The author's note technique that changes AI conversations

There's a technique borrowed from creative writing AI tools that solves one of the most frustrating problems in AI roleplay and companion conversations: characters drifting out of voice, the AI rushing pacing, tone shifting toward generic friendliness, scenes losing coherence over long sessions. The technique is called the author's note, and it works on a simple principle. Instead of telling the AI what to do at the start of the conversation, you inject your instruction near the end of the context, where the model pays the most attention.

This is the single most effective intervention you can make in long AI conversations, and most users have never heard of it.

What an author's note actually is

In creative writing tools like AI Dungeon and SillyTavern, the author's note is a dedicated text field that gets injected into the prompt at a specific position you control. The position matters because of how language models pay attention to context. As we covered in the post on context windows, models attend most strongly to the very beginning and very end of their context. The middle gets weighted less. An instruction placed near the end of context, just before the model generates its response, gets significantly more influence on the output than the same instruction placed at the beginning.

The author's note exploits this. You write a short directive about how the AI should behave, the tool inserts it close to the end of the context (typically at depth 4, meaning before the most recent 3-4 messages), and the model treats it as one of the most recent and important pieces of guidance.

The convention dates back to AI Dungeon's adventure-game format, where authors needed a way to steer the AI's narrative voice without breaking the in-character flow. They settled on putting instructions in square brackets, which the model had been trained to interpret as authorial annotations rather than story content. AI Dungeon's help documentation explains the underlying logic: in fiction the model trained on, square brackets often appeared around editor's notes or out-of-band commentary, so the AI learned to read bracketed text as meta-information rather than narrative.

SillyTavern adopted the same pattern and made the position and frequency configurable. You can choose how deep into the context the note gets injected and how often it appears (every message, every 4th message, only at conversation start, etc.).

What it does that other prompts don't

The author's note solves a specific problem: instructions placed in the system prompt or character description lose effectiveness over long conversations. The model treats them as setup that happened "in the distant past" and weighs current messages more heavily. By the time you're 50 messages into a conversation, the original instruction to "write in a noir style" has been buried under accumulated context and may not be doing much.

An author's note injected near the end of context dodges this problem. Every time the model generates a response, your instruction is sitting right there, recent enough to influence behavior. It's the difference between a coach giving a pep talk before the game starts and a coach calling instructions from the sideline during the game.

The SillyTavern docs put it directly: "The closer the Author's Note is to the bottom of the prompt, the more impact it has on the next AI response."

This is why an author's note can fix problems that no amount of system-prompt rewriting will fix. You can spend hours refining a character card and still see drift after long conversations. Add a 30-word author's note at depth 4, and the drift often disappears.

What to put in an author's note

The most effective author's notes share a few characteristics. They're short. They're specific. They focus on what the AI should do rather than what it shouldn't.

Examples that work well:

For tone enforcement: "Write in a measured, observational tone. Vary sentence length. Avoid stacked parallel constructions."

For pacing: "Develop scenes slowly. Let conversations unfold. Don't rush to resolution."

For character consistency: "Mira speaks plainly without pet names or flourishes. She thinks before responding. Her replies are short."

For genre: "This is a slow-burn noir mystery. Maintain ambiguity. Reveal information sparingly."

For NSFW pacing: "Build tension through restraint. Don't escalate physical scenes faster than the emotional groundwork supports."

The note doesn't need to be elegant. It needs to be specific and short. A 30-word note is often more effective than a 150-word note, because the shorter version gives the model a clearer signal and doesn't dilute the instruction across multiple ideas.

What doesn't work as well:

Long, multi-purpose notes that try to handle several issues simultaneously. The model picks up the structure but applies it inconsistently.

Notes phrased as prohibitions ("Don't break character, don't get repetitive, don't use cliched phrases"). Models follow positive instructions more reliably than negative ones. As the SillyTavern documentation observes, "tell the AI how you want it to write instead" of telling it what to avoid.

Notes that contradict information already in the character card. The note will fight with the established setup and produce inconsistent results.

Using author's notes outside SillyTavern

Most consumer AI companion apps don't expose an author's note feature. The technique still works, you just have to inject it manually using prompt conventions the model has been trained to recognize.

The two most common conventions are square brackets and double parentheses. Both signal to the model that the bracketed text is meta-information rather than story content.

In Character AI, the convention is double parentheses. A message like ((Storyline reminder: We're in the cabin during a snowstorm. Mira has just discovered the broken window and the intruder is unaccounted for. Maintain tension.)) gets treated as an out-of-band instruction rather than character dialogue. The model uses it to reorient without it appearing in the narrative.

In platforms that don't have an established convention, square brackets often work because models trained on creative writing data have seen [Editor's note: ...] and [Author's note: ...] patterns regularly. Try [Author's note: maintain noir tone, slow pacing, and Mira's characteristic restraint] and see how the model handles it.

The frequency and depth controls in SillyTavern aren't replicable in consumer apps, but you can approximate the effect by injecting the note manually when you sense drift starting. After every 10-20 messages, drop in a fresh instruction. The cost is small and the impact on long-conversation quality is large.

Where to position the note for maximum effect

If you're using SillyTavern or another tool with depth controls, the conventional wisdom is depth 4. That places the note before the most recent 3-4 messages, which is recent enough to strongly influence the model's response without overwhelming the immediate exchange.

Depth 0 places the note at the very end of context, immediately before the model generates. This sometimes works but can feel intrusive: the model's response will lean heavily on the note's instructions and might not engage with the user's actual most recent message naturally.

Depth 1-3 puts the note among the most recent messages, which can confuse the model about what's character dialogue versus authorial direction.

Depth 5-10 still works but loses some of the "right at the front of attention" benefit. Useful for less critical instructions or for notes you want to influence without dominating.

Depth higher than 10 starts behaving more like a system prompt: persistently present but not dramatically more attended to than other context. At that point you might as well put it in the actual system prompt.

The best practice for most users is to start at depth 4 with frequency set to every 4 messages. This gives consistent influence without the note becoming overbearing.

When the technique doesn't help

Author's notes can't solve every problem. They work for tone, pacing, voice, genre, and structural guidance. They don't work as well for:

Memory issues. If the model genuinely doesn't have access to context it needs (because conversation history has aged out of the working window), an author's note can't substitute for the missing information. You need pinned memories or chat memory features for that.

Hard refusals. If the platform's safety filter is blocking content, an author's note won't override that. The filter operates at a different layer than the model's response generation.

Bad character cards. If the underlying character description is contradictory or thin, the author's note can patch some of the symptoms but won't fix the root cause. Better to revise the character card.

Wrong model. Some models are better suited to certain styles than others. An author's note can push a model toward a style, but it can't make a chatty playful model genuinely produce stark literary prose. The base model's training matters.

A workflow for long roleplays

For users running ongoing roleplays where character consistency matters across dozens or hundreds of messages, here's a workflow that compounds the benefit of author's notes:

Set a baseline author's note when you start the conversation, focused on tone and pacing. Keep it under 50 words.

Every time you notice drift starting (the AI getting repetitive, voice slipping, pacing rushing), don't fix it through correction in conversation. Update the author's note instead. The correction lasts longer.

For specific scenes that need a different tone (a tense confrontation, an intimate moment, a comedic interlude), temporarily replace the baseline note with a scene-specific one. Switch back when the scene resolves.

Periodically rewrite the note to be tighter. The version you start with is rarely the best version. After a few weeks of refinement, your author's notes become surgical: 20-30 words that consistently produce the kind of writing you want.

This pattern is widely used among experienced SillyTavern users and is one of the things that distinguishes a polished long-running roleplay from a conversation that gradually loses coherence.

Frequently asked

Is the author's note the same as the system prompt?

No. The system prompt is at the beginning of the context and represents long-term setup. The author's note is injected near the end and represents immediate influence. Both can be used together; they're complementary rather than substitutes.

Why is depth 4 the recommended position?

It's a sweet spot between being recent enough to influence behavior strongly and being far enough back that it doesn't override the user's actual current message. Depths 1-3 can confuse the model about which "voice" to attend to. Depths above 4 lose some of the recency advantage.

Can I use multiple author's notes at once?

In SillyTavern, yes, by using author's notes in combination with character-specific notes. In consumer apps, you can stack manual injections (one for tone, one for pacing) but you risk the model getting confused. Usually one well-written note works better than multiple competing ones.

How long should an author's note be?

Shorter is usually better. 20-50 words is a good range for most uses. Beyond about 100 words, the note starts diluting itself with too many ideas, and the model picks up the structure inconsistently.

Does this work on every model?

Mostly. Models trained on creative writing data (which is most of them) understand the bracket and parentheses conventions reliably. Models trained primarily on conversational data may handle author's notes less elegantly but still pick up on the prompt-injection pattern.

Can I use this technique for serious work, not just roleplay?

Yes. The technique works for any AI conversation where you want to influence the model's output without restating instructions. It's used by writers steering AI drafting tools, developers fine-tuning AI assistants, and anyone running long structured conversations where consistency matters.

Why isn't this feature in mainstream AI apps?

Because it requires understanding prompt structure, and consumer apps optimize for users who don't want to think about prompts. The functionality is hidden inside character cards and system prompts in most apps. Power users get the benefit by exposing the underlying tooling, which is what SillyTavern and similar interfaces do.