Chatting with the assistant

The Assistant panel is where your website actually gets built. You type what you want in everyday language and the assistant makes the change, then the preview updates so you can see it.

Choosing your AI model

A new website starts on the model we currently set as the default for new websites. To change it, open your website's Settings → AI Settings, pick a model, then choose Save default model. The choice applies only to this website, starting with your next request. It does not delete your pages or chat history, and it does not change an update already running. If we later change the default for new websites, your website keeps the model it has.

The model choices and Model cost guide use the same current catalog. Compare the strengths and prices shown for each model. When supported models change, both lists update together.

OpenAI offers Sol. Anthropic offers Claude Fable 5.1 for demanding builds and complex redesigns, and Opus 5 for advanced website development. Models marked unavailable cannot currently be selected.

Some models are premium, and your organization's plan decides whether you can use them: Claude Fable 5.1 is included from the Builder plan up. On Free and Starter it stays in the list, marked Premium and locked, with a note naming the plan that unlocks it and, if you manage billing, a link to upgrade. If your website was set to a premium model before your organization moved to a smaller plan, the page says so and the assistant answers nothing on that website until you pick a standard model or upgrade — it is never quietly switched to a different model. A new website in an organization without premium models starts on a standard model even when the platform's default is a premium one. See Organization plans.

Terra, Luna, Sonnet and Haiku have been discontinued. Existing Terra/Luna websites move to Sol; Sonnet/Haiku websites move to Opus. These replacements use more credits, with rates shown in the cost guide. Your pages, chat history and past usage are preserved.

The Model cost guide shows your charged rates, including the 20% markup. Input is the context the model reads; output is the text and code it creates. Cached context has separate, lower read rates. Claude's cache writes use a five-minute lifetime. Very large OpenAI requests have a long-context surcharge; that surcharge does not apply to Claude. All models draw from this site's same AI-credit allowance. Fable uses credits fastest.

Changing the builder model does not change the separate image-generation service used for site icons and branding.

Your first message

If your website is empty, start with your business rather than with web design. Something like:

I run a family Italian restaurant in Portland called Vitello's. We do lunch and dinner, Tuesday to Sunday, and we're known for our wood-fired pizza. I need a homepage, a menu page, and a page with our hours and address.

That's enough. The assistant writes the pages, the wording, and the design, and you refine it from there.

Press Enter to send, Shift+Enter to add a line without sending. While a build is running you'll see its progress in the panel — and the message box stays open, so you can line up your next request without waiting (see Queuing your next request, below).

Queuing your next request

The assistant works through your website one request at a time, but you don't have to wait for one to finish before asking for the next. While it is working, the message box stays open: type what you want next and press Enter (the hint under the box reads Enter to queue). The message appears in the conversation straight away, marked Queued, and goes to the assistant the moment the current update finishes. Queue as many as you like — they run in the order you typed them, each as its own update, with its own reply.

A queued message is yours to take back until it goes: press Remove under it. Photos and files you attached to a queued message are already uploaded, so they travel with it when its turn comes. Reloading the page keeps the queue; it carries on once the update that was running has landed.

What your message is waiting for

While a message waits, a notice right above the message box says what is running and who started it — the request a teammate asked Max for, in their words, a colleague's import or save, a publish — with how long it has been going. Anyone who can edit the website can press Stop there to end that work; a stopped request of Max's is undone the way it always is, and your waiting message then goes.

The same notice appears when something else is holding the website even before you have typed anything, so Enter to queue never comes as a surprise: a request of Max's running in another conversation or another tab, a teammate's save, an import or a publish in progress. Your message waits behind it and goes by itself.

The pages finishing in the background after an update are not one of those things. The line at the top says so while it happens, and nothing waits behind it.

Stopping the update that is running does not touch the queue: the next queued message goes ahead as soon as the stop has finished and the website is back the way it was. If the next message should not run after a stop, remove it first.

Never turned away for being busy

A message is never refused because the website is busy. If it was sent while something had just taken hold of the website — a save another member made a moment earlier, say — it is kept, marked Queued, and sent again by itself a few seconds later. The same goes when Max is already working on several of your websites at once, when an earlier request on this website is still being wrapped up, or when the platform itself is at capacity: the notice above the box says which it is, and the message goes as soon as there is room.

On a very large website — thousands of pages — Max's change is saved at once, but the pages are rebuilt in the background: the bar above the preview says the website is compiling, the preview shows the change when it finishes, and Publish waits until then.

If Max did start an update and found the website held when it came to write — another update finished at the same moment, say — it tells you so in its reply, changes nothing, and you are not charged: the reply reads Not charged · the website was busy where its credits would be, and the request goes back to the front of the queue to be sent again on its own once the website is free.

If the assistant can't take the next message when its turn comes — your website has run out of AI credits, say — the queue pauses and tells you why, with the message kept at the front and a Try again button for once that is sorted out. When the reason is credits, the notice also carries the fix: Buy credits if you manage billing, or the people who can buy them, with an email link each, if you don't. An organization that still bills per website is sent to that website's Billing page instead. A message you send while nothing is running is kept the same way, so you never retype it. If Max is still finishing another request, the message waits and goes by itself (see Never turned away for being busy, above). If it was refused because the model needs a larger plan, the notice links to AI Settings, where you can pick another model.

When the assistant turns a message away

Max never spends credits it cannot bill, so a message is refused, kept, and shown with the fix in the builder, over the website you are working on. You stay on the same page. After the fix, press Try again and the message goes without retyping.

• Out of credits. An owner or Admin sees Buy credits, which opens the packs right there. Anyone else sees the owner and each Admin by name with an email link, and no button.
• A model your plan does not include. An owner or Admin sees Upgrade plan, naming the plan that includes it. Anyone else sees who to ask, and can pick another model in AI Settings.
• A paused organization (no plan, or a payment that failed). An owner or Admin sees Upgrade plan, or Update payment details when the payment failed. Your websites and their files are kept. Anyone else sees who to ask.
• Several websites on an old per-website plan. Past the free allowance the builder is paused while the organization moves to an organization plan. An owner or Admin sees Contact us, an email to us about the organization's billing. Anyone else sees who to ask.

The same credit notice appears when a clone, the logo generator, the branding image generator or the site icon generator is refused for credits. Max in the form designer shows the same notice above its box, keeps what you typed so you can send it again after the fix, and clears the notice on your next message.

Speaking a prompt

The microphone beside the message box lets you talk instead of type. Click it, allow the microphone if the browser asks, and speak. The words appear in the box as you go so you can edit them before sending. Click the microphone again to stop. It is there on the website builder and on the form designer; browsers that cannot transcribe speech hide the button.

Stopping the assistant

If a build is heading somewhere you didn't intend, press the Stop button under the reply being written — it sits beside the update's label, Quick edit or Complex update, in the conversation. It stops the work on our side, not just in your browser, so you can describe what you actually wanted straight away. (The same Stop is on the amber bar across the top of the builder while the assistant works.)

Stopping also undoes the build. Your website goes back to exactly how it was before that message: pages the assistant had already rewritten are restored, and pages it had created are removed. Nothing is left half-finished, so you can describe what you actually wanted without first untangling a change you stopped part-way through.

That undo is itself recorded, so nothing is lost for good: if you stopped by mistake and want something the build had written, the revision history in the Files panel still has it. A stopped build is not charged against your AI credits.

Usually the website is back before the button finishes. If the assistant happened to be mid-save at that exact moment, the undo finishes a few seconds later — the preview catches up on its own.

The Stop button is there whenever the assistant is working on that website, including when you have just reloaded the page in the middle of a build — then it sits beside the Max is still working on this line.

Working on more than one website

Each website builds on its own. If the assistant is busy on one of your websites you can open another and keep working there — one build never blocks another. Within a single website, though, the assistant handles one message at a time: anything you send while it's mid-build waits its turn as a queued message (see Queuing your next request, above). Press Enter to send, Shift+Enter to add a line without sending. You can stop a build with the Stop button under its reply if it's heading somewhere you didn't intend.

One limit applies across websites, and it makes a message wait rather than refuse it: Max builds at most three of one person's websites at once. Each website runs one request at a time, and a website never has to wait for another — chat on one website, switch to another and chat there, and both updates run side by side, on an organization plan as on any other. When a message does wait, the notice above the message box says why, and it goes as soon as the other work is done.

Short answers

The assistant keeps its replies short. A question gets a sentence or two, a change gets a line or so saying what changed, and a new website gets a brief list. It skips recaps and suggestions you didn't ask for, but it always tells you what you need to know: a page it still has to build, a detail it needs from you, or a change it couldn't make. Want more detail? Just ask — "explain why" works.

It also knows which AI model your website uses, so "which model are we on?" gets a straight answer.

While it's working

The reply begins as three pulsing dots — the assistant is composing, the same way a messaging app shows the other person typing. When the reply is ready it takes the place of the dots.

Underneath every reply is a More Details row. Open it to watch the assistant work step by step — the section it's building, the page it's writing, the search-engine files it's setting up — with a tick beside each step as it finishes. It stays there after the reply lands, so you can look back at how a change was made at any time. Leave it closed and the panel stays quiet; the reply is the part that matters.

Existing websites show Quick edit, Website update, or Complex update according to the scope of the request. A small change to one selected element can be a Quick edit; centering, aligning, or moving something — like the menu — uses Website update, because where it sits depends on styles shared with other pages. Matching a live site's layout and behavior requires broader research. These labels do not change your selected AI model.

You can ask directly for a clear change such as “Align all breadcrumbs left.” You do not need to split that instruction into extra messages. The preview updates as soon as the changed pages are ready, even if the assistant is finishing its reply.

Leaving while your first website is built

A brand-new website can take several minutes to build. While it does, a bar under the canvas tells you it is safe to leave: the build runs on our servers, not in your browser, so closing the tab or navigating away does not stop it.

Press Email me when ready in that bar and the bar confirms the address we will write to — the one on your account. When the build finishes, you get one email with a link straight back to the builder. Only the person who pressed the button is emailed. If you stop the build, or it does not finish, no email is sent; open the builder to see what happened.

When you started from an existing website or a design image, the build is several rounds rather than one: the first version, then passes that compare it with what you gave us. The email waits for the last of them, and while they run a banner across the top of the canvas — with a pulsing amber dot — says We're still putting the final touches on your website, so a website that has appeared but is still changing never reads as a finished one.

If your browser loses the live view of a build

Your build does not run in your browser's connection: it runs on our servers, and your browser watches it. A long build — cloning a website from a URL is the usual one — can outlast any one connection, and when that happens the builder quietly reopens it and picks the reply up from the beginning, so the words you were reading never disappear. Nothing is lost. If a connection cannot be reopened, the builder keeps watching for the build and brings in the finished work the moment it lands. You do not need to refresh, resend, or do anything at all; the working line simply keeps saying Max is still working on this, because Max is.

Once in a while you may see Waiting for a free worker under your message for a few seconds: every build runs on its own worker, and yours is waiting for one to come free.

The one case that is genuinely different is a message that never reached Max in the first place. The builder can tell, and says so plainly: it asks you to send that message again.

If a build stops part-way through

Builds are allowed to take as long as they need. A big first website can run for many minutes, and the builder simply waits: it never gives up on a build that is still working, however long that is.

Very occasionally a build stops before it is finished — the machine running it goes away, the connection to the AI model drops partway through its answer, or it runs past the limit on how long any one build may take. When that happens:

• You keep what it built. Every page the assistant had already finished is saved on your website. Nothing is undone; only pressing stop yourself undoes a build.
• You are not charged for it. A build that does not finish is not billed against your AI credits, the same as a build you stop.
• You carry on from there. Your message stays in the conversation with what happened underneath it and a Try again button, which sends the same message again — with the same photos and files. The assistant can see everything already on the website, so it picks up from what is there rather than starting over — you may find it only has a page or two left to add.
• If the connection was lost, we send it again for you. When the machine running your build went away because the platform itself was being updated, or the connection to the AI model dropped while it was answering, and you still have the builder open, it sends your message again on its own — you will see The connection to Max was lost with a note that your message is being sent again, and the build carry on from what was already saved. This usually happens within a minute, and it still happens if the builder reloads itself to pick up the update. It does this once per message, and a few times at most before asking you to take over. (A connection that drops while the model is still thinking, before it has written anything, is simply reconnected — you will not notice it at all.)

The message you see says which of the two happened: whether the pages it finished are saved, or whether it stopped before changing anything at all.

A build that stops because of a problem on our side says so instead — Max hit an unexpected problem on our side — with a reference you can quote to support. Our team is alerted to every one of those. Press Try again when you are ready; everything already built is still there.

Every build that doesn't finish stays in the chat

A build that fails, loses its connection, runs out of time, or that you stop never vanishes. Your message stays in the conversation where you sent it, and right below it a note says plainly what happened — for example The connection to Max was lost, Max ran out of time, Something went wrong, or You stopped this update — with a sentence on what it means for your website. It is still there after you reload the page or come back tomorrow, so you never have to wonder whether a request went through.

If your website's first build doesn't finish

When the very first build of a new website doesn't finish, there is no website to show yet, so the builder doesn't pretend there is. The preview says your website will appear there, and the chat ends with The site creation did not fully complete, followed by the reason. You have two ways on:

• Send a new prompt. Describe the website again — or something different — in the message box, and the assistant builds from there.
• Start over. Press Start over to put the website back exactly as it was the moment you created it: a clean starting page, a fresh conversation, and the four starting points (Start from a template, Clone an existing website, Describe it with AI, Start from an image, Blank canvas) to choose from again.

Starting over deletes nothing. The earlier attempt stays in Version history, marked Started over on the new starting page, and the earlier conversation stays in your chat history. Start over is only offered while the website has no pages of its own yet; once it does, use Version history to go back instead.

If the AI provider cannot connect

A big build sends the AI a lot to read, and it can take a while to start answering. The builder waits for it and keeps your view of the build alive meanwhile. If a Claude connection briefly fails, the assistant automatically retries opening it. If it still cannot connect, your message stays in the conversation with a Try again button. Any changes already saved to your website remain saved. Check the preview before trying again so you can describe what still needs doing.

When the assistant corrects itself

Occasionally the assistant reaches for a tool it doesn't have — the same way anyone might reach for the wrong drawer. It used to lose the whole reply when that happened. Now it is told what it actually has to work with and carries on, usually within the same message, so you see a finished reply rather than an error.

That correction is limited. If the assistant keeps missing, the message ends without the change rather than trying forever, and your next message starts fresh. If a request ends that way twice, describe the outcome you want rather than the mechanism — "the contact page needs a phone number under the address" instead of "edit the contact file" — which is what the assistant is best at anyway.

You are charged for the work the assistant does either way, including the step it spends correcting itself; a step is a small part of a typical message. Stopping a build still stops it immediately, even mid-correction, and a stopped build is not charged.

Asking for changes

Once a first draft exists, work in small, specific steps. The assistant uses the relevant parts of your website for each request, so you can refer to things by name:

• "Make the header photo taller and put the phone number on top of it."
• "The menu page needs a Desserts section — tiramisu, cannoli, and affogato, $9 each."
• "Use a warmer shade of green for the buttons."
• "Use Poppins for headings and Inter for body text."

Any family in the Google Fonts catalog is yours to ask for by name, on a new website or one you already have. The assistant adds it, and when your site is built the font files are copied onto the site itself, so your pages serve them from your own address and stay fast — no request to Google when a visitor arrives.

A few habits that make a real difference:

• One goal per message. Five unrelated requests in one paragraph get a muddier result than five short messages.
• Say which page. "On the contact page…" saves a guess.
• Give it your real words. Real hours, real prices, real menu items — the assistant writes better copy from facts than from placeholders. Text you paste in, a page you ask it to copy, or a section you ask it to move or restyle keeps every word — the assistant never shortens or summarizes your content unless you ask it to ("make this shorter").
• Say what's wrong, not just what to do. "This section feels cramped" is a useful instruction all by itself.

More detailed: how Max read your request

You write a request the way you'd say it; Max works from it the way a developer would write it down. Before Max starts on a request of more than a few words, it is spelled out as a short checklist, and Max is handed both — your message and the checklist. Open More detailed under your message to read it:

• The goal — your request in one plain sentence.
• When it's done — each thing you should be able to see once the change is made. A request that asks for several things is split into its parts, so "make the header see-through, then white with dark text when I scroll" becomes two results Max has to deliver, not one it can half-finish.
• Where — the page, the screen size and the moment it applies to: at phone width, after scrolling, while hovering.
• Stays as it is — what your request should leave alone.
• Assumed — where your message could be read two ways, the reading that was chosen.

Before replying, Max checks its work against each result on the list, and tells you plainly if one of them could not be done rather than reporting the whole request as finished.

Your own words always win: the checklist is a reading of your message, never a replacement for it. If More detailed shows the request was read differently from what you meant — the wrong page, an assumption you didn't intend — say so in your next message.

Good to know:

• It adds no wait. The checklist is written at the same moment Max works out how big the change is, not before it. If it isn't ready in time, Max simply starts without it.
• It uses none of your credits. Like the behind-the-scenes complexity check, it is paid for by the platform.
• Short requests don't get one. "Make this red" already says everything it means.
• You can switch it off for a website under Settings → AI Settings → Requests → Spell out my requests for Max. It is on by default.

Quick edits and larger updates

Small changes use a focused editing path. Select an element and ask for one specific change — for example, “Change this from blue to red” — to help Max find exactly what you mean. The preview refreshes as soon as the updated page is ready, even if the reply is still finishing.

On a large website Max updates the page you are looking at first, so the preview shows the change straight away. When a change touches something many pages share — the menu, the footer, a colour — Max then finishes the other pages before the update ends: the last step of the reply reads Finishing the other pages, and when Max is done, the whole website is.

On a very large website, where finishing every page takes longer than a few seconds, and after a change you make by hand, the other pages finish in the background instead, usually within a minute, and a quiet line at the top says Finishing the other pages in the background · You can keep working while they do. It is not something to wait for: your next message is sent straight away, a save goes through, and you can publish in the meantime.

A label below each update shows how it was handled:

• Quick edit — a small, targeted change.
• Website update — a broader change, such as updating a colour throughout your website.
• Complex update — a redesign or several related changes that need more work.

Plain language is enough: “Center the menu”, “Line the menu up with the logo”, or “Hide this on mobile” works without naming any code. Max knows which preview size you have open — desktop, tablet, or phone — so “the space on the left” is read at the width you're looking at. It looks for whatever is overriding the element before it writes, and a later request keeps earlier fixes and your own edits from Code mode. If a change didn't show up, just ask again — Max treats the repeat as a sign the first attempt was blocked, takes a closer look at what's overriding it, and tells you what it found.

If a change needs a broader approach than it first appeared to, Max continues in the same turn and shows a short note beneath the reply. The label and note stay in your conversation after a reload. Every update uses your website's selected model; the behind-the-scenes complexity check does not change that choice and is paid for by the platform.

These labels describe the work, not a guaranteed completion time. Larger updates can take longer. You can still stop an update using the Stop button beside its label.

Pointing at something on the page

When it's easier to show than to describe, use the cursor button beside the message box to select an element in the preview. Click the thing you want to change — a heading, a button, an image — and it gets attached to your next message, so you can simply write "make this smaller". Press Escape to clear the selection.

Pointing does more than save you a description. Along with the element, Max is told how it is actually laid out on your screen at that moment — its size, and how it and the boxes around it are arranged. So when something sits in the wrong place ("the spinner shows above the label instead of beside it"), select it first: Max starts from what your browser measured instead of guessing from the code.

Attaching photos and documents

The paperclip button attaches images and documents to a message. You can also drag files in from your computer: while the chat is open, drag them anywhere over the builder — the preview, the panels, the message box — and the whole window shows Add anything. Let go and they are attached to your message, exactly as if you had picked them with the paperclip. A picture dragged from another tab or web page attaches the same way. Dropping while Max is still working is fine too: the files wait for your next message.

Use it to hand over your logo, photos of your work, a menu PDF, or a screenshot of a layout you like. Images you attach are optimised into your website and then appear in the Media panel.

You can attach up to five files to one message. Images can be JPG, PNG, or WebP up to 50 MB each — full-page design mockups and tall screenshots included, up to 50,000 pixels on the longest side and 100 megapixels in all; documents can be PDF, DOCX, TXT, Markdown, or CSV up to 100 MB each. On an organization plan, your plan's largest file is the document limit instead (20 MB on Free, up to 250 MB on Artisan), and it caps images too when it is below 50 MB. A file over its limit is refused as soon as you add it, with a message naming the file and the limit. When your plan is the limit, owners and Admins also get an Upgrade plan button above the box for the plan that holds that file (Contact us when none does), and everyone else sees who to ask. If a message cannot be sent the chat says why and keeps your words and files so you can try again.

Adding or replacing a PDF

Attach the document and say where it goes — "update the SS Book link in the menu with this new file", or "add this menu to the footer". The assistant does not read what the document says; it keeps the file in your website's Library (listed in Media under Library, outside the website's versions, so it is never copied into new versions or exports) and points your links at it. When the new file replaces an old one, every link to the old file is updated and the old file stays in Media, so you can delete it there yourself once you are happy. Text, Markdown and CSV files are linked the same way but are not listed in Media.

Building from a design

If you already have a design — a full-page mockup from a designer, a screenshot of a website you like, a wireframe, even a photo of a sketch — attach it and say what you want: "build my homepage to look like this". The assistant treats the picture as the plan, not as a picture: it reads the whole page top to bottom, and builds real sections with live text you can edit, in the design's fonts, colours, layout, and spacing. A tall mockup is read in full-size pieces, so every word and measurement in it counts.

Three things to know:

• The picture never ends up on your page. Every heading, paragraph, and button in the design becomes real text, so search engines can read it, it fits a phone, and you can change any word later. If the assistant could not make out a word, it says which one.
• Photographs have to come from you. A photo in a mockup cannot be lifted out of it. The assistant uses a matching image from your Media panel when there is one, and otherwise holds the exact space with a placeholder in the same tone, then lists the photos to upload, section by section. Attach them in a later message and ask it to put them in place.
• It checks its own work. When the design is the first thing built on a new website, the assistant takes a screenshot of the result, compares it with your design, and fixes the differences it finds in one more turn — you will see that follow-up appear in the chat on its own. It is a normal turn, so it counts toward your website's AI usage.

For a website you already have live, Start from an existing website does better still: it reads the real pages rather than a picture of them.

Conversations

Each website keeps its conversations. The clock icon at the top of the panel lists them so you can return to an earlier one, and the plus icon starts a fresh conversation.

A new conversation is worth starting when you move to a genuinely different job — you've finished the homepage and now you're building a booking page. Staying in one conversation is better while you're refining the same thing, since the back-and-forth is context the assistant can use.

Who asked, and how long it took

When several people on your team work on the same website, each message in the chat shows a small picture of the person who sent it — the same avatar they have everywhere else in Max AI Builder. Hover over it (or move to it with the Tab key) to see their name and the date and time they sent the message, in your website's time zone. The very first message of a website's first conversation tells you who asked for the original build.

Someone who has since left your organization, or deleted their account, shows a plain placeholder that reads Former member — their name and picture are no longer shown. Messages sent before this feature arrived have no picture at all.

Under each finished reply, beside the time stamp, is how long the whole request took, from pressing send to the reply finishing — 48s, 2m 14s, 1h 03m. Updates still running, and ones that failed or were stopped, don't show one.

Beside it, on an organization that uses shared AI credits, is what the reply cost — Used 12 credits. It is the whole cost of that one request, taken from your organization's credits the moment the reply finished, and it matches the entry on the organization's Usage page. A reply that failed or that you stopped shows no cost, because it was not charged. See what things cost for what to expect before you ask.

You can't change who a message is credited to; everyone on the team, Viewers included, sees the same thing.

Then what?

Everything the assistant builds lands in your draft, and your dev site shows it right away. When you're happy with it, press Publish to put that version on your own domain. See Touring the builder for what the publish button's labels mean.