The agent dock takes natural-language instructions. There is no command syntax to learn — just describe what you want.
What good commands look like
| Goal | Effective command |
|---|---|
| Generate one image | “Generate a wide-angle shot of a mountain at sunrise, use @reference for the lighting.” |
| Animate a scene | “Animate @hero-render floating up and down over a soft purple gradient.” |
| Organise the board | “Group every image by colour, then put them into a horizontal frame.” |
| Find references | “Find five similar images to @hero and add the best three to the canvas.” |
| Build a video | “Open the video editor. Add @clip-a, @clip-b, @clip-c in order, fade between them, export at 1080p.” |
Short and concrete works better than long and abstract. Spell out what you want to happen, in the order you want it.
What the agent can act on
@references — explicit pointers to canvas elements. Always preferred.- Current selection — if you ask “move this to the top-right” while one element is selected, the agent assumes “this” is the selected element.
- Active board context — the agent knows which board you are on, the current view, the recent generation history.
What the agent cannot see (yet)
- Elements outside the active board (it cannot peek into other boards).
- Files on your computer outside the connected folder.
- Browser history, email, calendars, etc.
The Board Agent is sandboxed to canvas actions and AI providers you have configured.
When to break a task in two
The agent works best in 1–3 step tasks. If you ask for a 10-step plan it will usually do steps 1–4 well and degrade. Break complex tasks into separate prompts:
- “Find references for a winter character scene and add five to the canvas.”
- (review the results)
- “Pick the best three, group them, and lay them out in a horizontal frame.”
Cancelling
While the agent is working, the Send button turns into Stop. Click it to abort. Any pending action cards stay visible for review or removal.
Leave a Reply