Author's note: steering the story from inside the chat
An author's note is a short block of instructions or context that a roleplay frontend inserts near the end of the chat history, where the model pays the most attention.
What an author's note is
Instructions at the top of a long prompt fade in influence as the chat grows. The author's note solves that by placing text close to the end, a few messages back from the newest one, on every generation. It is the frontend equivalent of leaving a sticky note for a co-writer: "It is night now and raining. Keep replies under 300 words."
The idea comes from NovelAI and KoboldAI and is built into SillyTavern, which lets you set a note per chat, per character, or as a default for all chats.
How it works
The frontend builds the prompt as usual, then inserts the note text at a chosen depth: the number of messages from the end. Depth 0 puts it after the last message; depth 4 puts it four messages up. In Chat Completion mode SillyTavern sends it as a message with a role you pick (system, user or assistant). An insertion frequency setting controls whether it is added every turn or only every N turns.
The note is resent each time, so its tokens are billed on every request. It can also carry lorebook entries whose position is set to the author's note.
What to put in it
- Current scene state: time, place, who is present, what just happened.
- Style rules that drift: length, tense, point of view, "do not speak for {{user}}".
- Temporary direction: "Mara is hiding that she is injured."
- Pacing: "Slow down, let the conversation breathe."
Keep it short. Two to five lines tend to work better than a long block, because a large note in the middle of the recent history can confuse the model about what is dialogue and what is instruction. Writing it in brackets, such as [Style: ...], helps mark it as out-of-story text.
Common mistakes
- Depth 0 with a user role. The model may answer the note as if you had said it in character.
- Forgetting it is on. A per-chat note from last week can keep steering a new scene.
- Contradicting the card. When the note and the card disagree, the note usually wins because it is closer to the end, which can surprise you.
Author's notes with Wild West API
The author's note is pure prompt assembly done by the frontend, so it works with every model. With chat completions it arrives as one more message in the messages array. It pairs well with a short system prompt: put stable rules up top and drifting ones in the note.
FAQ
What depth should an author's note be at?
Depth 2 to 4 is a common choice. It is close enough to the end to matter but not mistaken for the latest message.
Does the author's note cost tokens?
Yes. It is part of the prompt on every request it is inserted into, so it is billed as input each time.