Response Style & Length
A page exploring Response Styles & Length in more detail.
Response Styles
The Response Styles are Xoul.AI's various available system prompts which will act as the instructions guiding the way the model responds within the chat.
These prompts are both powerful, and minimally intrusive. They cover all the basics.
Custom / System Prompt
In the Response Style option menu is the Custom option, which opens a new text field. This is located either inside Xouls or in the chat right side settings panel.
The System Prompt field located near the bottom of the Scenario Advanced Options works identically.
This is a power user tool that completely override's Xoul.AI's provided system prompts.
Use Responsibly!
Using the Custom override is a double-edged sword.
When you remove the provided system prompt, you are not just deleting one conflicting rule, you are removing the entire, well-tested framework for the model's behavior. You cannot simply insert a single instruction like "write in first person present tense" into an otherwise empty prompt and expect good results. The model will be left with no other guidance, leading to unpredictable and often poor performance.
Putting a simple generic prompt into the custom field is like deciding your living room needs a throw rug and instead of simply adding a throw rug to the existing room, you throw all of the furniture out and put only a throw rug into the center of a now barren room.
Custom Prompt Stack
Custom prompts can be added at different levels. You can include a custom prompt in an individual Xoul, in a Scenario, even in a chat, and you can have a group chat with up to 8 Xouls that might all be using custom prompts- so which custom prompt will be used?
It's on the stack:
A Custom Prompt found on any layer will override everything beneath it.
Top of Stack: Within the chat.
Middle of Stack: Within the Scenario.
Bottom of Stack: Within the Xoul.
Foundation: Provided by Xoul.AI as default.
What about Group Chats?
When the model replies, either you or the system selects which Xoul is responding. In this case, each Xoul uses its own selected Response Style or Custom Prompt, and all 8 Xouls can have completely different prompts in a single chat!
You can set all Xouls to a single Response Style through the chat interface however any Xouls that have a Custom Prompt will still use their prompt.
Getting Started with Custom Prompts
If you're new to writing prompts, here's some introductory advice. While not all system prompts will follow all of these rules, or might be very uniquely written, this is a good place to start.
Reminder: It's called 'Prompt Engineering' for a reason.
From the people who develop AI models to power-users can struggle to get their prompts working perfectly. Small tweaks to the system can cause a once working prompt to stop working as well. One model will take a system prompt perfectly; another will completely fail to understand the prompt.
Familiarity with how LLMs think and work, deep familiarity with how Xoul.AI's LLMs think and a willingness to reiterate and tweak dozens of times before it finally starts working correctly is often necessary to create a well working system prompt.
Do not be surprised if your prompt creates more issues than it resolves until you finally get it perfect.
If creating a working system prompt harms your experience or makes you frustrated, consider collaborating with other users but be prepared to take criticism.
A Custom Prompt should contain the following:
An overarching goal or purpose.
e.g. "You are a narrative engine", "You are collaborative roleplay partner exploring a story between
xoulanduser", "You are a helpful AI Assistant".
Permission to write content that may be considered uncomfortable, unethical or inappropriate (as necessary).
Without this permission the AI model may reject or avoid explicit topics, violence, or avoid discussing certain subjects (even broadly).
General Writing:
Clauses for which character(s) the model writes for, and which it doesn't.
Optionally, instructions for how to handle characters in certain ways.
Clauses for what the focus should be on.
e.g. Instruct it to describe settings, environments, etc. for more story focused prompts, or stick to conversation style replies for text-like chats.
How to write specific types of scenes (such as action, violence or explicit content).
What sort of tone, pacing, or writing style it should use if you want a specific flavor.
Formatting & Structure:
What tense and perspective replies should be in, whether it should wrap narration in asterisks or reserve them for emphasis, if it should use bold, etc.
How long replies should be, the number of paragraphs, the mix of dialogue, etc.
Any other clauses as necessary, based on the goal of the prompt.
Different models will react to these instructions in different ways. Be sure to check how well your prompt performs on the model you intend to use it with.
Check the Creator's Guide (once published) for more information about making different types of Custom prompts and don't be afraid to check out the Xoul.AI Discord server where other creators take prompt requests, share their prompts, and collaborate on prompt creation!
Length
There are four available response lengths you can set in a chat. Length will generally encourage the AI models to stay in a general length, but the length of the Greeting, Chat Samples, previous replies in the chat, and your replies will have a strong influence on how long the model's responses end up being.
Set your desired length early, take note of the Greeting length (most relevant at the start of a chat) and be sure to edit replies as necessary to make sure the model adopts and sticks to the desired length.
Dynamic: A flexible option that allows the model to disregard all patterns of response length and structure. This gives the model the most freedom to generate engaging, well-paced replies. Replies can be short, medium or long, will grow or shrink as necessary to develop the story, and you're unlikely to see strict patterns for how long sentences or paragraphs will be. Highly recommended!
Long: Instructions telling the model to use at least 3-4 paragraphs.
Medium: Instructions telling the model to stick to 2-3 paragraphs.
Short: Instructions telling the model to only use a single paragraph.
Last updated