVES]

The density rule

Shorter prompts render better than longer ones. Target 280 to 400 words for a single-shot video scene; never over 600 even for multi-shot sequences. Every word must do work. When in doubt, trust the reference image to carry visual information and cut the redundant description. A bloated colour paragraph with a vague light note yields a flat image every time.

The teachable unitWhen documenting any lever for a student, always give the same four-part shape: what it controls · the engineering value · the vocabulary · the failure mode. That structure is how a look becomes transferable instead of a lucky accident.