Talking with the Assistant

The Composer, the prompt at the bottom of a Canvas (and of a Space, see The Space’s Composer), has a mode. Its button sits at the start of the prompt, and its icon shows the mode you’re in. Press it to choose another:

Ask (the sparkle) talks with the Assistant. It’s where you start.
Image makes an image from what you type, straight away (see Making images).
Video makes a video, straight away (see Making videos).

The Composer always opens in Ask, which costs nothing. Image and Video are for one prompt: once you send, the Composer goes back to Ask. If what you sent is refused, it comes back in its mode, with why. In Ask there’s no price beside the send button, because talking costs nothing. Settings (the sliders button) offers Auto (the default), OpenAI or Claude to answer you. Your choice is remembered in this browser. If the one you chose isn’t available, the Assistant says so rather than using the other; Auto always uses what’s available. Image and Video have their own price and model menu, because sending there makes something.

Asking

In Ask, write to the Assistant as you would to a person: “I’m thinking of a poster for a jazz night”, “How could we make this more cinematic?”, “Generate an image of a cat on a windowsill”. Its reply appears as it’s written, in a bubble just above the Composer.

The bubble stays while the reply is being written, and while your pointer is on it. If it has a prompt card to use, or asks you something, it stays until you dismiss it (×) or ask something else. Otherwise it goes when you click elsewhere, or after a little while. Show reply brings it back.

Open conversation (the arrows at its top) docks the Assistant right with the whole conversation. Docking the Composer right does the same. Nothing is lost when a bubble goes: every message is kept in the conversation.

Where you put the Composer is where it stays, everywhere: docked right or floating at the bottom, it’s the same on a Canvas, in an image’s editor, and in a Space’s Work and Assets. Floating, it sits in the same place in the window in each. Docked, drag the Assistant’s left edge to make it wider or narrower, from 320 pixels up to half the window; the handle that moves it sits halfway down, just outside its left edge. Your choices are remembered in this browser.

The Assistant works out what you mean:

A question or an idea gets an answer, ideas, or one short question back.
A suggestion comes as a prompt card: the prompt it would send, and whether it’s an image or a video. Use puts the prompt in the Composer, in that mode, for you to change and send. Make it makes it as it is. A card is made once, and then shows that it was made.
A clear request (“Generate…”, “Make this into a video”) is made straight away, with no “are you sure?”. It’s priced and charged like anything you make in Image or Video.

If you’re not sure the Assistant will read you right, ask. When it’s unsure, it proposes rather than making.

What’s made from the conversation appears there with its progress and result, and on the Canvas as usual. So do the images and videos you make in Image or Video mode. The Canvas’s conversation also shows what was made on the Canvas before you started talking, and what the editor and Iterate make.

Working from images

Images you’ve chosen with Use as reference, dragged onto the Composer from the Canvas or the library, or pasted into it, are sent with what you ask, named as the Composer shows them (Image 1, Image 2). The Assistant looks at them, so you can ask “I need a prompt to make something like this” and get one straight away. Once your message is sent they’re let go, ready for the next one. To show the Assistant what’s wrong, take a screenshot and paste it into the prompt (Ctrl/⌘ + V): it’s added as a reference, like a dragged image, in Ask, Image and Video. A pasted image is kept only as a reference: it isn’t added to your assets or placed on the Canvas, and one you never send is removed after 30 days. Pasting text still pastes text. What the Assistant proposes or makes from that message is made with them, and Use puts them back in the prompt along with the card’s prompt. In an image’s editor, the image itself is what the conversation is about; references are the other images you’ve chosen.

Seeing and changing the Canvas

The Assistant knows what’s on the Canvas: every image, video, shape, note, text and frame, what it says, and how each image was made. It knows what you’ve selected, so “make this one blue” works. To keep replies quick, it looks only as far as your question needs:

When you ask about layout, or several things (“line these up”, “put a note next to the photo”), it also sees what’s around you, where everything is, and a picture of your view.
When you ask about everything (“tidy up the whole canvas”), it sees the whole Canvas.
It can also look further, or closer at particular images, whenever it needs to.

Ask it to change the Canvas and it does, straight away:

Add shapes (rectangles and ellipses, with a fill, outline and words), text, sticky notes and frames.
Arrange: move, resize, line up, space out and group things.
Select things for you (“select all the cat pictures”).
Delete things. A deletion always asks first: Delete, or Keep to leave things as they are.

Each change shows in the conversation as it’s done, with Undo to put things back as they were. Undo works as long as nothing it changed has been changed again since. It’s also one step of the Canvas’s own undo (Ctrl+Z), and everyone on the Canvas sees it at once. Only the Canvas’s own conversation changes the Canvas: an image’s conversation in its editor can look at it, but not change it.

Making, and what was made

While an image or video is being made, the conversation shows its shape, then its picture once it’s ready. Cancel stops it: you’re never charged for it, and its place on the Canvas goes. On a Canvas, a node still being made has Cancel in its menu too (right-click it, or press and hold). An image the Assistant made opens in its editor when you click it.

Hover a message (on a touch screen, they’re always there) to Copy it, or Reuse prompt to put what you said back in the Composer, in the mode you said it in, with the images it was sent with. A proposed prompt has Copy beside Use. Images in the conversation have Use as reference, to add them to what you send next.

An image or video’s Metadata (right-click it on the Canvas) shows the prompt that made it, with Reuse prompt: the prompt goes back in the Composer, with the model it was made with (if you chose one) and the images it was made from. Reuse prompt is also in its menu on the Canvas (right-click it, or press and hold it on a touch screen), beside Metadata, for anything the Assistant made.

Paste a prompt on its own into Ask (one that made something before, say), and the Assistant offers to make it, as a card with Make it, in the medium it describes. Asked to rewrite a prompt you’ve pasted, it proposes the new one: nothing is made until you choose Make it.

An image’s own conversation

Open an image in its editor and the Assistant is there with that image’s own conversation, docked or floating as you last left the Composer. Docked, the editor makes room for it. The Assistant can see the image as it is now, and knows how it was made. Ask it about the image (“What would make this moodier?”), use its cards, or switch to Image to change the image now, or to Video to animate it. A video is placed beside the image on the Canvas.

Ask it to change the image and it does, using the editor’s own tools:

Remove the background, delete a mask, split a layer into its subject and background, and show or hide a layer are done straight away.
Relight, Upscale and extending the image to another shape (“make it widescreen”, which fills the new areas) cost money, so they ask first, with their price: press Relight, Upscale or Fill to make it, or Not now to leave the image as it is.
A change described in words (“make the jacket red”) edits the image, or just the regions you’ve drawn, as Image mode would.

The Assistant knows the image’s layers, so “hide the relit layer” works. If you don’t name one, it uses the layer you’ve chosen, or else the picture itself. Each change shows in the conversation as it’s done. Undo it in the editor (Ctrl+Z), as you would any change there.

Each image keeps its own conversation. Close the editor and open it again, even days later, and its conversation is there. It’s the same conversation wherever you open the image: from a Canvas, from the Space’s library, or by its own address, and even when the image is on several Canvases. Other images’ conversations, and the Canvas’s, are kept apart. An image made from another (by Iterate, or a change made beside it) starts a conversation of its own.

In the Space’s library, the Composer talks in the Space’s own conversation. Every conversation in a Space knows the Space and its newest images and videos.

Limits

Everyone who can edit a Canvas can talk with the Assistant there and sees its conversations. People who can only view it see no Assistant.

Talking with the Assistant is free during the Public Beta, whichever model answers. You can send up to 300 messages a day. Once they’re used, the Assistant says it’s back tomorrow.

Not available yet: audio or 3D modes.