The Slides AI

DataMind's AI, Kartlos, is also the author inside Slides: with a presentation open in the editor, the AI panel beside it builds and reworks the deck from your request, and every change lands on the slides as it works. This page covers how that author behaves: how it composes slides, how it reads a deck before touching it, and the rules its changes follow. Opening the panel and the permission it needs are on Write slides with AI.

What it works on

One conversation is bound to one presentation: the deck you have open in the editor, where its edits appear live on your screen.

• It never saves on its own. Your editor's save setting decides when its changes are written; if the editor saves only when you press Save, the AI tells you once that its changes are held in your browser until you save.
• If the deck has a newer version on the server, it refuses to edit and asks you to refresh the deck first, so nothing it does is discarded. Saving and version history covers refresh, overwrite and revert.

How it composes a slide

A slide is composed from a layout, never from hand-placed boxes. The AI names a layout and hands over the words and the numbers; the layout engine works out every position from the deck's size and the number of items, which is why its slides line up with the deck's margins and why overlapping elements and off-slide frames do not appear.

• There are fourteen layouts, from Title, Section break and Bullet list to KPI grid, Chart with takeaway, Table and Closing. Slide layouts describes each one and what it holds.
• Content a layout cannot accept builds nothing for that slide, and the reply says what to change. A layout that can place only part of the content reports what it left out.
• It edits the deck directly: you never receive slide markup, code or a file to paste in.
• It moves or resizes an element by hand only when you ask for something a layout cannot express, or to repair a problem a check reported.

How it reads your deck

Before it writes, it reads, and it reads no more than the request needs.

• It starts from the deck's outline: one row per slide, with what the slide claims, its layout and what sits on it.
• It reads only the slides a request touches, in windows of up to twelve slides per read. The whole deck is read only when the task genuinely needs every slide at once.
• A read too large to return comes back refused and asks for a narrower read instead, rather than a trimmed answer that hides what is missing.
• An element's vector markup can be read back up to 16 KB; a larger vector is refused rather than shortened.

How it edits, in steps

• It writes slides in batches, so the deck is built in front of you: slides appear one after another instead of arriving after one long silence.
• It adds up to twelve slides in one step, and builds a deck of more than about fifteen slides over two rounds.
• Every change is one undo step: press Ctrl+Z once to take back a change the AI made, exactly as with your own edits. A pass that touches several slides is still one step.
• When it builds a fresh deck, it fixes the final order at the end of the run, so the sequence does not depend on the order the slides happened to land in.

Motion, page numbers and the theme

Motion is part of what the layouts themselves emit, not a separate pass you have to request. Animations and transitions covers the editor side; this is what the AI's own work carries.

• A slide it builds arrives already choreographed: entrances stagger down the slide, KPI figures count up, and the statement layouts (Title, Section break, Quote, Closing) carry a slow ambient backdrop.
• The engine wires the deck's transitions for you: a resting Fade of 400 ms at ordinary junctions, and a Morph of 620 ms where the deck's travelling accent motif crosses from one section into the next. They are set again as the deck changes, so there is no choreography pass to ask for.
• On request it can still add a deliberate touch, such as pairing two elements so one morphs into the other between two slides.
• Ask for a plain or static deck and it honours that fully: no entrances, no count up, no backdrop, no travelling motif and no transitions, while the composition stays exactly the same. If plain was meant that broadly, the page numbers are left out too.
• Page numbers: on, in the Outline numeral style at the bottom left, skipped on the first slide. It can set another shipped style, or hide numbering on chosen slides.
• Page numbers are never drawn as text or vector elements: the AI switches on the deck's own automatic numbering, which follows each slide's position, so the numbers stay right when slides move.
• Colours are theme tokens, such as token:ink, token:accent and the deck's twelve named chart colours, never fixed values, so its slides match the theme and change with it. It cannot freeze the deck's palette; that choice lives in the deck's theme settings. See Design a deck.

How it checks its work

Before it tells you a deck is done, it grades it.

• The check measures overlapping elements, frames off the slide or into its margin, text and tables taller than their boxes, empty elements, slides with no headline, and colour pairs below the WCAG AA contrast ratio.
• Findings come worst first, and each one names the elements it is about and the fix for it. The AI repairs what it can, then re-checks the slides it touched.
• A clean report is the evidence it needs; it does not call a deck finished without one.

A deck from a chat

A deck can also start in a chat: a Mind builds a presentation from your conversation, analysis first, and it is saved in Artifacts when it is ready. That flow is covered on Charts, KPI cards and diagrams in answers.

What it does not do

• It does not publish, duplicate, move or delete a presentation. It says so plainly and points you to the control that does.
• It does not run queries, and it cannot bind an element to one. To put live figures on a slide, bind the element yourself; Fill slides from your data covers that.
• It does not export the deck. Exporting is yours, from the deck's export controls; Export and import covers them.
• An older editor may not carry every tool the AI can use. When a call is refused with not implemented in this editor build, the AI falls back to reading and updating slides one at a time instead of retrying it.

Questions, answered

Why does it ask me to refresh the deck before it edits?

Because the deck has a newer version on the server, so any edit now would be discarded. Refresh the deck (overwrite or revert in the save panel), then ask again. See Saving and version history.

How are elements placed on a slide it builds?

By the layout, not by coordinates. The AI names a layout and supplies the words and the numbers, and the layout engine computes every frame from the deck's size and the item count. That is why its slides line up with the deck's margins and overlapping elements do not appear.

Can I ask for a plain, static deck?

Yes. Ask for plain, static or no animations and every baked effect is left out, including the transitions, while the composition stays the same. If plain was meant to include the page furniture, the page numbers are left out too.

How many slides can it add at once?

It adds up to twelve slides in one step, and builds a deck of more than about fifteen slides over two rounds. Either way you watch the slides arrive as they are written.

Why are my page numbers wrong after I move slides?

They follow the slides, so they stay right. The AI does not draw page numbers itself; it switches on the deck's automatic numbering, which is drawn from each slide's position. A number built from a text box is what goes stale on a reorder, and that is the approach it avoids.

Will its slides match my theme?

Yes. It styles with the deck's theme tokens rather than fixed colours, so its slides follow the theme, chart colours included, and they change with it. It cannot freeze the deck's palette; that choice is in the deck's theme settings. See Design a deck.

How does it know a deck is finished?

It checks the deck, repairs what the check reports and re-checks the slides it touched. A clean report is the only evidence it accepts, so it does not call a deck done without one.

Why did nothing change?

The request matched what the deck already held, for example a value that was already set. The AI reports that instead of passing it off as a change, for example No change: the deck already held every value asked for.