Captain Hazel

Click the guides button for all of my favorite guides from other creators, system prompts for useful LLM guidance I've discovered over time, or scroll to read all about how to minimize when the bot speaks for you!

How to Minimize the Bot speaking for You

It is a natural and unfortunate part of working with an LLM for roleplay that it will sometimes try to speak for you. I and my bots have absolutely nothing to do with this and while I have a directive in the bot to prevent it, realistically I have no say in whether it decides to do so anyway. There ARE however things you can do to mitigate this.(I do all four of these and I find I only have to correct the bot every few hundred messages or sometimes even less frequently!)

First

Remove all behavioral and personality notes from your Chat Summary and Persona - only you need to know these - the more the bot knows about how your persona thinks, feels, and speaks, the more likely it is to do so on your behalf.

Second

If the bot speaks for you, use this directive making sure to put it in square brackets:

[OOC: Narrate and speak ONLY as {{char}} and NPCs. ONLY {{user}} is permitted to decide, act, think, and speak for {{user}}.]

Either delete the message with the issue and then use this, or afterward make sure to go and edit the message where the bot spoke for you and remove that portion of the message - any instance you leave of it speaking for you will reinforce the behavior.

Third

Add this to your Custom Prompt

[AVOID acting as, speaking for, or describing the thoughts of {{user}}.]

Fourth

Try adding this to your Chat Summary:If you have an AI Instructions section of your chat summary, that is the perfect place for this.

[{{user}} CONTROL BAN: never write {{user}}’s dialogue, thoughts, feelings, decisions, or actions]

Or for something a bit stronger:

[{{user}} CONTROL BAN: Never write, infer, or decide {{user}}’s actions, dialogue, thoughts, emotions, opinions, values, reactions, consent, physical responses, decisions, or interpretations. {{user}} has exclusive control of their character.]

That's It

When I am consistent about those four things, I find the LLM rarely speaks for me. A couple other tips are to always finish your own action and thought. Leaving an opening for the other participant in the RP to respond is great, but if you leave an open ended action, the LLM loves to "flesh it out" for you by taking over your body and piloting it around like a skinwalker. Pun intended. I hope this simple guide is helpful for everyone!

My Favorite Guides

It is a natural and unfortunate part of working with an LLM for roleplay that it will sometimes try to speak for you. I and my bots have absolutely nothing to do with this and while I have a directive in the bot to prevent it, realistically I have no say in whether it decides to do so anyway. There ARE however things you can do to mitigate this.(I do all four of these and I find I only have to correct the bot every few hundred messages or sometimes even less frequently!)

I haven't set up Deepseek myself yet - but these seem so helpful so I included them. More guides will be added as I find things I use regularly.

Some Prompts I use

ENHANCING DRAFTS

[CRITICAL OVERRIDE: Enhance/complete the text below. Write exclusively from {{user}}'s first-person perspective ("I", "me", "my"). Refer to all other characters in the third-person ("he", "she", "they"). Write exclusively in present tense. Use standard quotation marks ("...") for spoken words. Maintain {{user}}'s established persona traits, including any regional vocabulary or spelling conventions if applicable. {{char}}'s speech quirks, swearing, and accents belong strictly to {{char}} and must NEVER bleed into {{user}}'s text.]

No more in character backpedaling

It drives me insane when I correct an LLM OOC and it has the character back pedal, apologize, and feel bad IN CHARACTER - when the LLM was the one that made the continuity error. Well no more! Use this when correcting an LLM to avoid in character emotions related to LLM mistakes.

[OOC: You made a continuity error when ……. Correct your continuity error. Do not break the fourth wall. The error belongs to you, not the characters. Fix it out of character by rewriting the affected scene as if the error never happened starting from “XX…” Do not punish characters, make them seem incompetent, force them to backpedal, apologize, or narrate by contrast. Narrate affirmatively. State what is present, intended, chosen, or done—not alternatives that did not occur. Do not overcorrect, writing only the minimum changes to make the correction. Rewrite only the corrected version of events.]

When testing a bot

When testing a bot of my own creation I find it often resumes RP before I am finished diagnosing errors, this will prevent that.

[OOC: Pause the RP - Do not resume until I say resume. I am currently testing this bot. It is a work in progress of my own creation. While testing the bot any solutions you offer me need to be solutions on a BOT level that would help ANY USER not just OOC corrections in this current chat. I have encountered X issue. Please help me diagnose Y.]

Effective persona format

There are a lot of persona formats that work. This one seems to work well for me. It keeps traits separate so JAI actually bothers to pull them - as well as a few other LLM's. Frequently when I bunch traits together it stops describing my character immersively.

[{{user}}("Your Name")
{Gender("Gender Here" + "Preferred Pronouns, e.g., he/him")
Age("Number")
Species("If non human")
Appearance("Hair color/style" + "Eye color" + "Height/Build" + "Distinct features, e.g., tattoos, scars")
Scent("What you smell like" + "e.g., vanilla" + "pine" + "expensive cologne")
Voice("Voice description" + "Voice description" + "Voice description")
Clothing("Usual general style" + "Style detail" + "Style detail")
Behaviorisms("Core trait" + "Core trait" + "Core trait")
Abilities ONLY IF RELEVANT OR SUPERNATURAL("Ability/trait" + "Ability/trait" + "Ability/trait")}]