How a prompt is assembled from a shot

From shot to prompt basics

You type one line into a shot: “Ivo Renk walks the pier toward the lighthouse, dusk.”

What comes back is a prompt that already knows what Ivo looks like, which picture of him to send, that the picture governs his face and costume, pose, the environment, what the pier and the lighthouse are, the Look the project agreed on, and how long the prompt is allowed to be for the model you chose.

Then it does the same for every other shot in the list for any given project with an attached brief.

Why not just ask your own LLM?

Your LLM will write you a good prompt. What it cannot do is attach the right picture of Ivo to the request, tell the model that the picture controls his face and not the background, notice that you have no picture of him in that costume, keep him consistent across forty shots with different states, and size every prompt to the limits of the model you are actually sending it to.

What a Key Prompt pulls in

Every character, prop and location the shot names. A name in the row’s References field, or an @token in the description, is resolved to its brief entry, to a state, and to a view. Each one brings three things:

its description from the brief, so the shot is specific even with no picture;
its picture, if one is on file for that state;
its reference contract, a sentence saying what that picture is allowed to control. A character’s reads “controls identity and costume only — do not copy pose, background, or lighting from it”; a location’s blocks the camera against the anchor you gave it and rules out the angles you ruled out.

The contract is what makes a reference useful. Without one, a picture is an instruction to copy all of it; with one, it governs exactly the part it should.

The brief, as it is confirmed right now. The Look, the Intent and the entries the shot names are read at the moment the composition runs, so confirm the blocks, then compose.

The scene’s own Look, if it has one. A scene that overrides Light, Emulsion or Color hands its values to every shot in it; a scene that inherits hands nothing and the project’s Look stands.

The shot line, every time. A shot-list row composes from its Description, the line you typed, and never from its own last composition or from an edit you made in the Studio. Composing the same row twice gives the same kind of answer instead of drifting further each time.

What reaches the model

Attached, or held

Before the agent writes anything, Lorvara decides what the generation will receive, and the composition is written to match.

Attached means a picture is on file and the target model will carry it. The prompt gets the @entity name and the contract.

Held means it will not be carried: there is no picture on file for the state the shot resolved to, or the chosen model takes no reference images at all, or the shot names more references than the model can attach. A held reference keeps its written description and drops its token and its contract, whatever the agent wrote. The row names the entry that had no picture, so you know what to add.

On the shot card this is the difference between a green reference chip with a tick and an amber one.

Sized for the model

A composition is written for the project’s default model in that mode, and the record remembers which.

The prompt fits that model. Every model has a hard limit on its prompt field, and they range from a couple of thousand characters to tens of thousands. A prompt composed for a generous model can be too long for a stricter one. To use it there, ask the Prompt Agent in the Studio conversation to refactor it to that model’s limit.

The agent is given the room, not the cap. Lorvara appends text to what the agent writes on the way out (the reference tokens in the target’s spelling and a line naming each attached picture), so the budget the agent writes to is the model’s limit with that measured and taken off. It is measured from the pictures this request actually carries.

How it works underneath

Names now, tokens at send

A stored prompt names a picture by what it is (@ivo-renk, @the-mast) and not by a position in a list. The token the model reads (@Image2, @Element2, or no token at all) is written when the request is sent, from the list the request actually carries.

So a prompt composed for one model opens on another and still reads correctly: the names stay, only the spelling changes. It is also why the @ picker in the Studio inserts a name rather than a number.

The last line states the order. A composed prompt ends with “Reference images, in the order sent: …”, built from the same list the request sends, so the positions the prompt describes are the positions that arrive. If you later attach a picture that pushes the others down, the Studio notices that the stated order no longer matches and offers to renumber it. See The model picker.

Two passes

Lorvara asks the Prompt Agent twice to compose one shot, and charges you once.

The first pass only reads: it compares the shot against your brief and works out what is unclear or missing. The second pass writes the prompt, with those open points in front of it, and settles each one on its best judgment for the shot. Neither pass asks you anything — this is not the Studio conversation.

The split is deliberate. Asked to read and write at the same time, a model answers the parts it is sure of and quietly invents the rest. Making it finish reading first means every gap is named before the writing starts, so each one is decided rather than slipped past.