One direction, many grammars
A prompt is a translation, and every model speaks a different language. The Prompts chapter holds two things: the Prompt Pack — the direction written out in full, structured prose — and the Platform Outputs, where the same direction is re-rendered in the native grammar of each model you selected for the brief.
This is the correction to the industry’s laziest habit: writing one generic prompt and pasting it everywhere. Midjourney thinks in compressed keyword phrases and parameter flags; a video model wants motion and duration; a photorealism model wants materials named outright. Feed them all the same sentence and each one hears a different fraction of it.
The engine instead holds one schema — the direction itself — and lets a per-model adapter do the speaking. The vision never changes between tabs. Only the accent does.
The Prompt Pack
The pack is built directly from your creative schema, clause by clause: the core theme, the subject and posture, the environment and its materials, the lighting style, direction, contrast and visibility logic, the lens, angle, framing and crop, the wardrobe and fabric, the compositional rules, the palette, the realism level. Nothing is invented at prompt time; everything traces back to a decision you made in the brief.
It also carries craft you did not have to write. A photographic realism layer disciplines surfaces — highlight rolloff, shadow depth, material separation. Where the brief signals it, a body-casting clause is embedded, because models default to sample-size bodies unless directed otherwise, and the House does not leave casting to a default.
Every prompt is checked before it reaches you: if any value failed to resolve, generation stops rather than shipping a broken prompt. What you copy is whole or it does not exist.
How an adapter thinks
Take the Midjourney adapter as the worked example. It reads the camera crop off the schema and turns it into an aspect-ratio flag; it reads the emotional register and sets the stylize value — a surreal brief pushes stylisation up, an architectural one holds it down. Lighting styles are translated into the model’s own vocabulary: a Rembrandt setup becomes ratio language, a split light becomes sculptural division.
Fabric gets the same treatment. Silk becomes highlight migration; leather becomes specularity; wool becomes fibre-depth definition — the physical behaviour the model should render, not just the noun. The suffix pins version, raw style and quality so the output is repeatable rather than lucky.
Each platform tab then offers several readings of the direction: a primary, a darker alternative, a cinematic variant, a minimal one, a simplified fallback and a high-detail pass for the frame that has to survive scrutiny at full zoom. Same schema throughout — six ways of asking for it.
Generate here, or carry it out
Some platforms the engine can drive directly. Where a model is natively supported, the Generate button loads the prompt straight into the in-house generator with the right model already selected — no copying, no re-typing, no drift between what you read and what you ran.
Others — Midjourney and Sora among them — live outside. For those, the button becomes a copy action, and the pack is your export: lift the prompt and carry it into the tool’s own interface. Either way the copy buttons sit on every prompt, because the pack is meant to leave the building. It is a deliverable, not a private note.
Reading prompts as a director
Do not treat the pack as a black box to paste from. Read it. Every clause in the prompt is one of your decisions made legible, which means the pack doubles as an audit of the brief: if a prompt reads wrong, the direction is wrong, and the place to fix it is the schema, not the prompt text.
And when a model disappoints, compare its adapter output against the others before blaming the direction. The same vision spoken five ways will fail in five different accents — the failures tell you which model misheard, and which clause it misheard. That is intelligence you can use on the next brief.