The assistant & director

Two kinds of help that know your project: an assistant that has read your pages, and a director that can run a stage for you. Neither spends a GPU without you knowing.

The project assistant

Every planning pass writes notes into the project — what each page says and shows, what each panel is, what the story does. The assistant chats from those notes rather than re-reading the images, so it is cheap to ask and it knows your comic: who is in this panel, why this shot was dropped, what a page's gutter implies.

It can read any file's insights and propose the next step, with the inputs and settings filled in. Running the step is your click. It also remembers per file, so a conversation about one panel stays with that panel.

The director

Point the director at a stage, or a run of stages, and it works through the loop you would otherwise do by hand: inspect the stage, choose the single next step, run it, wait, judge the result, repeat — until it judges the stage done or a cap is hit.

Control is deterministic and enforced in code, never in a prompt. The model only ever chooses content — which step, which inputs, which settings — never whether to keep going. Three options in the run panel:

Scope

Which stages to walk. Gate between them to park at every boundary and look before it continues.

Scale

How many inputs a step takes — so a whole stage can be judged on three panels before it runs on two hundred.

Resume

Skip the inputs an earlier run already finished.

Pause, resume and stop change the run's state in the database, which the loop checks between steps; stop also cancels the job in flight. A restart of the studio resumes an active run rather than orphaning it.

Re-run with notes

When a planning pass reports a defect on a file — a hole filled badly, a plate with a border still on it — the assistant can re-run the step that made the file with a note built from the defect, under a budget cap. It is the same fix you would type, typed for you.

What it costs

Chat and proposals cost a few credits of language-model time and no GPU. A director run costs exactly what the steps it runs would cost by hand; the sample size is how you bound it.

Last updated 2026-09-17. Credit figures are from the studio's own reference runs; the estimate on the run button is what you are charged.