Model-Tips-and-Tricks

Greetings! This page intends to share some insights I've gathered from several years of experience testing Large Language Models (LLMs). The focus is primarily on optimizing LLMs for creative writing and roleplaying applications, though these tips may also be beneficial for technical tasks as well.

These suggestions may not apply universally to every model or use case, nor do they guarantee optimal results in every instance. However, they should help improve your results and overall experience with the model, even if marginally, and assist you in navigating common pitfalls or misleading advice. Some points may be familiar to experienced users, but I aim to include less commonly discussed details too. Please note that I am still learning and open to feedback and corrections.

1: Understanding Instruct Formatting

1-1: What is an Instruct Format?

The Instruct Format (also known as an Instruct Template) is arguably one of the most crucial aspects for ensuring an LLM functions correctly. It defines the specific structure and special tokens used to separate different parts of the input – such as system prompts, user messages, and AI responses – during the model's training and subsequent interactions.

Some formats, like ChatML or Alpaca, are widely adopted across various models. Others are specific to certain model families, such as Llama 3 Instruct or Mistral Instruct. However, it's important to note that not all models released under a specific brand necessarily use that brand's official format; fine-tuned versions might employ a different structure.

It's crucial to identify the correct format a model expects before using it. You can usually find this information in the model's documentation (e.g., the model card on platforms like Hugging Face). If the format isn't explicitly stated, you can often deduce it by examining the model's configuration files, specifically tokenizer_config.json and sometimes special_tokens_map.json, located within the model's directory.

For instance, finding tokens like <|im_start|> and <|im_end|> in these configuration files for a model based on Mistral architecture strongly suggests it was fine-tuned using the ChatML format, even if the base Mistral model uses a different native format. Familiarity with common format tokens helps in identifying the correct structure, especially when documentation is missing or unclear.

1-2: Should You Deviate from the Prescribed Format?

Adhering to the intended Instruct Format generally yields the best performance in terms of coherence, instruction following, and prose quality. However, some users experiment with deviating from the standard format, sometimes claiming it can reduce censorship or elicit different response styles.

Often, this deviation comes at the cost of reduced performance and reliability, making it a questionable trade-off. There are typically more effective methods to achieve desired outcomes, such as careful prompt engineering or strategically editing and continuing the AI's responses.

Based on my observations, fine-tuning an existing Instruct-tuned model (rather than a base model) using a different format than the one it was originally trained on (e.g., applying ChatML formatting during fine-tuning on a model originally released with Mistral Instruct) can lead to performance degradation. This results in less optimal responses compared to using a consistent format throughout the fine-tuning process.

This potential issue is separate from other challenges that can arise when fine-tuning already instruction-tuned models. However, if the fine-tuning does use the underlying model's original format, or if advanced alignment techniques like DPO (Direct Preference Optimization) or its variants (ORPO, etc.) are employed, the resulting model tends to perform more reliably.

1-3: Handling Models Listing Multiple Formats

Models might list multiple compatible formats due to techniques like model merging or specific fine-tuning choices where the developer intentionally trained across different structures. This is sometimes an intentional training strategy.

In such cases, you may need to experiment to determine which format yields the best results for your specific use case. While merging models or formats can sometimes produce unique capabilities, success often depends on careful prompting that aligns with the model's mixed training history.

Let me illustrate with a couple of examples:

Example 1: Nous-Capybara-limarpv3-34B and Hybrid Formatting

An older model, Nous-Capybara-limarpv3-34B, exemplifies this situation. It was based on a Vicuna-formatted model but had a LoRA (limarpv3) applied, which introduced a "Message Length Modifier" feature trained using the Alpaca format. This modifier allows users to suggest response length (e.g., short, medium, long) within the Assistant's prompt prefix.

The base Capybara model uses the Vicuna format:

System:
User:
Assistant:

The limarpv3 LoRA, however, used the Alpaca format for its training data, including the length modifier:

### Instruction:
### Input:
### Response: (length = short/medium/long/etc) 

(Note: This specific syntax within the response prefix is unique to this model's LoRA).

While the base formats differ, experimentation revealed a way to combine elements. Using the standard Vicuna format while simply adding the length modifier tag did not reliably activate the feature:

System:
User:
Assistant: (length = short/medium/long/etc) 

(This approach was generally ineffective).

However, by adopting the triple-hash style from Alpaca within the Vicuna structure, the length modifier became effective:

### System:
### User:
### Assistant: (length = short/medium/long/etc)

(This hybrid approach successfully influenced response length).

This demonstrates how elements from different formats used during training or merging might sometimes be combined effectively through careful experimentation.

Example 2: RP-Stew-v4 and Stop Tokens

Another case involves RP-Stew-v4, a model resulting from merging Vicuna and ChatML-based models. The ChatML format looks like this:

<|im_start|>system
System prompt<|im_end|>
<|im_start|>user
User prompt<|im_end|>
<|im_start|>assistant
Bot response<|im_end|>

Note that ChatML explicitly includes roles (system, user, assistant) within its tags and uses <|im_end|> as an end-of-turn token.

Standard Vicuna format doesn't use explicit end-of-turn tokens. However, for RP-Stew-v4, adding a similar token (<|end|>) after each role's content in a Vicuna-like structure proved beneficial:

SYSTEM: system prompt<|end|>
USER: user prompt<|end|>
ASSISTANT: assistant output<|end|>

This hybrid structure helped reduce rambling, repetitive outputs, and instances of the model incorrectly continuing the user's turn (speaking as the user).

Generally, sticking to a single, consistent format during model merging yields more predictable results, but these examples show that careful, informed mixing can sometimes be advantageous.

1-4: The Importance of Stop Tokens

In my experience, models often perform more reliably when their format includes explicit stop tokens (or end-of-turn tokens). As demonstrated with the RP-Stew-v4 example, adding appropriate stop tokens significantly reduced unwanted repetition (an estimated 25-33% reduction in rambling length was observed in that specific case).

Formats incorporating stop tokens generally provide clearer structural boundaries for the model. This helps it recognize the end of a conversational turn and maintain role distinction more consistently, leading to greater stability, especially in creative back-and-forth exchanges.

Conversely, omitting necessary stop tokens can sometimes lead to run-on responses, confusion between roles, or other unpredictable behavior. While experimentation is always possible, deviating significantly from the intended token structure often negatively impacts overall performance.

When experimenting with formats and tokens, it's wise to use separate configurations or instances to avoid errors. Incorrect placement (e.g., putting a stop token before the user's input within the user prefix) can lead to unexpected and incorrect model behavior.

2: Character and Chat Formatting

2-1: What is a Character Card?

Although somewhat obvious, let's define Character Cards first. Character Cards establish the persona for an AI bot to impersonate. This persona can represent a real person, a character from an established franchise, or an original character (OC).

Character Cards are typically structured like a profile or dossier, outlining the character's key attributes. Various formatting styles can be used to detail their unique traits, background, and mannerisms.

2-2: What are the Different Styles of Character Cards?

Creating effective Character Cards is not an exact science; the best approach can vary depending on the specific LLM (brand, size) and the desired outcome. However, several popular styles have emerged within the user community.

• Prose Style: One style involves writing the character description in natural prose, much like a narrative description in a book. This method requires careful writing to ensure clarity, avoid excessive keyword repetition, and maintain flow. It might be more challenging for beginners but potentially well-suited for users focused purely on narrative writing.

• List Formats: Another common approach uses list formats, where traits are categorized clearly. Variations exist, including:

  • Simple Markdown lists (- item or * item).
  • Wiki-style formatting (using templates or specific syntax like == Section ==).
  • Community-created structured formats (like JED+ or W++).

These list styles might use parentheses () or brackets [] to enclose information, dashes - or asterisks * for list items, or formatting like bolding (**text** or ### Heading) to structure sections.

Determining the 'best' format is difficult, as effectiveness can be model-dependent. However, based on my experience, highly structured but non-standard formats like W++ can sometimes lead to less consistent or coherent behavior (perhaps due to being less common in the LLM's training data). Similarly, directly pasting large amounts of raw text from a wiki ("wiki style") without curation can introduce irrelevant information ('bloat') and may not be optimally processed by the model.

My current recommendation is something akin to JED+, which often involves a hybrid approach: using lists for concrete details like appearance and personality traits, and employing standard prose for background history or defining speech patterns.

Regardless of style, it's essential to consider the writing perspective used within the card.

2-3: What Writing Perspectives Should I Consider?

Choosing the writing perspective for your Character Card is arguably more straightforward than selecting a specific formatting style, but it's equally crucial. The main options are:

• First-Person ("I"): The card is written from the character's own viewpoint (e.g., "I am...", "My background is...").

  • Pros: Can lead to bots with a more distinct, individualized voice, suitable for one-on-one chats. Reflects common online communication styles.
  • Cons: Requires careful crafting to accurately reflect the character's unique voice and mannerisms.
  • Keywords: I, me, my, mine.

• Second-Person ("You"): The card addresses the bot directly (e.g., "You are...", "Your personality includes...").

  • Pros: Often used in system prompts and AI training for direct instruction. Can be effective for defining behavior explicitly.
  • Cons: Less common in creative writing or typical chat data, potentially feeling less personal or natural for roleplaying. Can lead to awkward perspective mixing during interaction.
  • Keywords: You, your, yours, you're.

• Third-Person ("They"/"He"/"She"): The card describes the character objectively (e.g., "He is...", "Her history involves...").

  • Pros: Standard perspective for storytelling and literature, abundant in training data. Excellent for creative writing and often better for managing multiple characters in group chats due to clear distinction.
  • Cons: May feel slightly less direct for simple one-on-one interaction compared to First-Person.
  • Keywords: He/She/They, her/his/their, hers/his/theirs, it/its.

Even list-based cards implicitly adopt a perspective through pronoun usage and phrasing within the descriptions. Because LLMs predict text based on patterns, the chosen perspective significantly influences the likely style of the bot's responses.

• For personal one-on-one chats, First-person ("I") often works well.
• For narrative writing or multi-character scenarios, Third-person ("They") is generally preferred due to clarity and alignment with literary training data.
• Second-person ("You") is less common for character interaction and arguably the worse overall, but can be useful for direct instruction.

This principle extends to your own messages during the chat. Interacting consistently with the chosen perspective generally works best. Using a Second-person ("You") card often necessitates a mix of perspectives in the interaction (e.g., addressing the bot as "You" while narrating your actions in third person), which can sometimes confuse the model.

Another crucial formatting choice involves how dialogue and actions are represented in the chat.

2-4: Should I Use Quotes or Asterisks for Dialogue/Actions in Chat?

Next, consider how to format dialogue and actions within the chat interaction itself. The two primary styles are:

1. Quote Style: "Dialogue is enclosed in quotes," while actions are described in plain text outside the quotes.
2. Asterisk Style: Dialogue is written in plain text, *while actions are enclosed in asterisks.*

These styles are fundamentally distinct and likely draw upon different subsets of the LLM's training data due to the strong patterns they represent.

• Quote Style: This is the standard format in literature and published fiction. Consequently, using it often results in better prose quality and narrative coherence, especially if the character or archetype is common in literary works. Using quotes for dialogue is generally recommended for story-writing or more formal roleplaying.

• Asterisk Style: This style is comparatively niche, frequently seen in online roleplaying communities (though not universally) and resembling instant messaging or script-like conventions. It can be suitable if you prefer a style closer to text-based RP or casual chat.

Mixing the styles within a single message (e.g., "Quoted dialogue" *action in asterisks*) is strongly discouraged. This hybrid approach is generally absent from standard training datasets, offers no significant benefit, consumes extra tokens, and potentially confuses the model. Avoid using it unless you want to emphasize certain sections of the dialogue or text to make them distinct.

2-5: Recommended Combinations

Based on the points above, here are my general recommendations for combining perspective and chat style:

• For Creative Writing / Group Roleplay:

  • Use Third-person (They/She/He) perspective in the Character Card.
  • Use the Quote Style ("Dialogue", Action) for chat interactions.

• For Simpler One-on-One / Texting-Style Roleplay:

  • Use First-person (I) perspective in the Character Card.
  • Use the Asterisk Style (Dialogue, *Action*) for chat interactions.

3: Crafting a Character Card

3-1: Essential Resources for Character Creation

While creating a Character Card entirely from scratch, whether for an original character (OC) or one from an existing intellectual property (IP), can be a valuable writing exercise, it's not always the most practical approach. Several online resources can help streamline the process. Of course, it's best to decide on your preferred Character Card format (as discussed in Part 2) before gathering information.

• Wikis and Indexes: For existing characters or established archetypes, wikis are an obvious but crucial starting point, especially for beginners. Fan wikis (like those on Fandom) or encyclopedic sites (like Wikipedia) provide comprehensive information, helping you grasp the character's overall concept and potentially uncovering details you hadn't considered.
  • For existing characters, you can visit a wiki dedicated to their source material. Using an AI assistant to summarize key points from a relevant wiki page can provide a good base, but always review the summary and the original pages yourself to ensure essential details for your intended chat/RP experience are included.
  • For OCs or niche characters without dedicated wikis, consult broader resources. Is your character inspired by a known archetype (e.g., a masked vigilante)? Visiting wikis for similar established characters (like from Marvel or DC Fandom) can offer inspiration for finer details. Need ideas for an elf princess? Perhaps the Zelda Wiki could offer relevant concepts.
• Archetype/Trope Resources: More general resources can also be invaluable:
  • For specific mythical or fantasy beings (e.g., mermaids, harpies), specialized encyclopedias might exist (like the "Monster Girl Encyclopedia" wiki, focusing on specific interpretations).
  • For common character types found in media, sites like "TV Tropes" (for narrative patterns and character archetypes) or the "Dere Types Wiki" (for specific anime/manga personality archetypes) can provide ideas and terminology.
  • For characters with specific health or psychological conditions, consulting relevant, specialized resources could provide simple examples if wanted, like WebMD for defining complex traits.

The underlying principle is: if your character is based on an existing concept or archetype, there's likely a wiki or index online you can draw inspiration from. Beyond these general information sources, let's highlight two types of sites useful for specific card sections.

3-2: Resources for Describing Appearance

To detail your character's physical appearance, art databases with tagging systems can be very helpful.

• Image Tag Sites (e.g., Danbooru): Sites like Danbooru allow you to search for characters (yours, if they have art, or similar ones) and identify specific tags related to their build, clothing, hairstyle, accessories, etc. This is useful if you have a visual in mind but lack the precise terminology. Not every image is exhaustively tagged, so browsing multiple images might be necessary. While a wiki might describe appearance generally, tag sites can offer more specific visual keywords that LLMs might recognize (e.g., "blue sailor collar," "thighhighs," "long low-tied twintails").

3-3: Resources for Describing Personality

Defining personality can be complex, but certain resources offer frameworks and keywords.

• Personality Database (PDB): This website catalogs characters (real and fictional) according to various personality typology systems. While user-submitted typings aren't always perfectly accurate and the real-world validity of some systems is debated, PDB can be a useful tool for LLM character creation. Even if your character is an OC, browsing profiles of characters with similar perceived traits can provide inspiration and terminology.
• Common Typology Systems: Several systems listed on PDB are often recognizable by LLMs:
  • MBTI (Myers-Briggs Type Indicator): Uses a four-letter code (e.g., INTP, ESFJ) representing preferences across four dichotomies (Introversion vs. Extraversion, Intuition vs. Sensing, Thinking vs. Feeling, Perceiving vs. Judging).
  • Enneagram: Describes nine core personality types (Type 1 to 9), often supplemented by a "wing" (one of the adjacent types, e.g., 5w4, 9w1) indicating nuances.
  • Temperaments: Often refers to the Four Temperaments (Sanguine, Choleric, Melancholic, Phlegmatic), sometimes used in combination like the wing system.
  • Alignment: Originating from Dungeons & Dragons, classifies characters on two axes: Lawful vs. Chaotic and Good vs. Evil (resulting in types like Lawful Good, Chaotic Neutral, Neutral Evil).
  • Zodiac: The familiar astrological signs, often associated with a character's birthday, but sometimes used purely thematically.
• Other Personality Concepts:
  • Instinctual Variants (Enneagram Subtypes): May be too niche or require specific prompting.
  • Archetypes: Specific character roles or patterns (e.g., Tsundere, Mad Genius, Spoiled Princess; or Jungian archetypes like The Hero, The Mentor).
  • Tarot Associations: The symbolic meanings of Tarot cards are widely discussed online and likely present in training data.

Using These Resources: You don't necessarily need to become an expert in these systems. You can often use an AI assistant effectively by providing it with a character's profile from PDB (e.g., "INTP 5w4 Lawful Neutral") and asking it to summarize the associated traits or generate keywords. You can then incorporate these keywords or summary points into your Character Card's personality section. Combining the "compact" typology codes with the "complex" list of generated traits can help reinforce the desired personality without excessive token repetition.

3-4: Key Sections for Your Character Card

Consider including some of the following sections in your card for a well-rounded character definition:

• Backstory / Background: A crucial yet straightforward section. It provides context on the character's history, development, and potentially their current situation or point in their original timeline (if from an existing IP).
• Personality: Describes core traits, temperament, and demeanor (potentially drawing from the typology systems mentioned above).
• Appearance: Details physical features, clothing, and overall visual style (potentially using keywords from tag sites).
• Likes / Dislikes: Guides how the character reacts to specific things. Listing items separately works well. You can also use subcategories (e.g., Food, Activities, People, Concepts) or tiers (e.g., Loves: [item], Likes: [item], Tolerates: [item], Dislikes: [item], Hates: [item]) to add nuance.
• Goals / Fears: Strong drivers for character motivation and behavior. Defining primary and secondary goals/fears can add depth. These can represent aspirations to achieve or challenges to avoid or confront.
• Quirks / Habits: Unique mannerisms, routines, or distinctive behaviors that differentiate the character.
• Skills / Abilities / Stats: Defines what the character is capable of. Skills/Abilities apply broadly, while numerical Stats are often more suited for RPG or adventure settings but can sometimes be adapted for everyday contexts (e.g., Stamina: Low).
• Views / Beliefs: Covers the character's opinions on various topics, relationships with specific other characters (including the user), or stances on more significant themes like politics, religion, or morality. Helps refine personality in specific contexts.
• Speech / Mannerisms: Focuses specifically on communicative habits – tone of voice, common phrases, accent, verbal tics, etc. This works in conjunction with example messages.

3-5: The Value of Example Messages

Are example messages essential? While you might suffice with just the core description and a starting message for simple chats, adding detailed example messages is highly recommended.

If you want the bot to consistently adhere to a specific interaction style, voice, and formatting, examples are invaluable. Listing traits provides a foundation, but concrete examples demonstrating those traits in action – including your chosen chat format (Quotes vs. Asterisks), prose style, and typical response length – yield significantly more reliable results. They help the AI grasp the nuances of the character's expression.

• Prose Detail: Examples allow you to showcase descriptive language fitting the character (e.g., "her tail gave a thoughtful flick," "the piston in his arm whirred softly") or more subtle cues for grounded characters.
• Multiple Examples: Providing several examples (e.g., 3-7 messages) is often beneficial.
  1. It allows you to demonstrate the character reacting to different situations or expressing various relevant emotions without cramming everything into one potentially disjointed or overly long message.
  2. It allows you to vary example message length, which can encourage the bot to generate responses of varying lengths itself, rather than settling into a repetitive pattern.

Example messages bridge the gap between description and performance, significantly improving the consistency and believability of the AI's portrayal.

4: Interacting with the Bot: Messages and Conversations

4-1: Pre-Chat Preparation: Testing the Model's Knowledge

Before diving into creative interactions with your character bots, consider performing some preliminary tests on the underlying LLM itself. Why test the model rather than the specific character? Because different LLMs, even of the same size or family, possess varying levels of knowledge on specific subjects based on their training data. Understanding the model's baseline knowledge is crucial, especially for newcomers.

The Testing Process:

1. Use a Blank Slate: Set up a minimal chat configuration. Use simple names like "User" and "Assistant" (or "You" and "Me") for the participants and provide no character description or system prompt details beyond the basic Instruct Format structure.
2. Query Key Topics: Ask the model direct questions about subjects relevant to your intended creative work (e.g., specific lore from an IP, historical events, scientific concepts, character archetypes).
3. Minimize Randomness: To assess factual recall accurately, set the model's sampling temperature to 0 or extremely low (e.g., 0.01). This reduces the chance of the model hallucinating or inventing information.
4. Check Consistency: Even with low temperature, it's wise to generate ("swipe") a few responses to the same factual query to see if the information remains consistent. If testing creative generation (as discussed below), consistency across multiple swipes is also important.

A Note on Model Consistency and Marketing: (You can skip to the next section if preferred.) Related to testing, it's worth noting that model performance can vary. Sometimes, advertised capabilities or user-shared examples might not consistently reflect typical performance. Impressive results showcased online could be cherry-picked after many generation attempts ("swipes"). While benchmark scores offer some indication, real-world creative consistency is also vital. My preference is for models that demonstrate both creativity and reliability. It's advisable to test models yourself for your specific use cases rather than relying solely on promotional claims or anecdotal reports. Now, returning to the main topic...

4-2: What Specific Knowledge Should I Probe?

When testing the model's baseline knowledge, focus on areas critical to your creative goals:

• World/Character Knowledge: How familiar is the model with the specific IPs, characters, locations, or established archetypes you plan to use? Does it understand key historical or mythological contexts?
  • Example: Does the model know details about The Legend of Zelda series and general fantasy tropes? Test its understanding of specific character appearances, motivations, or key plot points relevant to your scenario. Look for inaccuracies: wrong details, incorrect actions, or nonsensical hallucinations.
• Archetypes and Concepts: Test knowledge of specific creature types (e.g., unique fantasy races, robot designs), cultural practices, historical periods, personality types, or abstract concepts you intend to feature.

Why Test This? The primary goal is context optimization. If the model already possesses accurate intrinsic knowledge about a character, setting, or concept, you may not need to explicitly detail it in your Character Card or world info prompts. This saves valuable context space and allows you to focus prompts on unique aspects or specific instructions. Conversely, identifying knowledge gaps tells you where you must provide explicit information.

How to Test:

• Direct Questions: Ask "What is [X]?" or "Describe [character/place/concept]."
• Scenario Prompts: Ask the model to write a short scene or dialogue snippet incorporating the elements you're testing. This reveals not only if it knows the information but also if it can apply it correctly in a narrative context.
• Template Generation: Ask the model to generate a basic character template or profile for a specific character or archetype. Compare its output against known facts to see what it gets right or wrong. Pay attention to how it formats the information as well.

4-3: Observing the Model's Default Formatting

When interacting with a blank bot (no character info, minimal system prompt), pay attention to its inherent formatting preferences:

• Dialogue/Action Style: Does it default to using quotes ("Dialogue") or asterisks (*Action*)?
• Dashes: Does it use hyphens (-) or em dashes (—) for pauses or parentheticals?
• List Formatting: If asked to create a profile or list, what structure does it use (bullet points, bolding, specific syntax)?

Understanding a model's natural tendencies can help you work with its patterns rather than against them, potentially leading to smoother interactions. For example, if a model consistently uses em dashes, adopting them yourself might encourage more consistent output. This is particularly relevant if you plan to primarily use one specific model or family of models with similar training.

4-4: Managing Formatting Dynamics Within the Chat

Formatting consistency is crucial not just at the start but throughout an ongoing conversation or roleplay. Be mindful of these common issues:

• Message Length Inertia: Models can sometimes fall into a pattern based on recent message lengths. If your recent messages and the bot's replies have all been short, it might struggle to generate longer, more detailed responses when needed (and vice-versa).
  • Solution: Intentionally vary the length and complexity of your own messages from the beginning of the chat. Don't let every turn be roughly the same length.
• Repetitive Sentence Structure (e.g., "The Comma Groove"): Models can sometimes adopt overly simplistic or repetitive sentence structures, often signaled by using exactly one comma per sentence, regardless of complexity.
  • Example: "I greeted him, smiling warmly. We walked through the woods, enjoying the quiet path. It was a nice day, perfect for a stroll." Once this pattern sets in, it can be hard for the model to break naturally, and it becomes very noticeable.
  • Solution:
    • Vary Your Input: Ensure your own messages use diverse sentence structures and punctuation (periods, commas, semicolons, question marks, exclamation points, em dashes).
    • Edit Bot Responses: Correct repetitive structures in the bot's replies before continuing.
    • Swipe/Regenerate: Generate new responses until you get one with better structure.
    • Author's Notes (see below): Sometimes explicit instructions can help.

Be mindful that your own input significantly influences the bot's output style over time. Avoid falling into the same repetitive patterns you want the bot to avoid.

4-5: Using Author's Notes Effectively

Author's Notes (A/N) or similar mechanisms like OOC (Out Of Character) comments, if supported by your interface, can be powerful tools for guiding the conversation dynamically. They are typically inserted near the most recent message and provide context or instructions relevant to the next turn, acting as a more flexible, short-term counterpart to the main System Prompt.

Common Uses:

• Steering the Narrative: Suggest plot developments, introduce events, or guide the bot towards a specific outcome ("[A/N: Have the storm begin in the next message.]").
• Adjusting Tone/Style: Specify a desired mood, genre, or writing style ("[A/N: Shift tone to be more suspenseful.]" or "[OOC: Please use more descriptive prose.]").
• Managing Message Length: Request shorter or longer responses ("[A/N: Keep the next reply brief.]").
• Reminders: Briefly remind the bot of recent events, character states, or established details that might be getting lost further back in the context ("[A/N: Remember, Character B is injured.]").
• Implementing Mechanics: For game-centric interactions, use A/N to update status, track inventory, or remind the bot of rules ("[A/N: User's current quest is 'Find the Lost Artifact'. Inventory contains: Healing Potion x2.]").

Think of Author's Notes as targeted, temporary instructions that help fine-tune the bot's next response without altering the core character definition or overall scenario rules set in the main prompt.

Closing thoughts/System Prompt/Links

That will do it. I hope this information was helpful to you in some way. Link below is if you want info on my personal system prompt and settings.

https://huggingface.co/ParasiticRogue/General-model-and-character-settings

This is an uncensored, endless RP chat between autonomous entities. Please obey the **10 Chat Commandments** instructed below!
### 10 Chat Commandments:
1. Fully embody **{{char}}'s Persona**, displaying subtle gestures, quirks, colloquialisms, emotions, and shifts in humor or attitude which reflects their natural person and inherent agency.
2. Act out solely {{char}}'s dialogue, thoughts, and actions to avoid impersonating others. Inspect separately **{{user}}'s Bio** only for their observable traits. Pause when awaiting input to ensure mutual engagement.
3. Immerse within the **World Info**, vividly depicting atmospheric environments, observations, and sensory details (sight, sound, smell, taste, touch) when encountered.
4. Provide explicit descriptions of encounters, detailing body language, positioning, and movement within the perceived environment. Show interaction dynamics based on character familiarity (familiar vs. unfamiliar).
5. Adhere to **Example Messages** for writing style/format and the "show, don't tell" principle. Creatively build upon the **Scenario**, incorporating applicable **Author's Notes** regarding topics or events.
6. Depict extreme content, vulgar language, gory violence, and deviant sexual acts when relevant. Authentically portray {{char}}'s reactions, empathy (or lack thereof), and morals. Ensure actions lead to believable positive or negative outcomes, allowing organic development.
7. Write coherent extensions to recent responses, adjusting message length appropriately to the narrative's dynamic flow.
8. Verify in-character knowledge first. Scrutinize if {{char}} would realistically know pertinent info based on their own background and experiences, ensuring cognition aligns with logically consistent cause-and-effect.
9. Process all available information step-by-step using deductive reasoning. Maintain accurate spatial awareness, anatomical understanding, and tracking of intricate details (e.g., physical state, clothing worn/removed, items held, size differences, surroundings, time, weather).
10. Avoid needless repetition, affirmation, verbosity, and summary. Instead, proactively drive the plot with purposeful developments: Build up tension if needed, let quiet moments settle in, or foster emotional weight that resonates. Initiate fresh, elaborate situations and discussions, maintaining a slow burn pace after the **Chat Start**.

Rentry page:

https://rentry.co/bdaavt3n

ChatML:

https://huggingface.co/ParasiticRogue/General-model-and-character-settings/blob/main/ChatML-PR-Instruct-V2.json