Using it in a workflow

It lives in the column down the right, and two controls open and close it:

• Once it is collapsed the canvas fills the screen and that AI Agent button floats at the far right of the top row — right where the column used to be, so it expands back in place.
• Collapsing only hides it: the half-typed message and the conversation so far are all still there (see 12.1).
• The open or closed state is remembered: next time you open this project it looks the way you left it.
• The header has one more button, Full-screen chat, just left of Collapse: it spreads this column across the whole workspace — that is, it switches to the chat view (see 1.4) — and the canvas steps aside. Exit fullscreen brings the workflow back.

• Click a node on the canvas and a reference to it turns up in the input box on the right.
• Marquee-select several and you get several: lasso five storyboard frames, then say "make them all low angle", and those five are what it gets.
• Deselect and the reference drops off by itself. Clicking the same node again and again never stacks up duplicates.
• So the easiest thing to say is this and these: "lower the camera on this one", "join these into one" — which one "this" is comes from the selection.

Use this when you don't want to disturb what's selected on the canvas: type an @ in the input box.

• Keep typing to search (names and prompts both match), then press Enter or click to pick one.
• Anything already wired in as a reference comes first — in one turn those are usually the ones you point at over and over.
• You can @ a model too: name it in the sentence, as in "use Seedance 2.0 for this one", and it stops picking a model itself for that turn.
• When the canvas holds a lot of nodes the list shows one page at a time, and the last row loads more.

Two other ways in: the plus (or pasting straight in) brings images or videos off your own machine (several at once if you like) — a strip of attachment cards appears above the text first, and once uploaded they land as nodes on the canvas and go out with this message (see 12.1); typing / picks a ready-made routine (see 12.2).

• All of it lands on the canvas, the one you have open — it can't switch canvases and it can't create one (see 12.3).
• You'll see a few empty cards appear first, then the images arrive one by one: the Agent first fills in each card's name, prompt and model, then starts the job. Queue and generation progress are drawn on the node (a breathing background and a percentage), not as a number to track in the conversation.
• For a video, which runs for minutes, the timeline first says submitted, then adds the result once everything has settled. Trust the node on the canvas.

Under every action in the conversation are cards for the nodes it touched. Click a card and the canvas moves the viewport there and selects it — exactly like clicking a row in Canvas elements on the left. Once a film gets big this is the quickest way to find something: no hunting around the canvas.

Some shots are too hard to pin down in words alone: who moves first, when the camera pushes in, whether two people overlap. Ask the Agent for a moving previz first, review it, and only then commit to the final generation.

How to start: select the relevant characters, set pieces or storyboard references, then tell the Agent the shot count, duration and purpose. For example:

What the Agent does: it places a previz node on the canvas, arranges character positions and actions, sets up each shot and its camera move, then produces something you can play. When a shot needs finer timing, it can split the action or camera move into stages inside that shot.

How to review: don't stop at whether the first impression feels smooth. Check every shot for four things—

The previz includes check reminders. A reminder is not a failure: it means “this shot may expose a visible problem,” and clicking it takes you to the shot. For small timing issues, ask the Agent to change one shot—“hold shot 3 for half a second before pushing in.” If the direction is wrong, redo the shot—“shot 5 should look in through the doorway.”

After you approve, the preview lands on the canvas with a draft “make the final video with this” node. Add style, lighting, wardrobe and sound requirements, then send; the later generation follows the previz's blocking, positions and camera movement. To fine-tune the timeline yourself, open the previz stage from the previz node (see 8.8).

Most of the structural work is within its reach now: wiring references, grouping and placing are written along with the node, and running a whole group, laying out storyboards in bulk are done with the node tools (see 12.3). What is genuinely yours: it cannot create a panel node like the style sample; the editing bay, the film study card and the previz stage it can all build — it can cut the videos on the canvas into one clip as you describe and render the finished cut straight out (see 7.3), a film study card it normally builds for you to open and run, though it can run one on your word as well (see 4.8), and the previz stage is under “Ask for a previz pass first” above; uploads plus interface-level things like history, toolboxes and your account are yours too. The split that still works well: let it generate the content, and the shape of the workflow is your call (see 9.1).