Recipes
Every recipe is a row with two halves. One half is locked before you arrive: the model, the setup and the prompt that took a week to get right. The other half is four questions only you can answer.
What it already knows
{{creative_prompt}}your answer{{slot_instructions}}your picture{{choice:angle}}your pick{{choice:background}}your pickProvider prompt templates remain private.
What it asks you
Listing ShotReady
Create·3
Four stops, then it runs, for 3 credits, the same as typing the prompt yourself.
Two documents
Both wear the same card, ask for what they need and cost credits to run. What differs is what is behind the press — one prompt, or a board of steps that feed each other.
Asks a few questions, fills one prompt, makes one thing per press.
The anatomy above, and most of the catalog. The model, the settings and the wording are decided; your answers drop into the slots the author left. One press, one picture.
Runs several steps on a board — write, then make — and can make many things per press.
A document with declared inputs and an ordered list of steps, drawn on a node canvas. A step can read what an earlier one produced, and one step can make many things. The rest of this page is about this one.
The workflow
Inputs down one side, steps down the other. Drawing a wire from one card to another is not decoration — it writes the read into the document, and cutting it takes the read back out. There are no conditions, no branches and no parallel joins, because the work this exists for does not need them.
Up to 8 declared inputs, filled before the run starts. The reader answers them once, in the column, and the run carries the answers from there.
The six things it can ask for
An image, a clip or a sound from what the account already holds — up to 8 of them. Uploads and finished generations are the same kind of reference here.
Free text, dropped into any step that names it. The author sets how long it may run; a generous bound is the signal the field should rest tall.
A whole number, inside bounds the author sets. This is what a step repeats over when the run is told to make several of something.
The engine and the voice, picked together. The step that speaks reads both off this one ask rather than binding a voice the author chose months ago.
Named identities off the account's own shelf — a character, a location, a product. Tagging one supplies its face and its description; the server resolves the bytes and seats them against what the model can carry.
A picker, not a catalog: up to 12 options the author chose. It is how a workflow lets the reader trade quality for price without rewriting the board.
An ordered list. Each step names the values it reads with a dotted path — an input, an earlier step's output, or the item it is repeating over — and a path nothing can produce is refused when the workflow is written, not when it runs.
The ten things it can do
In the order a score is written: read what came in, think about it, make a picture, a clip, a sound, and finish.
Measures the cuts, the keyframes and the pacing with ffmpeg, then asks a model one question only: which shots are one beat. The timings are measured, never guessed — there is no field an analyst could write a timestamp into.
One language-model call that writes text. It can look at the run's images, and it answers one string, a list of them, or records with fields the author declared. It is what turns one brief into six prompts.
One ordinary generation, through the same door the composer uses — same credits, same moderation, same library. Repeat it over a list and it makes one per item.
The same, for a clip: one generation per item, on whichever video model the board names or the reader picked.
A voice-over or a music track. It takes a script and a voice — never a reference image — which is why it is its own step rather than a third media kind.
One of the finishing tools — remove a background, upscale, join clips, burn subtitles. It binds a model and takes no prompt, because a tool makes a file and a step makes a decision.
One speech transcript per media object, keyed to the bytes rather than to the room they arrived in. A repeat ask for the same file costs nothing, because the answer is already stored against it.
The run stops and waits for you. It can show up to 4 of the things it has made so far, and it can quote what continuing will cost before you say yes.
Up to 6 tracks on a 9:16, 1:1 or 16:9 canvas — a sequence of clips, an audio bed at a volume you set, burnt-in captions in a style you pick. The multi-track document, written by the board instead of by hand.
Lays a film out as scenes on a board — a title, a description, a still prompt and a motion prompt per beat. It calls no provider and writes no bytes: the run stops at the board, and every clip is bought in the Studio, one scene at a time.
Wiring a later card into an earlier one moves the reader behind what it reads. The board cannot draw a document the engine could not run.
Point a step at a list and it runs once per item. Six prompts become six images in one press, each one an ordinary generation with its own row, its own credits and its own place in the library.
Which generations a step created are rows, written one at a time as each is made. A run that dies halfway resumes from the number of rows it finds, so a crash costs you time rather than a second charge for the same picture.
Limits and runs
The limits depend on whose recipe it is. Write one yourself and you get the wide ceiling; press somebody else's published workflow and the narrow one applies, because their document is spending your credits.
| What a workflow may do, and whose credits it spends. | A workflow you wrote | A workflow somebody published |
|---|---|---|
| Steps | 20 | 6 |
| Steps that write | 5 | 3 |
| Things one step may make | 30 | 10 |
| Runs at once | 3 | 3 |
The one row that is not about your bill: a step that writes runs on our language model and is not charged to you, so it has a ceiling of its own.
Pressing a workflow starts a run — a row with a status, not a spinner. The same board redraws as the record of it: each card carries what it made, so you are reading the document you wrote rather than a log of it.
The catalog
Every public recipe and every style one can be made of, in one wall. Free to read, credits to run, and each one carries the whole prompt so you only answer what it asks.