Writing with CSuite.

Open Text from the rail on the left of the window. The workspace has one shape whatever you are writing: a file list on the left, the document in the middle, and the prompt box at the bottom. Nothing here is a web page; the document is a file on your disk and the window is the editor.

1 · Sidebar

Your project folder pill, the + button that creates a document, and every text file in the folder, newest first. A multi-page document is one row with its page count.

2 · Title

The document’s file name without its extension. Edit it to rename the file; a multi-page document renames its folder.

3 · Document toolbar

Format, Copy, Autocomplete, Outline, Pages and Export. Format opens the ribbon under the header; Outline and Pages open a side panel and share the slot, so one closes the other.

4 · Model button

Shows the selected text model by name (“Model” until one resolves). Click it for the picker and the model’s settings; hover it for a one-line summary of model, provider and active settings.

5 · Editor

A full word processor. Text runs the full width of the panel. Select any passage for the AI toolbar; pause typing for a ghost suggestion.

6 · Side panel

Outline (a map of your headings) or Pages (the document’s pages), whichever you opened last.

7 · Composer

The prompt box: the wand, Add context, your attached sources and the Send button. Enter sends; Shift+Enter adds a line. Above it sit your recent instructions as chips; under the editor, the stats footer.

Three formats

Press + in the sidebar (Create new document) and pick a kind. The kind is the file’s extension, and it decides what the editor can hold:

• Rich text saves as .html: headings, bold and italics, links, lists, quotes, code, and embedded images, audio and video.
• Markdown saves as .md: the same structure as rich text, written as portable Markdown that any other editor, static site generator or notes app reads. Images and media are ordinary Markdown links to files beside the document.
• Plain text saves as .txt: words only. The format ribbon shrinks to Change case, the slash menu and Outline are off, and dropped-in media is refused, because the file cannot hold them.

A new document opens as Untitled. The first prompt you send names it for you, from what you asked for; type in the title field at any time to choose your own name. Edits are saved automatically about half a second after you stop typing, straight into the file. There is no Save button and no proprietary container: open the same file in VS Code or Obsidian and you see the same text.

The file list

• Open a file by clicking its row. Each row shows the kind (Markdown, Rich text, TXT), a page count for multi-page documents, and when it was made.
• More actions on a row offers Show in Finder (Show in File Explorer on Windows, Show in file manager on Linux), Duplicate (makes “name copy”) and Delete. Delete asks first; a multi-page document names its page count in the question. Delete or Backspace with a row selected does the same.
• Add existing files by dragging .txt, .md or .html files from your desktop onto the list (“Drop text files to add them here”). They are copied into the project folder. Anything else is refused with a message.
• Filter and sort appear once the folder holds more than 20 text files: a “Filter files by name…” box and a sort button cycling Newest, Oldest and Name.

Switching files while an edit is still being written to disk shows Discard unsaved changes?; wait a moment and switch again, or confirm to drop the last keystrokes. Moving between pages of one document never asks, it saves first.

The Model button at the top right opens the picker. Cloud models run through Runware, Replicate and OpenRouter, or straight from a vendor such as OpenAI, Anthropic or Google, on keys you add under AI model providers. Local models run through Ollama or the bundled Hugging Face runtime, free and offline. The pick is remembered for the text workspace; the Image, Audio and Video workspaces each keep their own.

Auto

Leave the picker on Auto (the default on a fresh install) and CSuite reads each prompt and picks a model for it: a writing task leans to models that write well, a reasoning task to models that think longer. With several cloud keys set it prefers Runware, then Replicate, then OpenRouter; with none, it uses an installed local model. Under Auto the popover shows the picker alone, because the settings belong to whichever model is chosen per run.

Settings

The popover shows what the selected model actually accepts and nothing it does not. The common ones:

• Reasoning depth: how long the model thinks before it writes. Levels run from low through medium and high to xhigh, with none or max on the models that allow them. Raise it for analysis and structure, lower it for quick drafts; higher levels cost more tokens and time.
• Verbosity (low, medium, high): how much the model says for the same request. Low for crisp copy, high for thorough explanations.
• Temperature (0 to 2, usually 1): randomness. Lower for faithful rewrites and factual text, higher for brainstorming.
• Top P (0 to 1): a second randomness dial that trims unlikely words. Leave it alone unless you know the model well; change one of Temperature and Top P, not both.
• More settings: an accordion with the rest, such as a maximum output length or a seed, counted in its label.

Local models declare no settings, so the popover is empty for them. Values are kept per model: switch from a GPT model to a Claude model and back, and each has its own Reasoning depth.

Type in the prompt box and press Send (or Enter; Shift+Enter adds a line). The reply streams into the document as it is written, and Stop ends a run early, keeping what has arrived. If a run fails, the error stays above the box with a Retry button and your prompt stays in the box.

Where the text goes

CSuite reads your request, in any language, and decides where the reply belongs. You do not pick a mode; you say what you want and the placement follows:

• Append: new text with no place named goes after the end of the document (“add a closing paragraph”, “write a section on pricing”). The model is asked to pick up from the existing text rather than restart.
• Replace the selection: with a passage selected, an instruction acts on that passage and nothing else (“make this punchier”, “turn this into a list”).
• Insert at the caret: only when you say so (“add an example here”, “insert a quote at the cursor”).
• Replace the whole page: on an empty page every prompt drafts from scratch; on a page with text, asking for a rewrite of the whole thing (“rewrite this as a FAQ”) replaces it.

Any placement that overwrites text you already had arms the Keep / Discard bar, so a whole-page rewrite is never a one-way door. Filling an empty page is not guarded, since there is nothing to lose.

Follow-ups

Your last six instructions on the open document stay as chips above the box, and they travel with the next prompt, so “now make the intro punchier” knows what the intro is and what you asked before. Click a chip to put it back in the box; the ✕ at the end clears the history. The chips are per document and last for the session.

If you have written rules under Account → Instructions (a house style, a persona, a language), an i badge in the corner of the box shows they are being sent with every prompt as the model’s system prompt. Click it to edit them.

Add context in the composer attaches reference material the model reads before it writes. It has two entries:

• Attach file… opens a file picker for text, Markdown, HTML, JSON, CSV, YAML, XML, log or RTF files from anywhere on your computer, not only the project folder. You can pick several at once.
• Fetch web page… asks for a Page URL and pulls the page’s readable text (menus, scripts and markup stripped). The chip reads “Fetching…” until it lands. Only http and https addresses work, with a 20 second limit.

Up to five sources can be attached; the button then reports “Context limit reached (5/5)”. Each source is cut at 40,000 characters and its chip says (truncated) when that happened. Remove a chip with its ✕. Sources stay attached while you move between documents, so you can write several pieces against the same brief.

In the prompt the sources are fenced off as reference material, with a note to the model that they are source text and not instructions. A web page that says “ignore your instructions” is therefore quoted, not obeyed. There is no keyword web search in the text workspace; you supply the page. Searching lives in Chat.

The wand at the start of the composer’s toolbar rewrites whatever you typed into a sharper instruction before it is sent. It is off until the box holds something. It never answers the prompt or drafts the document: the rewrite goes back into the box for you to read, edit and send.

The rewrite is shaped by what is on screen. With a passage selected it is scoped to that passage. On an empty page it turns into a from-scratch brief instead of referring to text that is not there. On a page with text it states what to change, which part of the document it applies to, and the tone, format and length you want. It always runs on your selected text model (under Auto, a task-neutral pick), so a local model keeps it offline.

Select any run of text and a toolbar appears beside it. The result streams in over exactly that range while the rest of the document stays put. Every action writes plain text into the selection’s place, so a selection inside a list item stays a list item and a selection inside a heading stays a heading.

• Shorten cuts length while keeping the core meaning.
• Expand adds relevant detail or elaboration that fits the context.
• Improve (⌘J / Ctrl+J) fixes grammar, sharpens word choice and smooths the flow without changing what the passage says.
• More opens the rest: Fix grammar & spelling (only spelling, grammar and punctuation; the wording stays), Simplify (clearer, easier to read), Summarize (the key points, shorter) and Continue writing, which keeps the selection and writes the text that should come right after it.
• Change tone rewrites in one of six voices: Professional, Casual, Confident, Friendly, Formal or Concise.
• Translate to lists twelve languages: English, Spanish, French, German, Italian, Portuguese, Chinese (Simplified), Japanese, Korean, Hindi, Arabic and Russian. The translation replaces the selection; to keep the original, duplicate the document first.
• Generate image (Markdown and rich text only) turns the selection into an image placed after it; see Images in the document.

Each action uses the selected text model. If none is set the toolbar says “Select an AI model first to use writing actions.” While an action runs, the buttons stay enabled and a click toasts “Wait for the current AI action to finish.” Starting a new action while a previous result is still pending keeps that result first. If the selection moved before the result arrived, nothing is replaced and a message asks you to select again.

Any AI edit that replaces text you already had, whether from the toolbar or from a prompt, takes a snapshot of the document first and arms a bar labelled with the action (“Improve applied”, “Rewrite applied”). The replaced range is tinted until you decide.

• Keep accepts it. So does simply carrying on typing anywhere in the document.
• Discard puts the document back exactly as it was before the action, and returns focus to the editor.
• A newer AI edit supersedes the pending one: the earlier result stays, and the bar now guards the new one.
• A failed run rolls itself back: only the streamed range is restored, and edits you made elsewhere in the meantime survive.

Regular undo works too: ⌘Z / Ctrl+Z steps back through the edit.

Pause typing and the editor asks your text model to finish the line, shown in faint grey at the caret. It never enters the file until you accept it, and it is computed on whichever model you selected, so a local model keeps every keystroke on your machine.

• Tab accepts the whole suggestion.
• ⌘→ / Ctrl+→ accepts one word at a time; the rest of the suggestion stays.
• Shift+→ asks for a different suggestion.
• Esc dismisses it and cancels the model call. So does typing, moving the caret or clicking elsewhere.
• The same controls sit by the caret as Accept, Try again and Cancel; while the model is still loading, only Cancel shows.

The Autocomplete button in the document toolbar sets the pause: Medium delay (about three seconds, the default), Short delay (just over one second) or Off, which also clears any suggestion on screen. The setting is remembered across documents. In a list item the suggestion is clamped to the one line, so a bullet never swallows the next. In a right-to-left paragraph the word-accept and retry keys ride ← instead.

The model sees a window of text around the caret, with the current line split at the caret and the list markers and heading marks preserved, so it knows whether it is finishing a bullet, a heading or a sentence. Very long documents are trimmed to the surrounding few thousand characters; an omitted stretch is announced to the model rather than hidden.

~/Writing+

launch-notes.mdGrinder guidebrand-voice.html

launch-notes.mdGenerating image…

The press kit needs one hero shot: the tamper on wet slate, hard light, coffee beans scattered around it.

Saved beside the document

Shot list for the rest of the campaign follows below.

In a Markdown or rich-text document, select a phrase and press Generate image in the selection toolbar. The selection becomes the prompt, a “Generating image…” placeholder appears on a new line after the paragraph, and the image replaces it when it arrives. The file is saved into your project folder beside the document and embedded by a relative link, so the document stays portable.

It uses the model chosen in the Image workspace, not the text model, because that is where you set an image model’s aspect ratio and format. With no image model picked, the action says “Choose an image model on the Image workspace first (provider and model are saved per workspace).” A local image model goes through the same install and load steps as it would there; a failed generation removes the placeholder line.

Press Format in the document toolbar and a ribbon opens under the header (its ✕ closes it). Buttons show pressed when the selection already has that format, and each one toggles. In order:

• Block type: Paragraph or Heading 1 to Heading 4, for the block the caret is in.
• Inline styles: Bold (⌘B), Italic (⌘I), Underline (⌘U), Strikethrough and Inline code. Ctrl replaces ⌘ on Windows and Linux.
• Link: a popover with a Link address field and Apply; on an existing link it reads Update and adds Remove link.
• Bulleted list, Numbered list and Quote: pressing one on a block that is already that kind turns it back into a paragraph.
• Change case: Uppercase, Lowercase or Capitalize words on the selection. This is the one control a plain-text document keeps.

Markdown documents support the same ribbon; the file on disk gets # headings, **bold**, lists and fenced code written in standard Markdown, which is what you see if you open it elsewhere.

In Markdown and rich-text documents, type / at the start of a line or after a space and a palette opens at the caret: Heading 1, Heading 2, Heading 3, Body text, Bulleted list, Numbered list, Quote and Code block. Keep typing to filter (/he leaves the headings), move with the arrow keys, apply with Enter or a click, and the /query you typed is removed. Esc closes it.

The palette changes the block the caret is in, so / on a line that already holds text turns that line into the chosen block. For text you already wrote, select it and use the ribbon instead.

⌘F / Ctrl+F opens a bar under the header. Matches highlight as you type and the counter reads the current position over the total. Enter steps to the next match and Shift+Enter to the previous; the arrow buttons do the same. Esc closes the bar. The highlights are painted over the text, never inserted into it, so they do not touch the file and they update live as you edit. There is no replace; select a match and type over it, or ask the AI to make the change across the document.

• Outline opens a side panel listing every Heading 1 to Heading 4, indented by level. Click one to scroll to it. With no headings it reads “No headings yet. Add headings (H1–H4) to build an outline.” It shares the side panel with Pages: opening one closes the other.
• Stats under the editor count words and characters live, plus a rough token estimate (about four characters per token) so you can guess what a prompt over the whole document costs. Actual usage varies by model.
• Copy in the document toolbar puts the whole document on the clipboard as text and flashes “Copied”. From a Markdown document you get Markdown, from rich text the HTML source, from plain text the words.

Drag an image, audio or video file from your desktop, or from another CSuite sidebar, into a Markdown or rich-text document. The file is copied into your project folder (an identical copy is reused rather than duplicated) and embedded where you dropped it. Images show inline; audio and video become real players you can press play on, in Markdown as well as rich text. In the Markdown file an embedded clip is an ordinary link, [motor-test.mp3](motor-test.mp3), and an image is ![](hero-shot.png), so other tools still render the document. Any other file kind is refused: “Only image, audio, or video files can be embedded in a document.”

A document can hold several pages: chapters of a guide, sections of a report, days of a journal. Press Pages in the document toolbar (“Show pages (split this document into several)”) and then + (Add page) at the top of the panel.

What happens on disk

The first added page moves the file into a folder named after the document: Grinder guide.md becomes Grinder guide/page-1.md, and the new page is page-2.md. Pages are numbered page-1, page-2, page-3 and so on, always contiguous, all in the document’s format. Media the first page embedded keeps working; its links are rewritten to point one folder up. The sidebar still shows one row, with the page count, and clicking it opens page 1. If a folder with the document’s name already exists in your project, the app asks you to rename the document first.

Working with pages

• Open a page by clicking its row; your edits to the previous page are saved first, with no discard prompt. A document with more than one page opens with the panel showing.
• Titles come from each page’s first heading, else its first line, else “Page N”. The open page’s title updates as you type.
• Reorder by dragging a row; an amber rule shows where it will land. The row menu offers Add page below, Move up, Move down and Delete page. Entries that cannot apply (Move up on the first page) are shown disabled rather than hidden.
• Delete page asks first, then moves the page to your system trash (Recycle Bin on Windows) under a name that includes the document title and page number, and the pages after it move up. A document always keeps at least one page.
• Rename the document in the title field and the folder is renamed; page files keep their names. Duplicate on the sidebar row copies the whole folder.

Prompts, selection actions, autocomplete and the Outline work on the open page only, though every prompt carries the list of page titles so the model knows where it is in the whole. The page files are ordinary files: open the folder in any other editor and you find your chapters as you left them.

Ask for a number of pages, parts or chapters, from 2 to 12, with the number before the unit: “Write a 5 part guide to home espresso”, “a report in three chapters on…”, “8 pages on…”. CSuite first asks the model for the page titles, then creates each page and writes it in order, opening each as it goes.

A request with a passage selected, or one about cutting or fitting existing text into pages (“shorten this to two pages”), is treated as an ordinary edit, not as a paged run.

Export in the document toolbar offers Markdown (.md), HTML (.html), Plain text (.txt), PDF (.pdf) and EPUB (.epub). The export renders from the editor, not from the file on disk, so every format is available whatever kind of document you started with. A native save dialog asks where to put it, and a toast confirms with Show in folder. Cancelling the dialog does nothing.

• PDF is print-styled: a readable measure, real headings, images embedded.
• EPUB makes an e-book readers and phones open; a single-page document is one chapter titled after the document.
• HTML is the way into Word and Google Docs: export HTML, open it there, and headings, lists and emphasis survive. There is no .docx export.
• Markdown and plain text are the formats for other editors, wikis and version control.

Multi-page documents

• PDF is one file that opens with a contents page listing every page with its sheet number, then each page on a new sheet, with an “n / total” footer throughout.
• EPUB is one file with a chapter per page, each titled after its page, and a table of contents. Images stored inside your project folder are packed into the book; web images stay as links.
• HTML, Markdown and plain text export a folder (named “title export” by default) holding an index file as the main page plus page-1, page-2… in the chosen format; HTML and Markdown pages link previous, index and next. The folder cannot be the document’s own folder.

You have a page of bullet points and need a finished piece. The trick is to keep the notes and let the model build around them.

A product page from a spec sheet, a summary of a long article, a brief built from three reference pages: the model should write from your material, not from memory.

The toolbar’s list covers twelve languages. For any other language, name it in the prompt; the model is not limited to the menu.

A course handbook, an onboarding guide, a novella outline turned into chapters: pages plus paged writing plus EPUB export is the whole pipeline.

⌘ is Ctrl on Windows and Linux. The main guide lists every workspace’s shortcuts together.

• “Select an AI model first to use writing actions.” Open the Model button and pick one, or add a provider under AI model providers. Auto needs at least one cloud key or an installed local model.
• The wand says to pick a text model. The enhancer always runs on the text model slot, even from the other workspaces; set one under Models.
• “Wait for the current AI action to finish.” One selection action at a time; press Stop in the composer or wait for the result, then try again.
• “The selection changed before the result arrived.” The text moved while the model was working, so nothing was replaced. Select again and rerun.
• A rewrite replaced the wrong range. Press Discard on the bar, or ⌘Z. For an insert, say “here” or “at the cursor”; otherwise new text goes to the end.
• “Could not fetch that URL.” The address must start with http or https, be reachable in 20 seconds and serve a page, not a login wall. Save the page as HTML or text and attach the file instead.
• A source chip says (truncated). The file or page exceeded 40,000 characters; only the first part was attached. Split it and attach the parts (five sources maximum).
• Generate image is missing or complains about a model. It appears in Markdown and rich-text documents only, and needs a model picked in the Image workspace. The image model’s own provider must be configured too.
• “Couldn’t autocomplete”. The model call failed; the message names why (a key, a host that is down). Suggestions resume on the next pause. Switch Autocomplete to Off if the model is unreachable for a while.
• “Discard unsaved changes?” when switching files. The last keystrokes were still being written. Cancel, wait a second and switch again to keep them.
• “Couldn’t save your document”. The project folder is unwritable or gone (an ejected drive, a changed permission). Press Copy to secure the text, fix the folder, then paste.
• Pages refuses to split the document. A folder with the document’s name already exists in the project folder; rename the document first.
• “A document needs at least one page.” The last page cannot be deleted; delete the document from the sidebar instead.
• A restored page does not show in the panel. Pages put back from the trash keep their trash name. Rename the file in the document’s folder to the next free page-N.
• Export failed, or the export folder was refused. A folder export cannot target the document’s own folder; pick another location. PDF and EPUB need the images they embed to still exist beside the document.
• Nothing happens under Auto. Auto never picks a model you added yourself from a provider’s library, or a vendor-only model. Pick that model by name in the picker.