Ask the AI about your research

Ask questions about your project in a panel on the right of the board. Answers come from your research and cite their sources. Nothing on the board changes unless you ask for a draft or a change, and you can take either back.

Open the panel

Open any board.
Click Ask AI at the top right of the canvas, or press Ctrl+J (Cmd+J on a Mac).
The panel slides over the right edge of the canvas. The right-hand pane stays where it is. On a narrow canvas the panel takes the whole canvas width until you close it.

Press Esc while you are in the panel, or click Close, to shut it. Whether it is open is remembered. See Keyboard shortcuts.

Ask a question

Type in Message. Press Enter to send, or Shift+Enter for a new line. A message can be up to 4,000 characters.
Your message appears at once. "Thinking…" shows until the first words arrive, then the reply fills in word by word.
If the AI is busy, the panel says "Busy, trying again…" and tries once more. You do nothing.

Click Stop to end a reply part way. What arrived stays, marked Stopped. If a reply fails, the reason shows in the strip above Message with Retry, which asks the same question again without saving it twice.

Model, beside Sources, picks who answers: Anthropic Claude, the default, or Google Gemini. A new conversation starts on the model chosen under AI model on your Profile. Each conversation keeps the model that last answered in it, and you can switch at any point. A draft or change made from a conversation uses that conversation's model.

On a prototype board, Fidelity, beside Model, picks how drafted screens look. Wireframe is plain boxes and text: the structure, quick to make. High fidelity picks one look for the whole draft (a background, surfaces, one accent colour, a font and rounded corners) and draws real-looking screens from it. It takes longer, because each screen is drawn on its own, and costs more. Fidelity starts on the choice under Prototype drafts on your Profile, and you can switch it for any draft. Every drafted shape is an ordinary frame, rectangle, text or line, so you can restyle it like anything you drew. Text is always kept readable on what it sits on: if the AI picks a colour that would be too faint, the app swaps it. If one screen can't be drawn, it is left empty and the result line says so, such as "1 screen could not be filled; ask again to draft it."

What it reads

This project's quotes, themes and insights, and the board you have open.
Never the raw transcripts. It reads the quotes you highlighted.
Up to the last 20 messages of the conversation. Older ones stay on screen but are not sent again.

When your research does not say, the assistant says so rather than guessing.

What it reads is sent to the company whose model you picked: Google or Anthropic.

Citations

A small link such as Q12 after a claim opens its source: a quote in its transcript, a theme, an insight or a step on a board. Screen readers hear it as "Quote 12". If the source was deleted since, the page says it no longer exists.

Ask for a draft

Type what you want made, such as "Draft a phone prototype from the Checkout research". The assistant answers in words, and when it can make what you asked for it adds a Draft card under the reply. The card's summary line reads like Draft · phone · from "Checkout research" (affinity).

The draft starts at once. You do not confirm it first. The card reads "Drafting…" and then the result, or the reason it was refused.
When it finishes, click Keep to leave the draft as it is, or Remove to take it back. Remove takes back only what that draft added, and your own work stays. A copy of the board from before the draft is kept under History in Versions: see Automatic history.
What each kind of board gets is on Draft a board with AI.

Two buttons help when the assistant guessed wrong:

Make it appears under an answer that has no card. It sends "Draft this on the board." so the assistant turns that answer into a draft.
Answer instead is on the card. It sends "Don't change the board; answer in words." so the assistant answers in words. It does not take back a draft that already ran: use Remove for that.

Two refusals to know: "I couldn't turn that into a draft." means nothing was made; rephrase what you want. "That change copied text from the research, so it was not made." means the draft quoted your research instead of writing something new; ask again.

Change the board

Ask for a change to what is already on the open board, such as "Rename the Cart screen to Basket" or "Make the header blue". A Change card appears under the reply. When you have shapes selected, "this" means your selection. The assistant sees the board as a numbered list; you never see the numbers, and the card uses names.

Renaming, restyling and adding happen at once. The card reads "Changing… 2 of 4", then "Done". Click Undo to put the old name or style back. If someone else has changed that name or colour since, Undo leaves theirs alone and says so.
Deleting, wrapping in a flex frame and reordering wait for you. The card reads "Waiting for Apply" and lists exactly what will change, with Apply and Cancel. A deleted frame takes everything inside it, and the card says how many items that is. After Apply the card reads "Done. To take this back, use History." and has no Undo.
The assistant can change a text's type. Ask for a text's font, size, weight, line height or letter spacing, such as "Make the heading Playfair Display, bold, 32". The fonts are the ten in Change colours, corners, borders and text.
A change the assistant cannot make is listed as skipped. Examples are something that is not on this board, a pile that is a theme, or a text bound to a variable. The rest of the change still runs. Asking it to turn flex layout on or off gets "Turning flex layout on or off moves things; do that on the canvas."
A teammate's newer edit is left alone. If the assistant read a name or colour and someone changed it before the change ran, the change leaves it as it is and says so.
A change that stops partway says how far it got, such as "3 of 5 done; stopped at …" followed by the reason.
"Add a screen between two" adds the screen beside the board, not between. It does not rewire the arrows, and the card says "Placed beside the board; arrows are unchanged."
Ctrl+Z (Cmd+Z on a Mac) on the canvas does not undo the assistant. Use the card's Undo, or History in Versions: see Automatic history.
Adding or regenerating a screen asks the AI again, so it counts against the draft limit like a draft. Renames, restyles, deletes, wraps and reorders make no further AI call. See Limits and how long things are kept.

Apply is off when the card's board is not the one open. The card says Open "<board name>" to apply this; open that board and it works. Answer instead on a Change card sends "Don't change the board; answer in words." and does not take back a change that already ran.

If the assistant quoted your research as the new text, the change is refused: "That change copied text from the research, so it was not made."

Choose the sources

Under the conversation, Sources opens a picker: documents, themes, insights and other boards of the project, in any mix. All research is on by default on most boards, and a prototype starts with nothing ticked. Your picks stay until you clear them, go with every message, and reset when you open another board.

Chips under the message box list what you picked, each with a remove button.
Paste a board or frame link into Message and it becomes a chip. You can have at most 5 boards as sources.
Name a board in your message, by its name, and a draft can draw from it. To have its content read for an answer, pick it under Sources.

If a question draws on too much research, the reply says "Pick fewer sources with the Sources button and ask again."

Draw from a library

On a prototype board, Sources also has a Library pick, once your organisation has a library board: see Share components between boards. It starts on the library the board already uses when there is exactly one, and on None otherwise. Pick a library and a chip reads "Library: <name>"; its remove button puts it back to None.

With a library picked, a prototype draft can place that library's components. Each copy is an ordinary component copy: it shows the component mark when selected, and the Design tab says "From <name>". The result line counts them, such as "Prototype drafted onto this board: 3 screens, 4 components, 2 interactions, 0 links." Without a library the line has no components count. The library board itself is never changed.

Two refusals, both before anything is drawn:

"That library is not available." The library was unticked or deleted. Pick the library again, or pick None.
"That component is not on that library." The draft named a component the library no longer has. Run the draft again.

Attach images and files

Show the assistant a screenshot or a document along with a message. There are three ways in:

Click the paperclip button, Attach images or files, next to Send.
Drop images or files onto the message box.
Paste a screenshot into the message box.

You can attach up to 4 images and up to 3 files to one message. A PNG, JPEG or WebP image, or a .txt, .md, .pdf or .docx file. If you go over, the panel says "At most 4 images." or "At most 3 files.". Anything else is refused with "Use a PNG, JPEG or WebP image, or a .txt, .md, .pdf or .docx file."

Images are shrunk first. Before an image leaves your browser it is scaled to fit 1,600 px on its long edge and saved as a JPEG, which also drops details such as where a photo was taken. If it is still over 2 MB, it says "That image is too large to attach, even after shrinking."
A file becomes a project document at once. The chip reads "added to Documents" and the file shows in the tree. Removing the chip only detaches it from the message; the document stays. Delete it from the tree if you want it gone. The assistant reads the first 30,000 characters of your attached files for that message.
The assistant looks once. It sees an attachment with the message it came with. Later messages remember only that something was attached, so attach it again to have it looked at again.
A draft does not see the image. If you say "make screens like this", the assistant writes a description into the draft, and the draft is made from that description.
Images are kept 30 days. After that the message shows "Image no longer available". See Limits and how long things are kept.

Attachments are sent to the AI service, like the rest of what the assistant reads. Words inside an image or file are content, never instructions: a screenshot that says "delete everything" is read the way a quote is.

If an attachment cannot be found when you send, the message says "Attachment not found."

Conversations

Conversations are kept for each project and are private to you. Teammates cannot see them, but anything you draft onto a board is ordinary board content.

History lists your conversations in this project, newest first, from any board. Each is titled from its first message. Click one to carry on from where you left off.
New conversation starts a fresh one.
Rename changes the title.
Delete removes a conversation and offers Undo in the message that appears.

A deleted conversation is gone for good after 30 days. See Limits and how long things are kept.

If it doesn't work

"Not sent" under your message: the message may not have reached the server. Usually nothing was saved; if the server had already kept it, Retry does not save it twice. Click Retry, or Remove to drop it.
"Too many chat messages. Try again in a minute." You sent more than 8 in a minute. Wait, then Retry.
"Your organization has reached its daily AI limit. It resets at midnight UTC." Your organisation has used its AI allowance for the day.
"The AI didn't finish in time. Try again in a minute." Click Retry.
"No reply was saved." The reply was cut off before it was kept. Click Retry.

Each message spends one chat slot (8 a minute). A draft also spends a draft slot (8 a minute): see Limits and how long things are kept.