Skip to content

AI Assistant

Writer Studio has one AI assistant: an agent you talk to in plain language. It can do the same things you can do in the app — search, open scenes, read the Book Wiki, edit scenes, run and read proofreading diagnostics — because you ask it, not because it has its own private buttons.

There are no AI-only features. Anything the assistant does, you can also do by hand, and the product is fully useful with AI turned off. AI is off by default.

Turning AI on and choosing a provider

AI does nothing until you enable it. The first time you enable it, you choose a provider — there is no default, because where your text goes is your decision.

Provider Where text goes
DeepSeek The provider's servers
GigaChat The provider's servers
YandexGPT The provider's servers
A custom OpenAI-compatible endpoint Wherever that endpoint points
Ollama (local) Stays on your device — nothing leaves

For cloud providers you bring your own key (BYOK): paste the key the provider gave you. The choice screen states plainly where your text will be sent. Pick Ollama if you want AI without sending any text off your machine.

Manage all of this in Settings → AI: enable or disable AI, change provider, and update your key.

Privacy. Your key and provider settings are device-level secrets. They are never stored inside the project folder, never exported, and never included in a backup — so copying or sharing a project never leaks your key. Manuscript text leaves your device only to the provider you chose, and only for an operation you asked for.

Talking to the assistant

Open the Agent panel, type a request in natural language, and the assistant acts through the same operations you would use. For example:

  • "Where does Olga first meet Igor?" — it searches and answers.
  • "Open the scenes where the harbor appears." — it navigates for you.
  • "Who is Captain Renn?" — it reads the Book Wiki.
  • "Rewrite this dialogue to be drier." — it proposes an edit.
  • "Run the language check on this chapter and list the issues." — it reads the proofreading diagnostics.

You never pick a command from a menu — you describe what you want. The same assistant handles everything.

Input shortcuts:

Key Action
Enter Send the message
Shift+Enter New line

Reviewing changes (you approve everything)

The assistant never changes your manuscript or your confirmed story knowledge on its own. When it wants to edit a scene or add to the Book Wiki, its output arrives as a suggestion you review first.

  • Apply — the change is written in. It is a normal edit, so a single undo reverses it.
  • Reject — nothing changes; your text stays exactly as it was.

Text edits show a diff so you can see exactly what would change, and you can expand the full suggested text before deciding. Suggested wiki entries appear as "suggested" and never become canonical until you confirm them. Available edit shapes include replacing text, appending to the end, and inserting at the start.

Asking your book

When you ask the assistant a question about your story, it answers from your manuscript using the same read and search tools you have — not from guesswork. Answers cite the scenes and entries they came from, and you can jump straight to that evidence.

The assistant follows the same rules your own view does: anything in the Trash is invisible to it, so trashed scenes never show up in answers or context.

Seeing what context was used

For any operation, you can see what project context the assistant used — which scenes, which wiki entries, which guidance. Nothing is hidden, so you can verify what was sent before trusting an answer.

AI Guide and suggested prompts

The AI Guide is a place to write standing instructions for the assistant — writing style, world rules, naming conventions, anything it should keep in mind. Guides attach at the level you want — project, book, part, chapter, or scene — and a deeper guide adds to the ones above it rather than replacing them. The assistant follows your guide, but it never overrides a direct request or the approval step.

Around the editor, the wiki, and the issues list you'll see suggested prompts — common asks, prefilled, one click from where you are. They are shortcuts into the one assistant, not separate features.

Conversations per book

Each book keeps its own list of conversations; chat history never mixes between books.

Action How
New conversation + in the panel header
Switch conversation Conversation list → pick from history
Rename Double-click the conversation name
Delete Conversation list → delete icon next to the name

Cost and context indicator

Below the input, the current session shows tokens sent and received and a rough cost estimate. It appears only while a conversation is active, so you always have a sense of what a request is costing.

Context window and compaction

A long conversation eventually fills its context window. When it's about 80% full, you'll see a warning, with two ways forward:

  • Compact — older messages are summarized and archived to free up room. You can undo a compaction. The archived part stays in the conversation as a collapsed "Archived" block.
  • New conversation — start with a clean context without losing your history.

Message controls

Action How
Copy a message Message context menu
Retry a request Retry under the last message that failed
Stop generation Stop in the input area

If a response takes too long, you'll see a notice that the reply didn't arrive in time — retry the request or start a new conversation with +.