Working with the AI

What the assistant can do — and what it cannot

The assistant always works on exactly one website: the one selected at the top. It reads its pages, proposes changes and only carries them out once you approve them.

You will find it in two places. In the editor it sits on the right under the “AI” tab, where it knows the open page and the selected block. In the “Brand & design” area the same column sits along the right-hand edge, there without reference to a single page — it helps with questions about the website as a whole.

What the assistant does not do

• Publishing. Your changes only go live once you click “Publish” in the editor.
• Changing the header and footer. Those areas are not part of the catalogue the assistant may touch.
• Setting colours, fonts, logo and favicon. That belongs in the “Brand & design” area.
• Connecting domains, inviting team members, changing plan or invoices.
• Uploading or deleting media. The AI generates images at the image field of a block and during a build via “Build” — not in the chat.
• Reaching content that is not yours. The assistant only ever sees the selected website of your account.

In headings the AI highlights at most one keyword using ==word== — on the website that appears in your brand gradient. Whole sentences are never marked. In body text, **bold**, *italic* and links are available as well.

Opening the AI column

In the editor, the inspector on the right has two tabs: “Properties” and “AI”.

1. 1

   Click the “AI” tab at the top right.

   The column is headed “AI assistant”.

2. 2

   On narrow screens, tap “Properties” in the bottom bar.

   The right-hand panel slides open; choose the “AI” tab there.

3. 3

   Alternatively press Ctrl+K (Cmd+K on a Mac) to open the command search.

   Then choose “Open AI assistant”.

Next to the title sits a small status. In three words it tells you whether the AI can work.

Below that you see your usage: “AI images”, “Text” and “Credits”. At the top right, a button opens “New chat” — it clears the history, any open proposals and notices and starts a new conversation without changing your website. As long as there is nothing to clear, it stays greyed out.

The box above the history explains how to start: “Describe what you want to change — e.g. “make the heading shorter” or “add a pricing table”. For a whole page, enter a short brief and choose Plan or Build.” The input field itself is labelled “Message to the AI”; while it is empty it reads “Message to the AI … (Enter sends, Shift+Enter = line break)”. When a page is open, the note “Page active” appears next to the buttons.

Phrasing an instruction

The more precise the instruction, the less you have to correct afterwards. The assistant knows which page is open and which block is selected — everything else you tell it.

1. 1

   First select the block you are talking about.

   The assistant receives its type as context, so “this heading” refers exactly to that block.

2. 2

   Name the target and the result in one sentence.

   That is: which block or page, which field, and what should come out of it.

3. 3

   Send with Enter; Shift+Enter adds a line break.

   The answer appears word by word in the history.

If you name a different page or block, the assistant looks for it first. If the name matches several results, it asks back and changes nothing — so it cannot hit the wrong thing.

So that answers stay tied to your content rather than being made up, the editor adds matching excerpts from this very website to the request — up to five excerpts totalling around 6,000 characters. That is why the assistant can also answer questions about your own copy.

Instructions that work

• “Shorten the heading in the selected block to six words at most.”
• “On the Services page, add three cards below the opener: maintenance, repair, inspection.”
• “Write a meta description of at most 160 characters for the home page.”
• “Make the page nicer” is less helpful — no concrete step can be derived from it.

Instruction, “Plan” or “Build”

Below the input field are two extra buttons: “Plan” and “Build”. Together with sending a normal message that gives you three ways of working — they differ in how much the AI touches at once.

Reviewing and applying proposals

When the AI changes something in your blocks, it puts the result as a card below the chat history. Each card names the change in one line, with “1 block affected” or “… blocks affected” underneath.

1. 1

   Read the description on the card.

   It states in a single line what the change is about.

2. 2

   Choose “Apply” to pull the change into the page.

   If you do not like it, choose “Dismiss”.

3. 3

   Check the confirmation in the history.

   After applying, a short line appears, for example “Change applied (1 block).”

What happens when you apply

If the proposal brings finished blocks with it, the editor appends them to the open page — existing blocks stay where they are. The first new block is selected right away, and the editor saves the previous state beforehand so that “Undo” brings it back.

If the assistant carries the change out itself — creating a page, say, or setting the meta information — the server writes a restore point first and notes the change in the log of the website. The editor then reloads the page. You will find both under “Version history” and “Activity”.

Approving several proposals at once

As soon as at least two applicable proposals are waiting, a bar appears above them with the count (“… proposals open”) and the buttons “Apply all” and “Dismiss all”. “Apply all” works through the cards one after another and reports at the end how many changes were applied. If a single one fails, exactly that card stays in place so you can try again; the page itself is reloaded only once.

While the bulk approval runs, the individual cards are locked. “Dismiss all” clears the list completely without changing anything on the website.

Cancelling a run

While the AI is answering or building, the send button turns into a red “Stop” button.

1. 1

   Choose “Stop” to cancel the running job.

   The button sits where “Send” usually is.

2. 2

   The chat stops cleanly.

   No error message appears and whatever was already written stays in place.

3. 3

   Rephrase your request and send it again.

   It often helps to select the block in question beforehand.

“New chat” also cancels a running job and starts over on top of that. Changes you have already applied are unaffected — you revert those with “Undo” or via “Version history”.

Limits of a conversation

A conversation with the assistant is deliberately limited — in length, in working steps and in changes. Those limits guard against endless loops and against an old proposal landing on a state that has long since moved on.

Once thirty changes have been applied in one conversation, the assistant asks you to start a new chat. That is not an error but a clean break: a fresh conversation starts without the old history and without open cards.

Switching website or page also starts a new conversation. Any open proposal cards disappear with it — they belonged to the previous page and would no longer fit.

Rewriting copy: the seven commands

In the “Properties” tab, every text field and every multi-line text field has a slim bar underneath that starts with “AI” (tooltip: “AI text commands”).

1. 1

   Select the block whose text you want to change.

   The inspector then shows its fields.

2. 2

   Find the right field in the “Properties” tab.

   It has to contain text already, otherwise the commands stay greyed out.

3. 3

   Choose one of the seven commands: “Shorter”, “Longer”, “Rephrase”, “More professional”, “Simpler”, “Bullet points” or “Fix”.

   Each command works on that one field only.

4. 4

   Wait a moment — the new text then replaces the content of the field.

   While it works, a small spinner turns inside the command you chose; the preview picks the change up straight away.

What is preserved

• The language of the text — more precisely, the language version you are currently editing.
• Existing highlights using ==word== as well as **bold**, *italic* and links.
• The meaning: no facts, figures or promises are invented, and words such as “guaranteed”, “error-free” or “100 %” are avoided.
• German text keeps its real umlauts; what comes back is only the revised text, with no preamble and no surrounding quotation marks.
• The tone you set for your brand applies here as well — see the next section.

One command processes up to 6,000 characters. If you do not like the result, revert it with “Undo” or choose a different command. Without a selected website the buttons stay greyed out; the tooltip then reads “Only available with a signed-in website”. Inside list entries — the cards of a grid, for instance — the bar is not available; you edit those texts by hand or through the chat.

The tone of the generated copy

How the AI sounds is something you set once: in the “Brand & design” area under “Brand voice”. The choice affects everything produced afterwards — building whole pages, answers in the chat and the text commands.

Below the tiles sits the field “Your own voice (optional)”. There you describe in your own words what your address is like — that you use a formal register, use jargon sparingly and emphasise your region. Up to 400 characters of it feed into every text instruction. “Save voice” stores both.

Whatever the tone, the rules stay in force: no absolute promises, no invented figures, no emojis. A playful tone therefore does not turn a statement into advertising — it only changes how it is phrased.

Generating images

Every image field offers four routes to an image: “URL”, “Generate with AI”, “Upload” and “Media library”.

1. 1

   At the image field, choose the “Generate with AI” tab.

   Without being signed in it reads “AI image generation requires signing in.” instead.

2. 2

   Under “Describe the image”, say what should be visible.

   For example: bright office with a team at a whiteboard.

3. 3

   Under “Purpose”, pick between “Hero”, “Card”, “Gallery”, “Portrait” and “Banner”.

   The choice determines format and composition. Without a choice of your own, “Hero” applies to wide areas and “Card” everywhere else.

4. 4

   Choose “Generate”.

   While it works, the button reads “Generating…”.

5. 5

   The finished image is placed into the block straight away.

   The preview shows it immediately.

What never ends up in the image

• Text, letters, numbers, watermarks, logos or signage — type inside an image could be neither translated nor read reliably.
• Recognisable faces of identifiable people. That protects personal rights and prevents invented customer photos.
• Collages, frames and mocked-up user interfaces.
• The look stays photorealistic, in natural light and with a restrained palette.

Your brand colours and fonts are added to the image description automatically. Every generated image also lands in your media library and can be reused via the “Media library” tab; the “Source” filter separates “All”, “AI images” and “Uploads” there.

Every image needs an alternative text — without one it is not accessible to people using a screen reader. If it is missing, the editor uses your image description. You enter it yourself via “Edit alt text & title”, and “AI: suggest alt text & title” proposes one; the media library chapter describes both in detail.

Importing a brand and getting a logo suggestion

In the “Brand & design” area, two more AI features help you get started: analysing an existing website and suggesting a logo. Both only propose — nothing is adopted until you click.

Importing a brand from a website

In the “Import brand from a website” section, enter the address of your existing website and choose “Analyse”. The server fetches the page itself — you do not have to copy anything — and reads its title, meta description, site name, theme colour and a text sample. The “Brand suggestion” is derived from that.

None of it is applied automatically: “Apply tone” sets the tone and brand voice, “Apply colours” the two colours. The analysis only reads publicly reachable addresses — internal addresses and pages behind a password are refused.

Getting a logo suggested

In the “Logo” section, switch to “AI logo”. The result is a vector graphic that stays sharp at any size.

“Generate” starts it, “Regenerate” fetches another version. The suggestion first appears in the preview only — light and dark. Only “Apply as website logo” puts it to use, and “Remove logo” takes it out again. Every version is checked before it is shown: only drawing elements are allowed, no program code, no external references and no opaque background.

Having posts written

In the “Blog” area, the AI suggests topics and writes complete posts from them. The procedure is described in the chapter on content, media and blog; here is what a generated post always brings with it.

The topic suggestions know the titles of your existing posts and do not repeat them. Generated posts land in your list as drafts — scheduled or published only if you set it that way.

The check after writing

Every finished post then goes through a check for inadmissible advertising claims: “guaranteed”, “100 %”, “error-free”, “no. 1”, “market leader”, “world-leading”, “unbeatable” and “risk-free” — including their German equivalents. If the check finds something, there is exactly one repair run. If anything is still left after that, the post is discarded rather than published.

Invented statistics, studies, figures and awards are ruled out as well. Even so, read the post through before publishing it: only you know whether the statements about your business hold true.

Quotas and cost

AI actions run on two separate budgets: the conversation in the chat on a chat balance of its own, everything else on the monthly quotas for AI text and AI images. You see both above the input field and in full in the “Account” area.

Counting happens per action, not by length of text; only the conversation in the chat is billed by size. Failed runs are not counted, and if a build breaks off, credits already charged are refunded. The monthly quotas start over with every calendar month; purchased credits stay in your balance.

When the quota is reached

Once a monthly quota is used up, instead of an error you get a note headed “Quota reached — upgrade now”. Below it, the affected area is named: “Your quota for … is used up for this period.” with “AI images”, “AI text” or “AI credits” in the gap — or, in general terms, “Your AI quota is used up for this period.” The history adds the short line “Quota reached.” If the chat balance is empty instead, the chat reports that the chat token quota for this month is used up and that a plan change or a chat token pack will help.

1. 1

   Use “Upgrade now” to reach “Choose plan” in the “Account” area.

   There you can see which quotas each plan includes.

2. 2

   Use “Later” to dismiss the note.

   You then carry on working without the AI.

3. 3

   Additional credit is available in “Account” under “AI credit packs”.

   That tops up the current month without changing your plan.

What reaches the model

For the AI to work, it has to see your instruction and the necessary context. What exactly that is, feature by feature, is listed here.

In doing so, the assistant only ever sees content from the selected website of your account. Other websites, other accounts and your contact requests are not part of it.

The retention periods named above are the defaults. What else XICflow processes is set out in the privacy policy; contact requests and how long they are kept are covered in the chapter on requests and team.

If something does not work

These are the messages you are most likely to see, together with what helps next.

• “The AI is busy right now. Please try again in a moment.”

  Too many requests came together in a short time — thirty AI requests per minute are possible per account. Wait a moment, then send the request again.

• “The AI is currently unreachable.”

  The service did not accept the request. Try again in a few minutes; nothing changes in the editor itself.

• “Connection to the AI backend failed. Please try again later.”

  The connection broke down on the way — often a network problem. Check your connection and send again.

• “Your session has expired. Please sign in again.”

  Sign in again and reopen the editor. Anything typed into the chat but not sent is lost.

• “The request was incomplete.”

  Something was missing from the request, such as the website. Select your website at the top left and send again.

• “That did not quite work. Please rephrase the request differently or more precisely.”

  Be more specific: name the block, the field and the result you want — and select the block in the editor first.

• “An error occurred.”

  The run ended prematurely. Send the request again; if that does not help, start over with “New chat”.

• “Apply failed. Please try again.”

  The card stays in place. Choose “Apply” again. If it keeps failing, the proposal is probably older than ten minutes — dismiss it and rephrase the instruction.

• “No applicable block in the proposal.”

  The card contains nothing that could be pulled into the page, so “Apply” stays greyed out. Dismiss it and describe more precisely which block should change and how.

• The assistant asks for a new chat

  Up to thirty changes can be applied within one conversation. After that, choose “New chat” at the top right and carry on there.

• “The AI build failed. Please try again.”

  Close the message and start “Build” again with a shorter, clearer brief. Credits already charged for a failed build are refunded.

• “No result received.” or “AI text command failed.”

  The text command came back without a usable result. Choose the command again or pick a different one; the text in the field stays untouched.

• “Text refinement is currently unavailable.”

  The service did not accept the command. Wait a moment; your quota is not charged for it.

• The text commands are greyed out

  The field is empty, or no website is selected — the tooltip then reads “Only available with a signed-in website”. Enter some text first, or select your website at the top left.

• “The AI returned no image.” or “Image generation failed.”

  Adjust the description, pick a different “Purpose” if needed and choose “Generate” again.

• “No usable SVG received. Please try again.”

  The logo suggestion came back in a form that did not pass the check. Choose “Regenerate”, ideally with a shorter hint.

• “Please enter a valid public http(s) address.”

  The brand analysis only reads publicly reachable addresses. Enter the address in full, including “https://”, and without password protection.

• The usage display looks out of date

  It refreshes after every AI action and when you switch back to the browser tab. The complete picture is in the “Account” area.

If none of that helps, start a fresh conversation with “New chat” and break the instruction into smaller steps. If a change seems to be missing, check “Version history” to see whether it arrived after all — and revert it there if needed.

Frequently asked questions