While you build a custom visual, the assistant can work beside you: it reads the visual and its data, writes and patches the code, and checks how the result renders. Every step appears as its own card under the reply, and this page explains what those cards say. For the panel itself, see The AI assistant; for the shared card controls and the other card families, see The assistant's step cards. To write a visual by hand, see Custom visuals and the visual library.
Each card is titled by what the step does:
The card's body depends on what the step returns: a compact summary of rows, a findings list, a change preview, or a render report. Its title and detail are built from the request as it arrives, sometimes quoting what you asked for, so you can follow the step while it is still being written; a long detail is cut with an ellipsis.
Cards are coloured by the outcome: a step that only reads keeps the default colour, a change that lands turns green, a card that reads No change: turns amber, and a failure turns red.
A step that reads settles into a compact card of short labelled rows. You may see Fields, Roles, Source, Tile, Save, Suggested, Rows, Profiled, Examples, Assets, Kind and Size, depending on what the step has to report. A row with nothing to show reads none or no source; the Save row shows the save state, adds unsaved edits when there are changes that are not saved yet, and ends with the autosave state; and a count of declared roles reads N declared. Open Details for the full result the step returned.
Some steps report findings, such as a check that went over the visual's code. Each finding names its code, file and line, describes the problem, carries a severity pill (error or warning) and suggests a fix after Fix:. A check with nothing to report reads Clean, sometimes with the profile it ran with in brackets. A long list is cut short with N more not shown.
When a step writes or patches code, the card previews the change:
Blocks start collapsed behind a toggle that lists them with their line counts, and the toggle adds writing… while the text is still arriving, so a running write can read like visual.css (12 ln) · writing…. A line under the toggle lists each part's size in bytes.
A preview step reports what was painted, then how the render went.
The report under the chips reads Rendered or Render failed, followed by the facts of the run: the phase and how long it took in milliseconds, the number of nodes drawn, empty or overflow when they apply, and counts of errors, CSP violations and findings. Problems get their own lines, naming the error and its line, or the blocked address and the directive that blocked it. When the render failed, the Report payload opens by itself.
If the sandbox could not run the preview at all, the card reads Sandbox unavailable: followed by the reason. If the change was applied but the frame has not reported back yet, the card reads Applied; the frame has not reported yet.
The step finished without changing your visual, and the summary after No change: is what it returned. It is not an error; the amber colour marks that no change was made.
The step did not go through, and the Refused payload opens with its error text. Read the text to see why.
The colour shows the outcome: green for a change that landed, amber for a step that changed nothing (No change:), and red for a failed step. Steps that only read keep the default colour.
The step checked and found nothing to report. When a profile name follows in brackets, that is the profile the check ran with.
They summarise the drawing: the frame size with the dataset it drew from, the pixel size as width x height, and the number of rows. deterministic means every run painted the same result; paints differ warns that the runs did not match. A chip after export: names a failed export's error.
The preview could not run in the sandbox, and the reason follows after the colon. Ask the assistant to address the reason, then preview again.
The change reached the frame, but the render report has not come back yet. Give it a moment: the report appears on the card when it arrives.
They preview the change: Search is the text being replaced and Replace is the new text. Each block is marked - when it holds the old text and + when it holds the new text.
The list was cut short to keep the card readable, and the line counts the findings left out. Ask the assistant about the rest when they matter.