> ## 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.
# Inside Quarkle Chat
> How Chat decides what to read, which tools it uses, and how it proposes edits.
[Quarkle Chat](/ai/quarkle-chat) looks like a chat box. Under the hood it's an **agent**. It can decide to go read parts of your book before it answers, rather than replying from memory. Here's how a single message becomes a grounded reply.
## What Chat always sees
Before you say anything, every conversation already carries a compact picture of your book.
* Your **[Story Atlas](/ai/story-atlas)** knowledge map, your **[Memory](/ai/memory)** (preferences), and your **Book Notes**. All three are always included, even if they're empty.
* If you have a chapter open: that chapter's **summary** (what's already on the page) and its **Chapter Notes** (your working instructions for it, beats, character mindset, tone, anything ephemeral and scoped to this chapter). Other chapters aren't summarized into context automatically; Chat reads them on demand with tools.
* Which **chapter** you're currently in.
* The conversation so far. Long chats are automatically **condensed** so older messages don't crowd out the context window.
## Reading on demand: the tools
Sometimes your question needs more than the summary: the exact wording of a scene, every mention of a character, a fact from three chapters back. Then the agent uses **tools** to fetch it. Think of these as a small toolkit for navigating your manuscript.
* **Browse and read chapters.** List them, open one, or read around your cursor.
* **Search.** Semantic search ("find the scene where they argue about the light") and literal search ("every line containing *Vellmoor*").
* **Read and update the Story Atlas.** Pull in a character note, or write a new fact back into it.
* **Search past chats.** Recall something you worked out in an earlier conversation.
* **Web search.** Look up a real-world fact when a question calls for it.
The tool chips you see above a reply are exactly this: a record of what Quarkle read before answering.
## Thinking only as hard as needed
To keep replies fast, Quarkle sizes the effort to the question. A quick check of one line shouldn't cost the same as untangling a subplot. So each message is first sized up. Is this a quick chat or a real edit? Simple, standard, or deep? That sets how many steps the agent may take and how long it may spend. A greeting gets answered instantly. A hard structural question gets a longer leash to read more and think more.
## How edits appear
When you ask for a change, Quarkle doesn't overwrite your text. It marks the proposed edit inline as it writes, and the Editor turns that into a **suggestion card** (an add, a rewrite, or a deletion) that you **Accept** or **Reject**. The words only land in your manuscript when you say so.
This is the same principle everywhere in Quarkle: the AI proposes, you decide. See [Quarkle Chat](/ai/quarkle-chat) for the day-to-day flow, and the [Toolkit](/ai/quarkle-chat) for how to steer what it reads.
One Run review, many editors.