AI assistant

The ✨ button on the toolbar opens the AI assistant in the right-hand column. You type what you want in your own words, it answers — and when the answer is a change to your page, it does not make it. It proposes it.

The thing to know first: it proposes, it does not apply

When the assistant wants to change something, the change appears on the canvas so you can see the real thing, with a bar at the foot of the panel:

Make the heading bigger and use the brand colour Reject Accept

• Accept keeps it. Nothing further is written — it is already there.
• Reject undoes it, exactly as many steps as the proposal took.

The bar also carries a What changed list — the real before-and-after, read off your document rather than from the assistant's own sentence:

Heading · font_size: 16 → 48 Heading: “October sale” → “October sale — 30% off”

That is the part worth reading before you press Accept. The sentence is the assistant talking; this list is your page. Click a line and that element is selected on the canvas — so you can check each line against the page without hunting for it.

For a colour or a size the list is still not enough to decide; you have to see both states on the actual page. The bar has Before / After: Before temporarily lifts the proposal so you see the page as it was, After puts it back. Accept and Reject mean the same whichever you are looking at — pressing Accept while on Before still accepts the change.

The moment a proposal lands, the canvas flashes an outline around every element it touched — and if the assistant built new blocks, the first one is selected so the canvas scrolls to it. You never have to wonder "did it do anything?" and go looking.

The bar does not disappear on its own. Your page does not change behind your back: what you see on the canvas is what will be saved, and you decide whether it stays.

If you have edited by hand since the proposal arrived, Reject will tell you it can no longer undo it — the most recent step is now your work, not the assistant's. Use Ctrl+Z to step back yourself.

What it can do

You say	It does
"Make this heading bigger and red"	Changes traits on the element you have selected
"Change the heading to October Sale"	Rewrites the words of an element that already exists
"Change every text colour on this page to my brand colour"	Changes many elements at once
"Add a hero section with a heading, a short paragraph and a button"	Builds a new block, words included
"Build a landing page for a new product"	Builds several sections
"Move the About section to the top"	Reorders, or moves an element into another block
"Make three of these cards"	Duplicates an element, settings and all
"Get rid of the About section"	Deletes an element — Reject rebuilds it whole
"How do I make a section full width?"	Looks it up in this manual and answers

It knows your brand colours. Ask for "my brand colour" and it uses the colour you named in your site's palette, rather than guessing at a similar hex code. Which means that when you later change that colour once, everything the assistant built changes with it.

It rewrites the words of elements that already exist, not only the words in blocks it builds itself. Say "change the heading in the second section" and it finds the right element and rewrites it — you do not have to select it first.

A duplicate keeps everything. The copy carries the original's colours, sizes, spacing, pictures and child elements. It is the right way to ask for "another one like this" — asking it to build a new one gets you a fresh one from scratch, because the assistant cannot read the current settings of an element you have not selected.

It will delete, but only when you say you want something gone. It used not to, because Reject could only undo a change and never rebuild one. Now it can: you watch the element disappear on the canvas, press Before to see the page without the deletion, and Reject brings the whole block back — what was inside it and the words in them included.

It will not delete to "tidy up" or to replace something it could have edited instead. And at most five elements at once — beyond that it asks, because Accept is one button and a page that has been cleared is hard to notice.

The inspector moves to the left

The assistant takes the right-hand column, so the inspector moves into the Layer panel on the left and stays there until you close the assistant. There is a back arrow at the top to return to the layer list; clicking a layer opens that element's traits again. Close the assistant and everything returns to where it was.

Conversations are kept

Closing the panel loses nothing. The history button at the top of the panel opens your saved conversations — click one to reopen it, and your next question continues that thread.

The + button starts a new conversation. It does not delete the old one, it only clears the screen: yesterday's work is still in the list.

A conversation stores what was said, not the proposals. A proposal describes your page as it was at that moment, and the page has moved on — so reopening an old conversation will never re-offer a change.

While the assistant is working

It does not answer straight away: it reads the manual, looks over the elements on your page, finds a photo. The panel says each step as it happens:

✓ Reading the manual: full width section ✓ Searching the page: heading ✓ Reading an element ○ Finding a photo: fresh bread bakery

The middle two are the assistant reading your page as it needs to — it does not need you to select or point at an element to learn how it is set. "Make that card look like this one" is answerable because it reads both itself.

Finished steps get a tick, the one in flight gets a spinner. It tells you what the assistant understood your question to be — which is the easiest moment to press Stop and ask differently, rather than waiting for a finished answer to the wrong question.

Pointing the assistant at an element

Select an element and the floating toolbar above it has a ✨ button. It attaches that element to the conversation: it appears as a chip above the box, and on the canvas it gets a dashed outline — you can always see what the assistant is holding.

Typing @ in the box opens a list of the page's elements to search and pick from, matched on their name, their type and their own words.

Why it is worth doing: an element that is merely on the page reaches the assistant as "a heading" — it does not know what size this one is set to, or what it says. An element you point at is sent with both its current settings and its words, which is what makes "make that heading the same size as this one" answerable — and "rewrite this description" answered instead of met with "which description?".

The chips stay after you send, unlike an attached file. You point at a card and then usually ask three things about it; having to point again each time would be tiresome. Press × to remove one, or start a new conversation and they are cleared.

Five elements at a time — each one carries a lot of detail. A chip carries the same name you see in the Layer panel — including a name you gave it — so an element is never called two things in two places.

The assistant points at the canvas

When the assistant talks about a particular element it selects it and scrolls it into view, instead of describing where it is and leaving you to find it on a long page. It can switch the canvas to another breakpoint too — to show you a problem that only happens on mobile, say.

This is not a change to your page, so there is no Accept/Reject and nothing to undo. The assistant is pointing. Click somewhere else and it is over.

Attaching a file or a picture

The paperclip beside the box adds something to your question: a text file, or a picture. Each one appears as a small chip above the box; × removes it.

A text file (.txt, .md, .csv, .json, .html, .svg…) is read as text and folded into the question. It is not stored anywhere — not in your media library, not as a document of your shop. It is context for that one question.

The assistant reads a limited amount of text. If the file is longer than the room left, the panel says before you press Send how much will be cut — and it cuts the end of the file, never the words you typed. Your words are what tell the assistant what to do with the file; dropping them would keep the data and lose the instruction.

A picture (.png, .jpg, .webp, .gif, .avif) is the opposite: it is uploaded into your media library the moment you pick it, before you press Send. It has to be — a picture on a page must be an asset your shop owns, so that it can be tracked where it is used and so it is still there after this conversation is closed.

Once a picture is attached, there are two things to do with it:

You say	It does
"Build a hero section with this image"	Puts that exact picture on the page, rather than finding a stock photo
"What is wrong with this screenshot?"	Looks at the picture and answers

The second is the one people do not expect: the assistant genuinely sees the picture — a screenshot of a competitor's page, a brand sheet, a photograph of your shopfront to take the colours from. Ask about it the way you would ask somebody sitting beside you.

An SVG counts as a text file, not a picture. That reads as odd and is right: an SVG is text, and an attached one is nearly always a logo you want discussed rather than pasted onto the page.

The practical limits: text files under 1 MB, pictures under 8 MB, and at most four pictures shown to the assistant per question. While a picture is still uploading the Send button waits — sending then would hand the assistant a picture that does not exist yet.

Stopping an answer that is going the wrong way

The Send button becomes a Stop button while the assistant is writing. Press it and the words already on screen stay (marked as a stopped answer), and you can immediately ask for what you actually meant.

That call still counts against your allowance — it really did run. The button saves your time, not your quota.

Your allowance, and using your own key

The gear button at the top of the panel opens its settings. There are two things there:

This month's allowance — "12 of 100 calls used", say, with the date it resets. The number comes from your plan. A plan that does not include the AI assistant says so plainly, and the fix is to upgrade rather than to wait.

Use your own key — enter your own API key (Gemini, OpenAI, Anthropic, or an OpenAI-compatible service). Once you have:

• The bill is yours, paid directly to that provider.
• No plan limit applies. The allowance exists for the platform's key; counting your calls in order to refuse them would be metering someone else's spend.

The key is stored encrypted and can never be read back — not even by whoever runs this site. That is why the box is always empty even when a key is saved: to change it, type a new one; to go back to the platform's key, clear the box and press Save.

Changing the model or the API address without touching the key box keeps the stored key. An empty box means "leave it alone", not "delete it".

When it says it cannot

The assistant is allowed to say no, and it says so rather than guessing:

• This element has no such trait. Each kind of element has its own set — a heading has a font size, a divider does not.
• The builder has no such block. It looks up the real element list instead of inventing a plausible-sounding name.
• The manual does not cover it. For how-to questions it answers only from this manual, so it will not describe a button that does not exist.

When a proposal is refused for a technical reason, the panel says plainly that nothing was changed.

New words land in the source language

If you are viewing a page in another language (see Selling in several languages), the assistant cannot edit the words of an existing element — editing there would overwrite the source every other language falls back to. It will tell you so.

Words in a new block the assistant builds land in the source language, so they show under every language until you translate them.