Usage 

What the chat tells you besides the answer 

• Nothing was saved in this step. An answer that claims a finished change ("done", "created", "saved") carries this note when the run behind it wrote no record. Check the page before relying on the answer, and ask the assistant again.
• The AI provider is not set up. Editors see this sentence when the provider cannot be used, for example because its API key is missing. Administrators see the technical message instead, with a button to the LLM providers or tasks module.
• This step is still waiting for a decision. A message such as "weiter" or "habe alles freigegeben" while an approval card is on screen does not start a new run and does not approve anything: decide on the card. Without a card, the step is decided in AI Tasks; if you cannot open AI Tasks, ask someone who may approve there. If the step was already decided in AI Tasks, the chat says so and names the records the run wrote.
• Rule-set check of the proposed change. When the chat's AI configuration has the rule-set check of nr-llm switched on in report mode, the text of a change the assistant proposes is checked against the configured rule set. A rule the text may break is listed below this line, with the passage it is about; so is a check that gave no answer, in which case read the text against the rules yourself. A change shorter than 30 words is not checked, so a missing hint does not mean that a text was checked and passed. The change is still proposed, and you decide on it as usual. In block mode a change that breaks a rule is not proposed at all: the chat ends with an error that names the rule.

Answering a question in the chat 

A step can stop and ask you something before it goes on, for example which point of a page it should start with. A proposed change is never asked this way: it waits on an approval card. The question appears as the assistant's message, and the possible answers appear as buttons directly above the input field. Press one to answer; the step continues with your answer, and the answer stays in the conversation as your message.

When the question allows an answer in your own words, the input field says Write your own answer and what you send is the answer. Otherwise a message you send ends the question and starts a new request. A question with several fields is shown as a small form with a Send answer button. If the chat cannot show a question, it says so.

When the question changed after it was shown, or your answer could not be taken, the question comes back with the reason above the buttons; answer again.

Approving a step in the chat 

A step that writes data waits for your decision on a card in the chat. The card is headed with what the step does, for example Create page draft or Delete page, followed by what would change: where, the current state, the new state and the consequences. The approve button carries the same name, so it says what pressing it does. Cancel leaves everything as it is. When a guided process proposes a change, two buttons take the place of Cancel and tell the assistant why the change is not taken: Another variant asks for a different proposal, Keep leaves the current content as it is and closes the open point, Skip moves on without this change and keeps it open; your choice stays in the conversation as your message. In a guided process, a message of your own while a proposal or a question waits cancels that step. The name is the first line of the step's preview when nr-llm marks that line as naming the change; otherwise it is the tool's editor action label, and without one the step is headed Planned step and approved with Carry out step. A turn with several steps is approved as a whole with Carry out all steps.

Administrators and users of AI > AI Tasks decide every approval. An editor whose backend group holds Approve own changes in the chat (under Custom module options) decides the approvals of their own conversations, and may approve a change only when the card shows its preview — the preview is built with the editor's own rights. Where the AI configuration requires a second approver, the card offers only the denial and a colleague approves the step in AI Tasks.

A message you send while you work in a TYPO3 workspace is carried out in that workspace, also when you switch to another one before the chat has finished. A content element the chat changes there is stored as a draft of the workspace: it appears in the Workspaces module, and the live site stays as it was until the workspace is published. The approval card names the workspace, for example Saved as a draft in workspace “Entwurf”; live stays unchanged until the workspace is published, and the target language where the step translates something. If you may not open that workspace, the card says that it is a draft in a workspace whose title is not visible to you. The chat reads the page as your workspace shows it. Changes nr-llm cannot make in a workspace are refused there with a message instead of being made live. When you may no longer work in the workspace the message was sent from, the turn fails and nothing is written.

A step that a rule of the AI configuration excludes, for example because the configuration requires a draft workspace and you work in the live workspace, is marked This step will not be carried out on the card, and the rule's explanation stands below it. Approving the card does not carry that step out. When every step of the card is excluded, the card offers no approval, only the way to cancel. A configuration that requires a draft workspace says so beside the card while you work in the live workspace; switch to a workspace in the backend and send the request again.

The tool's internal name, its arguments and the record identifiers are under Show technical details, which is closed until you open it.

The assistant also knows which tools exist but are not available to you, and why — switched off, for administrators only, or not part of this chat's configuration — so it can tell you to ask an administrator instead of claiming that nothing can do what you asked.

What an assignment changed 

Below the conversation, Änderungen dieses Auftrags (Changes of this assignment) lists the records the assistant's steps changed. It appears once a step has written something, in every conversation, not only in guided processes, and it covers all steps of all runs of the conversation, also runs that were continued after an approval. Open it with Enter or Space; it loads the list when you open it.

Each row names the record, with its table and uid. The title is a link to the edit form where you may edit the table, and Änderungen ansehen opens the record history. The badges state, each on its own:

• what happened: created, updated, moved, deleted or restored;
• the workspace the record is stored in: a draft in a named workspace, or the live workspace;
• the language, where the record has one;
• whether the record is hidden or not hidden. Not hidden says nothing about whether the record is public: a draft in a workspace is not live, and a start time, an end time or a user group can still keep it from visitors. These restrictions are separate badges;
• whether the record is the target of the step or was changed along with it (the content of a copied page, the overlay of a translation);
• that the step failed after this record was written, in which case check the record before you ask again.

Under the list a note states what the list covers: the changes TYPO3 reported for the steps of the assignment. A change made another way does not appear, so the list is not proof that nothing else changed. Further notes appear when the capture of a step was not complete (Weitere Auswirkungen nicht vollständig erfasst), when more records changed than the list holds, when runs are older than the capture (only the record each step named is known) and when runs could not be read.

You see a record only if you may read it: the table, the page and its permissions and mounts, the language and, for a draft, the workspace. Records you may not read, and records that no longer exist, are not listed and not named, only counted. The link to the Workspaces module appears for records in the workspace you work in, because that module shows your current workspace.

The list is read-only. It belongs to the conversation: editing a message removes the steps after it from the list as well, and a conversation older than the list shows the record each of its steps named.

Editing a message 

Every message you wrote as text has an Edit control below it while no answer is being generated. Change the text and choose Save and run again: the message is replaced, everything after it is removed, and the assistant answers the new wording. An attachment of the edited message stays attached. Escape leaves the editor without changes, Ctrl+Enter saves.

What the assistant knows about you 

• The page you are on. When you send a message while a page module (Page, List, ...) shows a page, the assistant is told that page's uid and title and which module is open. "Summarise this page" therefore works without naming a uid. A page you may not see is never passed on.
• Your backend language. The assistant answers in the language your backend is set to (User settings > Language), even when you write in another one. Ask for a different language in the message ("answer in English") and it uses that instead.

What the chat is about: the chips above the input 

When the chat is about one page, a row of chips above the input says which:

• The page chip (grey, with a page icon) shows the title of the page. Beside the page module (Layout) it is the page selected in the page tree, and it changes as soon as you select another one. In a guided process it is the process's page, which stays when the page tree moves on. Only a page you may see is named.
• The element chip (white, with a black circle and a yellow dot) appears in a guided process when the current point is about a content element, and changes with the point; a point about no element shows the page chip only. Beside the page module, pointing at it or focusing it with the keyboard marks the element in the page layout, and a click scrolls it into view. In the full-page chat module, in the popped-out panel and while the page module shows another page, the chip is only a label.
• A page property such as the meta description gets the same chip when the current point is about it, named by the field's label as in the page properties. The page module has no element for it, so this chip is always only a label, and an element marked in an earlier point is no longer marked.
• The (i) at the end of the row says what the assistant may change, for example "Bereich: ganze Seite „AI Chat Agent“. Geändert wird nur nach deiner Bestätigung." or, for a page property, "Der Assistant ändert nur die Seiteneigenschaft „…“ der Seite „AI Chat Agent“." It opens on pointing at it and on keyboard focus; Escape closes it.

A long title is shortened in the chip; the full title is in the chip's tooltip and in the text of the (i). There is no row for a skill that works on no page (Whether a skill works on a page), for a chat without a page, or while no page is selected. A conversation about a missing alt text shows its image and the sentence on what may change in the same row.

What the assistant is doing 

The Activity button (a pulse line in the floating panel, Activity in the module) lists the steps of the current turn while it runs: every call to the model and every tool the assistant uses, with its duration, a failed tool call marked as such, and — when a step needs your approval — which tool is waiting. After you decide, your decision and the steps that follow it are added to the same list. The list starts over with your next message. It shows the names of the tools, not their arguments or results; those are in the run's timeline under AI > AI Tasks.

In the expanded panel the list sits above the conversation; maximized, and in the module, it is a sidebar on the right.

Starting with a page and a skill 

A link into the chat module can start a conversation about one page: &pageUid=<page>, optionally &languageUid=<language> and &skill=<skill>. The chat checks that you may open the page and edit the language, opens the new conversation and replaces the parameters with the conversation's own, so reloading opens it again.

Typing / at the start of the input lists the skills available in the chat. Pick one with the arrow keys and Enter, or with the mouse; nothing is sent. The conversation's skill is shown above the input, and the × beside it removes it; in a guided process about a page this ends the process (see below). Every later answer in the conversation uses it.

When the chat's AI configuration needs a second person to approve every change, no skill can be started, and the chat says so: a guided process is decided on the chat's own approval card, where you cannot release your own change.

Whether a skill works on a page 

A skill states in its frontmatter whether it works on a page (the page key of nr-llm, ADR-025: The Skill Says Whether It Needs a Page):

page: required

The conversation is about one page. Without one, the chat asks for it: for a guided process, choose the page in the page tree and confirm it with Diese Seite prüfen; for any other skill, choose the page and then write what to do -- your message carries the page. Once set, the page stays, whatever page the page module shows later; Zur neuen Seite wechseln opens a new conversation on the other page.

page: optional

The chat passes along the page the page module shows with each message, when there is one; it does not ask for one.

page: none

The conversation is about no page. The chat does not ask for one, does not pass one along and does not keep one, even when the conversation is started with a page.

A skill without the key behaves as before: a guided process as required, any other skill as optional. The Frequent tasks widget follows the same declaration (dashboardQuickTasks).

Instructions for a conversation 

The Instructions button (sliders icon in the floating panel) opens a field for instructions that apply to this conversation only -- for example "answer in bullet points" or "you are reviewing the press section". They are added to the instructions an administrator configured for the chat, which stay in force: where the two contradict each other, the configured ones win. Your instructions take effect with the next message. The assistant's identity, its tools and permissions, and the language rules stay the same. The button is highlighted while a conversation has instructions; empty the field and save to remove them. Instructions cannot be changed while an answer is being generated.

Guided processes 

A skill can walk you through a page one point at a time. While it does:

• The header shows where it stands -- for example "Über uns · Deutsch · Punkt 2 von 5": the page and its language from the conversation, then the point. When space is short the page name is shortened first; the point is always shown in full. Before the first point the header says "Analyse läuft". It changes when the assistant stops to ask you or to propose a change, not while it works.
• The element is highlighted. With the floating panel open next to the page module, the content element the current point is about gets a frame and the label "Betrifft den aktuellen Vorschlag", and is scrolled into view; a screen reader hears which element is marked. The full-page chat module and the popped-out panel have no page module beside them and only show the progress. A point about a page property, such as the meta description, marks nothing in the page module; the chip above the input names it.
• Each proposal shows its outcome. A proposal is shown as a block with what is affected and, where nr-llm provides it, the current and the proposed value with its length. After you answer "Übernehmen", the chat says itself whether the change was saved, saved but worth checking, or not saved. At the end it summarises what was applied, what to check and what you skipped.
• The end is shown. When the process is finished, the header says so and the highlight goes away.
• At the end you choose what follows. Fertig ends the process; where the dashboard is installed, the chat links to it. Where the process can be started anew (see below), Andere Seite wählen starts it in a new conversation that asks for a page.
• Another page in the page module is noticed. When the page module beside the floating panel shows another page than the process, the chat says "Andere Seite ausgewählt". Bei „<Seite>“ bleiben keeps the process and does not ask again for that page; where the process can be started anew, Zur neuen Seite wechseln ends it and starts it on the new page.
• You can end it at any point. The × beside the progress ends the process, and so does removing the conversation's skill. A proposal still waiting for your answer is withdrawn and changes nothing; changes already applied stay, and so do the outcomes shown in the chat and the open points. The highlight goes away and the input gets the focus. While the assistant works or a decision is being carried out, the process cannot be ended.
• A decision that cannot be taken right now comes back. When nr-llm cannot check the process just then, the proposal stays and the chat asks you to try again. When the AI configuration requires a second person to approve every change, nr-llm ends the process at the first decision: the chat says so, the proposal is not applied, and the conversation can go on without the process.
• Skipped proposals stay open. A change you answer with Skip is kept as an open point of that page: the record and field it was about, never its text. The next time the process runs on the page, it is offered again. It stays open until a change to the same record and field is applied in full, in any conversation; a change that was only partly applied leaves it open. A proposal you leave behind with a message of your own is not kept.
• Kept content closes the open point. A change you answer with Keep is not offered again: the open points on that record and field are closed, and the tour summary lists the proposal under Kept.

A process started for a page -- from the dashboard's "Eine Seite verbessern", or by Zur neuen Seite wechseln -- begins at once: the chat sends Diese Seite prüfen for you, and you do not need to type anything. The process stays on that page; to work on another one, start it there. A process started without a page asks for one: choose the page in the page tree, then confirm it with Diese Seite prüfen in the floating panel. This page choice, Andere Seite wählen and Zur neuen Seite wechseln are offered only where nr-llm can start a run with an invocation of the process; until it can, they stay hidden.

None of this changes a record. An administrator can switch the tools behind it off as the tool group nr_mcp_agent.

Starting a new conversation 

Click the New conversation button to start a fresh chat. The previous conversation remains in the sidebar for later access.

Resuming a conversation 

Click any conversation in the sidebar to resume it. The full message history is loaded, and you can continue where you left off.

Pinning conversations 

Pin important conversations to prevent them from being auto-archived. Pinned conversations appear at the top of the lists.

Archiving conversations 

Archive conversations you no longer need actively. Archived conversations are hidden from the default sidebar view but can still be accessed.

Conversations are also auto-archived after a configurable period of inactivity (default: 30 days).

Exporting a conversation 

The Export button (a downward arrow in the floating panel) downloads the open conversation as a Markdown file named after its title and the date. The file holds what the chat shows as the conversation: your messages and the assistant's answers with their time, and the name of any attached file. Tool results and system notices are left out. The file is built in the browser from the conversation already on screen; nothing is sent to the server for it.

Chat as a column 

The Column mode button in the panel's header docks the chat as a column in the backend's content area: in modules with a page tree it sits between the tree and the module, in modules without one (the Dashboard, for example) it is the leftmost column. The column stays when you switch modules, and the module beside it gets the remaining width. The same button in the column's header turns it back into the floating panel.

• Width -- Drag the edge between the column and the module, or focus it with Tab and use Left and Right to step, Home and End for the narrowest and the widest width. The column is at least 320 px and at most half the content area wide.
• Selected page -- The chat follows the page selected in the page tree, as soon as you click it. In a module that is not about a page the chat sends no page.
• Minimize -- The column becomes a small bar in the bottom right corner with the chat's title and state, for example Waiting for your approval when a change waits for you. Its first button opens the column again.

Column mode and width are stored in your backend user settings, so they follow you to other browsers. Where the backend has no content area for the column (for example without access to any module), the chat stays a floating panel. While the content area is narrower than 640 px (twice the column's minimum), for example in a small window, the chat floats as well and docks again once there is room; your choice stays stored, and the Column mode button is not offered meanwhile. While the chat is sending a message or a decision, the Column mode and Open in a separate window buttons are disabled. The module menu is not collapsed automatically.

AI assistant dashboard 

Editors can put together a dashboard for working with the assistant from the widgets in the AI Assistant group (Add widget > AI Assistant), or start from the AI Assistant preset when creating a dashboard, which holds them in this order: AI Assistant, Edit a page, Create content, Recent chats, Recommended for your website, Improve a page, Check website. Frequent tasks is in the group, not in the preset.

Every widget opens with one sentence and offers one kind of action: a button, a grid of buttons, number tiles or a list whose rows are links. An action that concerns a page opens the page module with the chat as a column beside it (floating where the window is too narrow, see Chat as a column; a chat popped out into its own window stays there). Where the chat panel is not available, the same action opens the chat module.

AI Assistant

A question of your own, without choosing a tour or a page first: type it under What would you like to do? and press Enter or the round send button; Shift+Enter starts a new line. The chat opens as a column beside the dashboard (floating where the window is too narrow for a column, see Chat as a column) on a new conversation, and your question is its first message, as you typed it; the assistant answers from there. The column is for this opening only: your chosen chat layout is not changed, and the chat returns to it once you close it. A chat you popped out into its own window stays there and gets the question there. The send button can be pressed only while there is a question. Where the chat cannot take the question -- while it is still sending another message, or when the question is longer than the configured maximum -- the question stays in the widget and the widget says why. A question about a page while no page is selected makes the assistant ask which page you mean.

Edit a page

Choose a page to edit opens the page module with a new chat beside it. Choose the page in the page tree; the chat follows it, and whatever you then ask is about that page.

Create content

One button per kind of content the assistant can create for you: New page, and News where EXT:news is installed. A kind is offered only where you may create its records and the tool that creates them is enabled for your chat, and that the tool would not refuse for you (the default language, and for news its table settings and nr-llm's deny-list). A button opens a new chat with the start of the request in its input (I would like to create a new page, namely …); nothing is sent until you complete and send it, as the widget says below the buttons. Without any kind, the widget says that your administrator can enable the tools.

Improve a page

One row per guided tour (by default Optimise SEO and Improve content). A row opens the page module and starts the tour in the chat beside it; the tour asks for the page, which you choose in the page tree and confirm with Check this page.

Frequent tasks

Shortcuts into the chat, set by dashboardQuickTasks (none by default), one row each. A task that works on one page opens the page module and asks for the page there; any other task opens the chat module.

Recommended for your website

Number tiles for what the extension can count exactly, each with one action:

• Images without alt text you may fix, on how many pages and in how many records; View opens the list (see Missing alt texts).
• Pages without a meta description: page versions you may edit whose meta description is empty, in every language you may edit and in your workspace; hidden pages, folders, links, shortcuts, mount points and spacers are not counted, nor are pages below a page you may not see in the page tree. Each translation counts on its own, and the label says so (Pages and translations without a meta description). The number is kept for five minutes and recounted as soon as a page is saved. Start now starts the SEO tour on the most recently changed of them, beside the page module. The tile is left out where you may not edit the meta description or the SEO tour is not available to you.

A count of 0 is shown as done (No missing alt texts – great!, Every page you may edit has a meta description – great!). Broken links and outdated content are not counted: the extension has no exact source for them.

Below the tiles, the points guided tours left open: changes a tour proposed that an editor skipped, until a change to the same record and field is applied. Each row starts the matching tour on its page, in the language the point is about. A point is shown only when the editor may access its page, read the changed record and field, and edit that language; a point whose skill no longer exists is not shown.

Check website

Site-wide checks, one row each. Missing alt texts is always listed and opens the list, also when nothing is open. Pages without a meta description is listed while one is open and starts the SEO tour as the tile does.

Only tours and tasks the chat can actually start are offered: a configured skill that does not exist (yet), is not attached to the chat, or cannot run on this installation is left out, and a widget left empty says that nothing is set up yet. For a user the chat is not available to, each widget says so instead. Offering a page is not permission to change it: what the assistant does on the page is checked again when it does it.

Optimise an alt text with the assistant 

Optimise in a row opens the chat beside the list, on that image. The module menu folds to make room and unfolds again when you close the chat; the row the chat works on is marked, and its button reads In the chat. The chat starts with a card of the image: where it is used, its language, and that it has no alt text. The assistant then proposes a text for this use of the image, with these answers:

Apply

Saves the text. The chat confirms it with Alt text saved., and the image leaves the list. If it cannot be saved, the assistant says why in one sentence, the proposal stays, and Offer it again brings it back as a new proposal.

Another variant

The assistant proposes a different text. You can also write a wish, such as Shorter, and get a new proposal.

Mark as decorative

For an image that adds nothing to the content (see Decorative images), offered when you may set the mark. The assistant then proposes the mark; you confirm it once, with Mark as decorative on that proposal, and Create an alt text asks for a text instead. For an image that belongs to a section or a record the proposal says so, because such an image probably needs an alt text. Where the assistant itself finds the image decorative, it proposes the mark in the same way; where you may not set the mark, it does not, and such a proposal cannot be confirmed.

Skip

Saves nothing; the image stays open in the list.

Optimise the next hit below the conversation, or Optimise in another row, moves the conversation to the next image: a line Hit changed names it, and its card follows. The next hit is the one after the current image in the order of the list, across the tabs. Optimise while the assistant is still writing its proposal takes effect when it has finished, unless you close the chat before. Closing the chat gives the keyboard focus back to the row it was on, and unfolds the module menu it folded.

The assistant changes nothing but the alt text of this one use of the image, and only after you confirm it; a proposal for another image cannot be confirmed in this chat. It sees the image where the AI provider supports images and the privacy setting lets it go (Image analysis (privacy)), as a scaled copy. Otherwise the card says that the proposal rests on the page, the element and the file name only, and why: image analysis is switched off in the settings, the image lies in a folder whose images are never sent (the card names the folder), the provider cannot see images, or the image cannot be read. The preview of the image in the chat and in the list stays; it is shown in the backend only.

Optimise works in the live workspace and in a draft workspace, when an administrator has enabled the tools set_file_reference_alternative_text and set_file_reference_decorative in nr-llm's Tools module, and when you may approve your own changes in the chat (Approving a step in the chat). Otherwise the button stays in the list, and a note above the list says why it does not work.

In a draft workspace the list shows the images as that workspace shows them, and the alt text or the decorative mark you approve is saved there as a draft; live stays unchanged until the workspace is published. The approval card names the workspace. Two exceptions: an image whose alt text is set in live and emptied only in the workspace is not listed, and an image on a record whose page, or on an inline record whose owner, exists only in the workspace is listed but cannot be optimised there. Where an administrator turned versioningWS off or set versioningWS_alwaysAllowLiveEdit for sys_file_reference, Optimise works in the live workspace only, and the note above the list says so.