What is an author's note in AI roleplay?

Why it sits near the bottom

An app builds each prompt from the top down: system prompt, character card, persona, older history, then the newest messages. In a long story the system prompt ends up tens of thousands of tokens away from the reply being written, and models lean hardest on what they read last. An author's note solves that by riding along a fixed distance from the end of the chat.

That distance is its depth. At depth 4 the note goes above the last four messages, so the end of the prompt looks like this:

• ...older messages
• [Author's note]
• message 7
• message 8
• message 9
• message 10
• (the model writes message 11)

When message 11 arrives, the note moves down with it and stays four from the bottom. Chat depth explains the counting in detail.

An example at depth 4

Say you are in a slow-burn story with Sophie Lane and the scene keeps rushing toward a confession. A note like this, placed at depth 4, holds it back:

[Author's note: Slow burn. Sophie is drawn to {{user}} but will not admit it yet; she deflects with jokes and changes the subject. It is raining and the café closes in ten minutes. Keep replies under 250 words.]

The brackets and the label tell the model this is direction from outside the story, not something a character said. When the scene changes, you edit the note: the rain stops, the café is closed, a new tension takes over.

What belongs in a note

• The current state of the scene: where, when, who is present, what just happened.
• Tone for this stretch: dread, banter, grief, short sentences.
• A plot nudge: a stranger has followed them since the market.
• Style that keeps slipping: third person, past tense, no speaking for {{user}}.
• Temporary facts the character must remember right now, such as a lie they told two scenes ago.

Things that are always true about the character belong in the card, and world facts that only matter when mentioned belong in a lorebook.

Common mistakes

• Writing an essay. The note is read on every reply; one to three sentences is plenty.
• Depth 0. Right at the bottom, some models quote the note or answer it as if you said it. Depth 2 to 4 is safer.
• Forgetting to update it. A note that says it is raining keeps it raining for two hundred messages.
• Addressing the character ("Sophie, please be shy"). Write it as stage direction, not dialogue.
• Contradicting the card. If the card says bold and the note says shy, replies wobble between them.

Where you find it

The idea comes from story-writing tools such as KoboldAI and NovelAI, and SillyTavern made it standard in roleplay. In SillyTavern it lives in the options menu beside the message box, though labels can change between versions. You can keep a note for one chat, a note for one character that follows them into every chat, and a default note for new chats. Each has a placement (after the main prompt, or in the chat at a depth), a depth, a role for Chat Completion APIs, and an insertion frequency, where 1 means every message.

Apps without the feature give you much the same effect through a bracketed out-of-character line at the end of your own message; see steering the story with OOC. That works on the UnsavedAI website too.

Author's notes with UnsavedAI

When you connect SillyTavern to UnsavedAI with an API key, SillyTavern inserts the note into the messages before sending them, and the model reads it like any other text. There is a speed bonus to placing it in the chat rather than at the top: UnsavedAI reuses the unchanged start of a prompt between messages, so editing a note near the bottom costs little, while editing text near the top makes the whole chat be read again.

Questions people ask