> ## Documentation Index
> Fetch the complete documentation index at: https://docs.quarkle.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# How Quarkle's AI works
> A plain-language tour of what happens behind the Creative Companion: the context, the models, and the moving parts.
You don't need to know any of this to use Quarkle. But if you like to understand the machine, this section is for you. It's a plain-language look at what happens when you chat, run a review, or watch the Story Atlas update itself.
## Overview
Quarkle is built around one choice: **minimise AI writing, maximise AI critique.** Almost everything below serves that. The AI's job is to *read* your book carefully and give you sharp, specific feedback. It does not generate your novel. When it does propose words, they arrive as suggestions you accept or reject, never as silent changes.
## What "reading your book" really means
The key idea is **context**: what the AI can actually see when it answers you. Quarkle doesn't send your whole book to a model every time. Instead, each feature gets exactly what it needs.
* The **chapter** you're working on (or the passage you selected).
* Your **[Story Atlas](/ai/story-atlas)**: the knowledge map, chapter summaries, and character and world notes.
* Your **[Memory](/ai/memory)**: the preferences you've asked Quarkle to keep.
* For Chat, the ability to **go and fetch more**. It can read another chapter, search the manuscript, or look something up, only when it needs to.
This is why the Companion gets more useful as your book grows. The Atlas gives every feature a richer, always-current picture of your story.
## The moving parts
Which AI models power Quarkle, and how **Auto** picks one for each task.
The agent loop: how Chat decides what to read before it answers.
How one "Run review" becomes a coordinated pass by many editors.
How one long read becomes a full editorial dossier.
How the Story Atlas stays current without re-reading your whole book.
What's private, what's never used for training, and how mature content is handled.
## Where it all runs
Behind the Editor, Quarkle runs a streaming service that sits between you and the AI models. It assembles the right context, calls a model (or several), and **streams** the response back word by word. That's why Chat answers appear live, review comments arrive tier by tier, and rewrites unspool in place. When a task is better done in one shot, like a version's change summary, it's batched instead. The next pages open up each of these.
Which models Quarkle uses and how Auto picks one.