Creative boards

A board is where ideas get spread out. Talking to the agent is telling it what you want; a board is laying it out where you can see it. You place notes, images, videos, audio and group frames — and connecting two of them means the downstream one takes the upstream as input.

Drop an edge anywhere on the visible + target at either side of a node; it snaps there without requiring pixel-perfect placement on the node boundary.

What a connection gives you#

Connections are not decoration. The downstream node reshapes its form around what it receives:

Upstream → downstream	What arrives
Note → image / video	The text fills in the prompt
Note → audio	The text is what gets read aloud
Note → note	The upstream text is sent as material, kept separate from your instruction
Image → image	Attached as a reference image
Image → video	One image becomes the first frame; several become references
Image / video → note	The model looks at it — images directly, video as sampled frames
Audio → note	Its transcript, if one exists, becomes material

What flows in never overwrites what you typed: it only fills a box that is still empty, or one that still holds the previous auto-filled text.

What each node can do#

Select a node: the action bar sits above it, its form below.

• Image / video — while still empty, a prompt panel hangs underneath. Model, aspect ratio, resolution and duration are driven by the model descriptor, so switching models reshapes that row by itself. Once there is a result, the action bar offers Preview, Swap and Trim.
• Note — empty means "write it"; with text it means "change it the way I say", which is what this panel is used for most (you wrote something, now you want it half as long). Double-click still edits by hand.
• Audio — pick a voice and the text to read. Audio here is not a "generation" capability; it is TTS, so you choose a voice rather than a model.
• Group frame — fence things off and name them. Select a few nodes and press ⌘G to group them directly.

Prompt, model, duration, source references and run status live inside the node's own data. Closing an inspector, selecting another node or reopening the board therefore cannot reconstruct a different placeholder or lose the form. Failed jobs settle into a visible failed state instead of spinning forever. Models that declare a continuous duration range get a numeric input for that range.

Every node also exposes its run state directly: queued uses a light primary outline, running adds an animated glow, completed turns green, failed turns red, and cancelled uses a subdued dashed outline. The signal is not confined to a spinner in an empty media slot, so a failed rerun remains visible even when the node still has an older result. A new result or manual asset replacement rehydrates the form from that node's saved data; failure and cancellation keep the inputs intact for retry.

Referencing assets with @#

Press @ in a prompt to reference something from the asset library; the menu follows the caret. What you pick becomes an inline chip in the sentence — "put the person from @creator.png on the street" reads as one sentence. On submit it splits in two: the name stays in the prompt (with several images the model needs to know which one you meant) and the asset itself is sent as input.

Trimming#

Video gets a filmstrip, audio gets its waveform; drag either end to set the range. The problem with typing numbers is not the typing — it is that you do not know what second 3.2 looks like. When you do want frame accuracy, the number fields on the right are still faster.

The original is untouched. A trim produces a new asset in a new node — every step on a board should be one you can walk back from.

The AI assistant#

The robot icon in the top-right opens it. It is not a second AI: sessions, messages and confirmation cards are the same ones the global assistant uses; each message just carries one extra line saying which board you are looking at.

The top-left title is the current conversation. Click it to search or switch sessions; the title ellipsizes when space is tight instead of introducing another framed picker.

It can read the canvas (get_board) and change it (edit_board): add notes, rewrite text, connect, delete, move things. Every change goes through a confirmation card — nothing lands until you approve it. It submits granular operations rather than a whole canvas, because a model handed the whole thing gets it wrong on any non-trivial board: items go missing, or the positions you arranged by hand get flattened.

Undo#

⌘Z undoes, ⌘⇧Z redoes, and the toolbar has buttons too. A continuous drag collapses into one step — dragging a node emits dozens of position updates, and one-per-update would mean dozens of presses to get back.