The AI panel in Markdown Studio is an editor collaborator, not a chat window bolted to the side. It works on the document you have open — drafting blocks, rewriting selections, proposing continuations as aiHighlight ghost text — and it can reach the web through a search skill when a draft needs a fact instead of a guess. Like every AI surface in Rixoffice, it runs on the provider you configured, from the built-in Rix AI to a fully local Ollama model.
What it is good at
Four jobs cover most of the daily use:
- Drafting — give it a title or a one-line brief and it produces a first outline or opening paragraphs you can reshape.
- Rewriting — select a block and ask for shorter, plainer, friendlier or more formal; the suggestion replaces the block only when you accept it.
- Continuing — park the cursor mid-thought and let ghost text propose the next sentence; take it, edit it or type past it.
- Summarizing — ask for the short version of a long note, useful for turning meeting scribbles into something sendable.
Ghost text: suggestions, not edits
The aiHighlight mechanism keeps you the author. AI-proposed writing appears as faded ghost text:
- Place the cursor where the suggestion belongs.
- The proposal appears in ghost style — visibly not yet part of the document.
- Accept by continuing to type into it, or edit it as it lands.
- Dismiss by moving on with your own words; the ghost text vanishes without a trace.
Nothing enters the file unless you take it. That rule keeps autosave honest — the .md on disk never contains a suggestion you did not choose — and it keeps the writing voice yours, with the AI as a fast junior draftsman rather than a co-author.
The web search skill
Language models know a lot and know it as of whenever they were trained. The panel's web search skill closes the gap: ask it to check the latest release version, pull a current statistic or confirm a API endpoint, and it searches before answering, then grounds the draft in what it found. Use it for anything time-sensitive; skip it for stable knowledge where the model alone is fine.
Prompts that work in a notes app
| Prompt | What you get |
|---|---|
| outline this note | A heading structure for a rambling draft |
| rewrite the selected paragraph in plain language | A clearer version of the same block |
| summarize these notes in five bullets | A sendable short form |
| check the current version number of this library | A search-grounded fact in the draft |
| draft the intro from the frontmatter title | A first paragraph that matches the properties |
The last one is a small synergy worth knowing: the properties set in the frontmatter panel give the AI immediate context about what the document is.
Where the panel fits a writing session
A typical session shows the division of labor without anyone deciding it in advance:
- You type the opening — the part that needs your voice.
- The AI outlines the rest from the title, giving the piece a skeleton in seconds.
- You argue with the outline, delete half of it, and write the sections that matter.
- Ghost text offers continuations in the sections that are purely mechanical — transitions, definitions, the summary of what was just said.
- Web search verifies the two facts the piece leans on.
- You do the final pass, because the reader is trusting one author, not a committee.
Writers who try to delegate the interesting parts end up rewriting everything; writers who delegate only the filler finish drafts in a sitting. The panel is tuned for exactly that contract — fast where the work is generic, silent where it is yours.
Providers, privacy and cost
Open Settings, AI provider setup, and the panel follows: built-in Rix AI on a plan, your own API key on BYO plans, or Ollama for writing with no internet at all. With a local model, ghost text and rewrites work on a plane — the document, the prompt and the suggestion never leave the machine. Provider setup lives in AI provider setup, and BYO keys are never metered by Rixcart: your provider bills your usage directly. On plan credits, AI pages remain unlimited across the suite, so the limiter that matters here is taste, not quota — how much of a draft you want to be yours.
When the draft is done, export to DOCX takes it to Word users intact.
One boundary deserves respect regardless of provider: the panel works on the document you point it at, not on your disk. It reads the open file and proposes for it; it does not browse your folders or touch other notes. Combined with local storage of the files themselves, that keeps the blast radius of any session to exactly one document — yours, chosen, in front of you.
What to explore next
- Getting started with Markdown Studio — the panel in context of the whole app.
- AI provider setup — choose the engine behind the ghost text.
- Frontmatter and properties — metadata the AI can read.
- Export to DOCX — the handoff once the draft is done.