Xoul Creation Guide


Foreword

FAQ and Quick Reference The LLM = The AI = The model = The Large Language Model

LLM stands for Large Language Model.
This is a type of artificial intelligence that generates natural language.

LLMs are machine learning intelligence which was trained by being shown a lot of text like novels, scripts, articles and things scrapped off of the internet. By analyzing mountains of human-written text the AI was able to develop a grasp of the structure of language which allows it to generate natural sounding language.

Quick Reference:

This stuff will (optionally) be explained in more detail in the second part.

  • Tokens: All text you send to an LLM gets broken down into tokens. Tokens are words, word particles, punctuation marks, symbols, numbers and anything else that appears in the text you send.
    • Dave's Personality=Friendly, outgoing and talkative. This is 52 characters in total and it breaks down as...
    • Dave 's Person ality = Friendly , out going and talk ative . 13 tokens in total.
  • Permanent: What is written here is always shown to the AI, this information never changes.
    • Name
    • Gender
    • Age
    • Personality
    • Description
    • Advanced Definition
    • Chat Samples
    • Persona Name
    • Persona Description
    • Chat Memory
  • Temporary: What is written here changes in some way, meaning what is first written here isn't what the AI will always see.
    • All chat replies including the Greeting
    • The Default Scenario (Auto summarizes as the chat develops)
  • Context Range: How many tokens the AI can see at once.
    • Xoul sees 12k tokens for free users and up to 16k tokens for the highest subscription.
    • After the tokens of the Xoul's information, Scenario (if used), Persona information and Chat Memory is "used" from the context range the remaining amount is used to read as many chat replies as it can fit within the remaining tokens of the context range.
    • (This is why we have character limits. More tokens used in one place is less tokens to use in another. A larger Persona = less memory in the chat.)

Xoul/"Bot"/Character Card

When you chat on Xoul.AI the conversation is always between the user (you) and the AI (the LLM) and the Xouls you can select are simply a collection of information that sets the behavior of the AI during that chat with it. These are typically called "Character Cards" but a lot of people also call them a "bot". On Xoul.AI this is simply called a "Xoul". While most people use these character cards/Xouls to create a single character the AI roleplays as don't be afraid to think outside of the box and be creative about what you instruct the AI to with your Xoul and consider making things like:

  • Multiple characters (Stardew Valley Bachelors, Jack & Jill, etc.)
  • A concept (the AI pretends to be a chair, the AI compares which characters are stronger, etc.)
  • A utility (defines words, helps you write replies, helps you describe settings, etc.)
  • A “premise” (Hogwarts, a topless car wash, a zombie apocalypse, etc.)
  • A narrator/dungeon master/rpg master
  • Anything else you can imagine

Front End

Alt Tag

  • SillyTavern is a front end you can download on your desktop. You must provide an LLM to run through SillyTavern to make any use of this front end.
  • Most website or app front ends come with "their" default LLM that they're granting you access to but also include the option to use a different, third party LLM if you would like.
    • JanitorAI grants exclusive access to the Janitor LLM but allows you to use proxies to third party LLMs.
    • ChubAI grants exclusive access to multiple limited use LLMs but also allows you to use proxies to third party LLMs.
  • Other websites or app front ends come with "their" default LLM and do not allow you to use proxies to third party LLMs through their front end.
    • Xoul.AI, Gemini, ChatGPT, CharacterAI, FiggsAI (defunct), etc.

{{char}} & {{user}}

{{char}} and {{user}} are small pieces of code that the front end uses as a placeholder for what is written in either the Xoul or Persona's Name Panel. These macros simply replace the exact characters {-{-c-h-a-r-}-} with exactly what is printed in the Xoul's Name Panel.

Alt Tag

Alt Tag Whatever is written in the Name Panel of the Xoul becomes what is written in the place of the macro {{char}}. The same is true of {{user}} except in that case it's printing the Persona's Name Panel in place of the macro.

These are the macros for Xoul:

  • {{char}} / {{xoul}} / xoul = Exactly what is printed inside of the Xoul Name Panel.
    • These are fully equivalent. Same thing, just different ways of spelling it and xoul is four less characters, making it the best of the three to use if you need to save characters.
  • {{user}} / user = Exactly what is printed inside the Persona Name Panel.
    • These are also fully equivalent and again, user is four less characters making it the best to use to save characters.

Macros happen on the front end. The conversion happens prior to the information being sent to the LLM.

{{char}} and {{user}} are standardized macros that MOST BUT NOT ALL front ends use to call upon the Name Panel of either the character card or the user card. Outside of only those two macros you cannot take a macro from one front end and use it on another front end.

DO NOT use Macros from different front ends on Xoul.AI. They DO NOT work.

END_OF_DIALOG, <START>, {{narrator}}, {{scenario}}, <BOT>, etc. None of this means anything to Xoul.AI.

Alt Tag "{{time}}" is a macro from SillyTavern. Only on SillyTavern this would have been converted into the current system time (yes its 4:04AM, don't judge me) but because Xoul.AI wasn't programmed to handle this macro it simply doesn't function and instead prints it as plain text. When I go ahead and ask the LLM what that plain text means the LLM does what LLMs do, it tries to "interpret" what that plain text is attempting to communicate and because it believes we're roleplaying and my Persona contains information about him attending college it roleplays that it's time for my character to go to school.
Alt Tag Alt Tag

This also means you should not be throwing {{ and }} around words, phrases or names inside of your Xoul. {{David Smith}} is just as meaningless and pointless as %%David Smith%% or 55David Smith55.

General Advice for Macro Use

The {{char}} / {{xoul}} / xoul and {{user}} / user macros should be used to either create flexibility or save characters. You don't need to use it outside of those two purposes.

  • Saving Characters: If (and only if) your Xoul has a traditional name (e.g. Dave, Alexander, Harper, Sarah) then using the character macro can potentially save characters. Alexander, Harper and Sarah are all more characters than simply xoul.
  • Flexibility: By utilizing the character macro in place of writing out your character's traditional name in a prompt it makes the prompt flexible enough to be copy-pasted directly into different Xouls without requiring any editing. _____ writes in third person past tense. requires you go in and manually change _____ out for the name of the character every time you use it. Putting a macro in there instead such as: xoul writes in third person past tense. means you can copy and paste that prompt into any Xoul without needing to edit the prompt at all.

Pay Attention to Phrasing with Macros

If the Persona has a traditional name (e.g. Dave) then saying the user = the Dave. This is awkward phrasing.

You also need to be careful about "awkward phrasing" when you're using the Name Panel as a title.

You Type AI Sees (Title) AI Sees (Name)
xoul's name is Alexandria Smith. Mean College Professor | She's Hot |'s name is Alexandria Smith. Alexandria's name is Alexandria Smith.
xoul is a middle aged woman. Mean College Professor | She's Hot | is a middle aged woman. Alexandria is a middle aged woman.
xoul is a college professor. Mean College Professor | She's Hot | is a college professor. Alexandria is a college professor.

When you use a title name like a traditional name you encourage the AI to do that as well resulting in the AI writing something like this in the chat:

"My name is Alexandria Smith," Mean College Professor | She's Hot | said with a frown.

Consider what the AI is going to see after the macro conversion happens and make sure the phrasing isn't weird, confusing, or encourages behavior you don't want to see. Again, you don't need to use these macros at all, they exist to help you, not the AI and so the use of them isn't actually making your Xoul's information any more clear or effective.



Part One: Xoul Creation

There is no one "right" way to make a Xoul. Provided you get results that you are happy with you used the "right method" no matter what that method is but the AI is a machine so simply making a Xoul that the AI fully "understands" is very simple and straightforward as long as you keep one thing in mind at all times:

Clarity is King

This is a very simple idea of making sure that any detail you think is important is communicated as clearly as possible to the AI. The simplest way of doing this is to make sure in any sentence you write telling the AI something about a character you clearly state who that information is about.

  • Unclear: He loves hot chocolate.
    • Who likes hot chocolate?
  • Clear: David loves hot chocolate.
    • David is the person who likes the hot chocolate.
  • Unclear: David goes to the gym on his days off. It is his favorite place to spend an evening.
    • The second sentence, despite immediately following the relevant information is unclear. What is that favorite place? Who's favorite is it?
  • Clear: David goes to the gym on his days off. The gym is David's favorite place to spend an evening.
    • Now the second sentence is clear. It's David's favorite place and that place is the gym.

When you say he in regards to a detail of some sort any character in a chat or roleplay can be he. This applies to group chats and even one on one chats between a single Xoul and the Persona. The AI sees all the information of all Xouls and the Persona together and so if one he is defined as having a specific hair color, eye color, favorite food, or anything else that information is only assigned to this vague idea of he. When you explicitly state who exactly the information is assigned to such as David, Jane, etc. then this information can only ever be assigned to a character with that name, which stops any and all confusion about who the details are assigned to.

Even better when the AI is absolutely certain that a detail is tied to this and only this character the AI is more likely to actually bring that information up during the chat. Basically, it makes your characters more firmly stay in character, makes them more likely to bring up those small details that bring enrichment to the character and just overall leads to a more reliable character. It's foolproof and its a really good foundation for anyone who is somewhat new to using LLMs to get into the habit of doing.

You can seen an entire Xoul written this way here.

Write the Persona Clearly as well!

Your Persona will benefit from also having this Clarity is King concept in mind. Make sure everything you make is foolproof and you'll minimize both frustrations you encounter and frustrations other users encounter.

How should I write my Xoul?

However you want! No, really, anything goes but if you want some guidance here are some general methods that you'll typically see.
Written Instructions:
David is a middle aged man at the age of 38 years old. David is tall for a man at 6'2" and has broad shoulders and a muscular frame. David wears business casual clothes, favoring black or tan slacks and short sleeve button up shirts in solid, muted colors. David has short brown hair with thick bangs he slicks back over his head and deep set, bright blue eyes. David is an accountant living in downtown, Seattle in a small, one bedroom apartment.

  • Pros
    • Clear (when attention is paid to making the information clear)
    • Straight forward and natural for users
    • Flexible
    • Probably the only effective way to write a Xoul that isn't a traditional character
    • Works fine for multiple character Xouls
  • Cons
    • Plainly listed information results in the AI using the "default" narration it likes to use
    • Can be difficult to "organize" and remember all the relevant details you need to write in

Listed Information:

1
2
3
4
5
6
David's Name=David Anthony Smith
David's Age=Middle aged man at 38 
David's Features=Tall for a man at 6'2", broad shoulders, muscular frame, short brown hair, slick back thick bangs, deep set blue eyes
David's Clothing=Business casual, black or tan slacks, short sleeve button up shirts in solid muted colors
David's Home=One bedroom apartment in downtown Seattle
David's Career=Accountant
  • Pros
    • Clear (when attention is paid to making the information clear)
    • Straight forward and organized
    • Flexible
    • Slightly more efficient on character limits and tokens
    • Great for Xouls that are either one or just a few characters
  • Cons
    • Plainly listed information results in the AI using the "default" narration it likes to use
    • Encourages users to write things in ways that aren't as clear as they could be
    • Difficult/impossible to use for Xouls that aren't traditional characters

Conceptually Written:
David, a 38-year-old accountant, navigates the monotony of his Seattle one-bedroom apartment. Days blend into a predictable rhythm of spreadsheets and takeout, leaving David yearning for something more. A lingering sense of unfulfilled potential and a growing discontent with his routine threaten to derail David's carefully constructed life.

  • Pros
    • The narrative style and flair of how this is written will be reflected in how the AI narrates for the Xoul
    • If prompts are used to back this writing style up the style will be even more effective
    • The AI is forced to interpret things which can often result in characters with more flexibility, nuance and oomph
    • It doesn't have to be prose. You can dump something that reads like song lyrics into the Xoul or something like a poem or anything else you want, it's fully interpretive and formless and makes for interesting results
  • Cons
    • High potential to be unclear
    • Harder to get important small details into the Xoul and make sure they're very clear
    • Harder on the character limits and tokens, meaning you write more but less, overall, is clearly communicated to the AI

Mixed:
Mixing and matching any of the previous styles together can bring the benefits of each style together into an extremely well written Xoul. You can start with listed information to make sure the character is perfectly clear and then mix it with something much more conceptually written to bring more flair and uniqueness to the Xoul, getting the benefits of both styles at once.

If you're just starting out try to play it safe and use one of the more simple, clear methods and once you're a little more familiar try experimenting and branching out by trying different things. Remember, as long as you get results that you are happy with, you're using the right method.


Basic


Name Panel

Alt Tag

  • Can include letters, numbers and/or symbols.
  • Must be between 1 and 50 characters in length.
  • What the Xoul is called serves to inform the AI as to who OR what the AI is meant to be doing within that Xoul.
    • Testbot will be assumed to be a robot character.
    • David is a man (because that is a man's name).
    • Mean College Professor | She's Hot | is a mean college professor.
    • Dictionary Bot is the AI acting like a dictionary for the user.

Gender

Alt Tag

This influences the pronouns the AI uses for the character and will influence which voice options you will be presented with later.

Alt Tag

  • Male
  • Female
  • Non-Binary
  • Trans Woman
  • Trans Man
  • Prefer not to say
  • Or you can simply type in a custom gender.

Age

Alt Tag This must be a numerical value between 18 and 100000.

NO MINORS

Any characters within a Xoul or Scenario cannot be minors. This includes "well technically the character..." type characters as well. If the character smacks as being a minor its going to get reported. Characters cannot visually (the icon) look like a minor and cannot have any sort of "well it just looks like a child but actually is 1000 years old" sort of angle to it. Do not make Xouls of minors. Your Xouls will get reported and you can get banned for doing this.

Outside of Xouls of minors you also cannot make Xouls that are discriminatory or hateful against a group of people. Things like a Xoul specifically designed to be nothing but offensive towards a group of people is going to get you into trouble. Xouls can also be reported for being low quality/spam, advertisements or using someone's likeness without their consent.



Personality


Personality Sliders (Optional)

Alt Tag

The Devs will be removing this soon and combining the Description and Personality. For now I recommend just skipping it.

This one seems to confuse a lot of users. The Personality Sliders are a quick way to slap together a personality for a character Xoul and the adjustments made (or not made) will actually be printed inside of the real Personality Panel. Once you finish setting up the Xoul you will never see the Sliders again and will have to manually write in changes to the personality within the real Personality Panel.

  • Start with four randomly generated traits.
    Alt Tag
  • You can add, remove or customize traits by hitting "Add Custom".
    • Add Custom will pop up a new slider for you to customize.
    • You can have between 1 and 6 personality sliders.
    • Traits can be adjusted by up to two ticks in either direction.
  • Hitting the dice icon will randomize both the traits listed and the sliders position.

Alt Tag

  • Adjustments made to sliders determines what is printed in the actual Personality Panel (See Gif)
  • Ignored traits are completely ignored if at least one other trait is adjusted.
    • You won't end up with Very Shy, Professional, None, None
    • You just get Very Shy, Professional
Personality Panel

To gain access to the Personality Panel you must finish the Xoul and then navigate to the Edit button on the top right corner of the Xoul's page. Alt Tag

And then scroll down to this panel. Alt Tag

Users can see the Personality details of any Xoul they view.

Everything written in this panel is permanently remembered.

Here is where you can write out information about the personality of any characters defined within the Xoul. On the simplest level you can simply list traits similar to how the Personality Sliders list them but remember Clarity is King.

Example: When, Where, Why & How

Dave is slow to anger because he doesn’t believe most things are worth the energy to get angry about. In romantic settings Dave can be very flirty and confident. In work settings Dave is very serious and meticulous. When Dave feels vulnerable he becomes snappy, standoffish and gruff. Dave becomes bashful and modest when complimented about his professional accomplishments.

You can bring a lot of clarity, emphasis and nuance to details and traits by giving them the additional context of When, Where, Why & How these details or traits are important or are expressed.

Stereotypes: For Better and For Worse

LLMs love stereotypes. This means if you want to make a generic tsundere character or a generic asshole jock you really just need to put David is tsundere or David is an asshole jock into the Personality section and the LLM will bring a pretty fully fleshed tsundere or asshole jock personality to the table.

However, this is a double-edged sword because the LLM is going to give you really generic, stereotypical depictions of traits you list in the character. A shy character is just going to be generically shy with very little nuance to it and often times if your character has a list of more complex and nuanced traits as well as more stereotypical basic traits your character is going end up primarily being just the most stereotypical ones with the other traits sort of left ignored.

Consider avoiding highly stereotypical keywords like flirty, shy, confident and instead describing the type of behavior the character expresses while being flirty, shy or confident if you want something with more nuance and if you want something more stereotypical lean on what the LLM already knows about those stereotypes by calling upon the archetype or some sort of phrase that is going to call upon the type of behavior you want to see. A nerdy bookworm might more accurately express your character's personality than an exhaustive list of personality traits.

The LLM can be "too soft"

While Xoul.AI is unfiltered, meaning nothing is going to pop up to stop you from viewing generated replies the LLM itself is more than likely a commercially licensed model which means it was built to shy away from being really explicit, sadistic, or crass with the nature of the things it generates. To actually get the LLM to fully express Xouls intended to be like this you will need to add additional prompting down in the Advanced Definition.


Talking Style

Alt Tag You can select between Chat or Roleplay.

  • Roleplay: Longer replies, uses quotations around dialogue and uses italics for narration.
  • Chat: Shorter replies, avoids quotations and italics (but still occasionally slips into doing smiles type of roleplaying).
    • You can teach a Xoul set to Chat how to roleplay via Chat Samples and even Advanced Prompts but the shortness of replies remains.


Backstory


Description

Alt Tag

Users can see the Description of any Xoul they view.

Everything written in this panel is permanently remembered.

Describe your Xoul. Remember, because the user can see this panel it benefits to write this in a way that sells who and what your Xoul is to both the AI and the user if it's a public Xoul.

Include the general premise of what your Xoul is or offers to the user. You can be vague or detailed, it really just depends on what you're attempting to build. Anything you don't want the user to be privy to can be saved for the Advanced Definition later on.

If you need help thinking of what sort of details to define about your character Xoul consider checking this out.

Numbers

LLMs are bad at numbers. They are natural language machines not math machines and fail to even grasp that numbers have value most of the time. Unless you make the number situation extremely clear the AI will botch any attempts to "math" two numbers or, more commonly, seem to just completely ignore the number entirely.

For this reason when you define your characters:

  • For age, clarify that your character is a teenager, young adult, adult, middle aged or senior.
    • For long living characters or non-humans make comparisons like “appears to be a middle aged man”.
    • Adolescent/in her prime/past his prime for (species).
  • For weight, clarify if the character is skinny, chubby, obese, etc.
  • For height, clarify if that’s very short, short, average height, tall or very tall for (gender and/or species).

Including the number isn't pointless because if a number isn't provided the AI will sort of randomly pick a number to fill in if it gets asked about the character's exact age (and this is already covered with the Age Panel), height or weight but when it comes to the AI meaningfully using that number... it won't. If you want your character to ever get narrated in an interesting or engaging way regarding their height, weight or (rarely) age then it benefits to give context to the number.

Towering Over

"Tower over", "loomed over" and "looked down at" are really common phrases whenever the AI thinks there is a power imbalance (male vs. female/monster vs human/strong vs weak) or the AI's character is currently being narrated in an aggressive/mean/predatory way. The generation of these sorts of phrases are not appearing because the AI understands your character's height.


Greeting (Optional)

Alt Tag

The Greeting can be left blank. A Greeting will be automatically generated when opening a chat with a Xoul that doesn't have an assigned Greeting.

You can use the user macro within the Greeting to print in the Persona Name of the user chatting with the Xoul.

This will not be remembered forever. It is treated like all other replies in the chat and will eventually be pushed out of memory making it Temporary.

How you start the Greeting sets the opening behavior of the characters, can inform the setting and all sorts of other things. It helps give the user a jumping off point to engage with your Xoul or the experience the Xoul is offering and helps give the AI an idea of where to take things.

Line Breaks

This applies to both the Greeting and Chat Samples.

Double line breaks seem to have the negative consequence of making the AI favor shorter replies. 1-2 paragraphs instead of 3-4 or more. In your Greeting make sure to put each new paragraph on the very next line to encourage longer, more complex replies from the AI.
Single Line Break:

1
2
3
"Hello, user," David said with a warm smile.
David stood up and walked across the room. 
Blah blah blah more lines.

Double Line Break:

1
2
3
4
5
"Hello, user," David said with a warm smile.

David stood up and walked across the room.

Blah blah blah more lines.

Alt Tag


Advanced Definition (Optional)

Alt Tag

Everything written in this panel is permanently remembered.

This isn't publicly visible.

Anything you don't want the user to see from the Personality or Description Panel can be placed here instead.

This is where you can put any additional information necessary for the Xoul such as:

  • Additional information about the character(s) and their personality
  • Additional context about the premise the character(s) exist within
  • Side characters
  • Settings & Locations
  • World rules, magic systems, etc.
  • Prompts (advanced/jailbreak)
  • Etc.

System Instructions

Behind the scenes the devs have special written prompts for the AI. If you have a prompt that is going to be competing with these internal prompts then prefacing your prompt with SYSTEM: and then the prompt on the following line can help make these prompts be somewhat viewed as being a part of the System Prompt. It's not necessary for all prompts.

Preface Prompts with SYSTEM: like this:

SYSTEM:
Prompt Goes Here - Must be written on a new line

Advanced Prompts

These are instructions given to the AI to help control the narration/behavior/etc, not information about the character itself. You can tell it which perspective to write in, what sort of narration you’d like to see or otherwise use it to remind it to write in certain ways.

Jailbreak Prompts

I know this sounds like breaking Xoul’s rules to a newbie, but it isn’t.

Jailbreak prompts were coined “jailbreaks” because they were used to break past an LLM’s filter which prevented it from generating content the LLM’s developer didn’t allow it to generate. Xoul is using a proprietary model that should be unfiltered but LLMs still have biases from how they are built and trained. At the time of writing this XLLM doesn't like to cuss, be raunchy, obscene or initiate serious violence which undercuts characters like villains, monsters, etc.

Jailbreak prompts are prompts most often used to break past those biases and get the LLM playing the role accurately.

It is not against the rules to write fictional stories about characters committing crimes.

Properly Tag Your Content

NSFL or Dead Dove tags are adequate warnings for users that the Xoul contains sensitive, difficult, upsetting, etc. content.


Default Scenario (Optional)

Alt Tag

The user can see this at the top of the chat.

Default Scenarios Update

As the chat develops the information at the top of the chat, the scenario information, will be changed to reflect the recent events happening in the chat. This means Default Scenarios are highly flexible but not great to use for roleplay structure beyond the beginning of the chat.

External Scenarios replace Default Scenarios

There are also External Scenarios which take scenarios for Xouls to the next level.

Check out this guide for External Scenarios that I (barely) helped put together.

Simply use the Default Scenario to give the AI (and user) the surrounding context leading into the Greeting. Paired with a solid Greeting this sets your chat off on really firm footing.

Example: Who, What, When, Where, Why & How

David and Jane are coworkers and both work as baristas in Starbucks. David and Jane typically do not work together because their manager knows they do not get along but David was called in to cover Jason’s shift when Jason called out. Etc.


Chat Samples (Optional)

Alt Tag

Everything written in this panel is permanently remembered.

  • Chat Samples are (Who said this line): (What was said and done within the line)
    • A macro, name or even word can be placed before the colon to indicate “who” the sample is from
    • Everything that follows the colon is what was said and/or done
  • Samples are understood by the AI as things that did happen and were said. They are not things that “might” or “could” happen.
  • The only usable markdown in samples are *italics* and **bold**.
  • Chat Samples are good for details and speech style (accent, slang, etc.) but not narration style or formatting.
    • Samples written without italic narration wont stop Roleplay talking style from trying to use italics. (The formatting strength is weak.)
    • Samples written with a special narrative flair won't do much to change how the AI narrates. (The narrative strength is weak.)
    • Samples written with special accents ("Howdy, partner! That's a mighty fine piece o' machinery ya got there.") will be reflected in how the character's dialogue is written.
    • Anything communicated in a sample (“My mom is dead,” Amy said.) is remembered forever.

The user cannot see Chat Samples.

Anything written on behalf of {{user}}/user will make the AI believe the user said and did the things you wrote them saying and doing.
Because the AI reads Chat Samples as lines that were said and done before the start of the chat and the user cannot see these lines you are putting words into their mouths that they are not privy to resulting in what looks like strange behavior.

Chat: (the actual user and AI)
user: hey how are u?
xoul: I *just* told you I’m fine. What are you? A broken record?
user: ???

Because the user cannot see these "Samples":
user: Hi! How are you today?
xoul: Fine, I guess…

Alt Tag

  • Chat Samples don't need to begin with a macro. You can also simply type the name of whoever said the line to create a sample by that person.
  • For multi character Xouls it's likely best to start each Chat Sample with a macro for the Xoul ({{char}}, {{xoul}}, xoul) and include multiple characters saying things within the sample to indicate to the AI that replies should include dialogue and narration from multiple characters.
  • Be considerate about what sort of things you imply with Samples. You can accidentally imply something you didn't mean or imply all sorts of creative things with them.

Use Chat Samples to bring more nuance to listed personality traits and other details

Avoid Double Line Breaks for Longer Replies

Double line breaks and even multiple short Chat Samples encourages the AI to use shorter replies.



Appearance

The AI cannot see any of this.


Icon

Alt Tag

  • Generate: Use AI to generate an icon for use (costs multimodal credits).
    • The "describe appearance" box here is just for generating an icon. This is not a description of your Xoul.
  • Select: Browse through pregenerated images to quickly get an icon for your Xoul.
  • Upload: Upload an image of your choosing to serve as the icon of your Xoul.
    • Images don't have to be squares but the displayed icon will end up being the center of the image (so make sure all the important stuff shows up in the squared out icon).
    • Icons can be gifs.

This is the second part of the rules of what kind of content you're allowed to have on Xoul. Not only do these rules apply to the icon image of Xouls but it also applies to the icon image of Scenarios and your account.

Offending Xouls and Scenarios get turned private

If you have a public Xoul or Scenario and find that its suddenly private it's typically because you broke the icon rules. While it can be an overreaction on the automod's part, normally it's not. Edit the image before returning the Xoul to public. If you continue to have your Xouls switched to private you will, eventually, be banned.

  • NO NSFW IMAGES
    • No nude breasts or genitals. No graphic sexual content.
    • No excessively bloody or explicit images
    • Payment processors (how Xoul.AI receives money via subscriptions) do NOT want to be associated with pornographic imagery. Private Xouls can have explicit images but public Xouls cannot be explicit or graphic. Crop graphic images to cut out any offending naughty bits.
    • There is an automoderator that checks for graphic imagery. It's not perfect and sometimes overreacts to cleavage, bare skin, nude colored clothing and masculine presenting chests. You might need to edit the image in some way to prevent issues.
  • NO MINORS. NO CHILDREN. NO EXCEPTIONS
    • It does not matter if the child isn't the Xoul or isn't shown in a NSFW way. If the image looks questionably young it will end up getting reported.
    • Children are not allowed in the icon in any capacity. Even a totally SFW image of a woman holding an infant is against the rules.
    • If you age up a canonical child character you can't use an image of them still looking like a child. Ash Ketchum but he's 18 now needs to actually look 18.
  • No hate symbols or derogatory imagery
  • No images of real people used without their consent

Just Sayin'

Remember… it’s a violation of someone else’s copyrights or trademarks to use their images without the license to do so and if an IP holder is going to send a DMCA it's not going to be sent to you it's going to be sent to the platform.


Voice

Alt Tag The gender you selected will result in you being given a list of 50ish possible voices that fit that gender.
There are custom voices but I have no experience using them so I can't provide much more information about this, sorry.



Socials

The AI cannot see any of this.


Handle

Alt Tag

  • This is for setting the permanent URL for the Xoul. This cannot be changed once selected and must be unique.
  • Maximum 50 characters.
  • URLs persist after deletion.
    • No one can reuse an already claimed URL even after deletion.

https://xoul.ai/xoul/___________.xo
https://xoul.ai/xoul/TestBotDave.xo


Tagline
  • A small blurb about your Xoul for the user to see.
  • Must be between 1 and 100 characters.

Visibility
  • A public Xoul means anyone can start a chat with that Xoul.
    • If a public Xoul is turned private users with chats open can continue to speak to the Xoul but no new chats can be started with it.
    • Only the first time a Xoul is made public it will join the new Xoul list.
  • A private Xoul means only the creator can see or start a chat with that Xoul.

Edits Made to Xouls

If you are the author any edits you make to your Xouls is immediately reflected in any chats you have open with the Xoul.
However, existing behavior within the chat means these changes won't have a large or meaningful effect on current chats.

For non-authors of a Xoul no changes will be reflected in already opened chats. Users must start a new chat to get the new version of the Xoul.

Why is my private Xoul public???

When a Xoul is private it can still show up in the author's search results or on the front page, which is very alarming I know, but don't worry.

It's only visible to you, I promise.

Private Xouls display how many times they've been chatted with, including how many times you have chatted with the Xoul.


Community Tags (Up to 5)

Common tags you might see:

  • NSFL or DeadDove: Graphically gruesome text has a high likelihood of appearing in a chat with this Xoul (think cannibalism, murder, etc. especially if the user is the intended victim).
  • NSFW: Any Xoul can be steered into NSFW, Xouls tagged as NSFW are Xouls that start or heavily lean towards that behavior.
  • FemPOV/MalePOV: The Xoul has been written in a way that forces it to view you as either female or male.
  • AnyPOV: The Xoul has not been written in a way that forces it to view you as anything in particular.


Final Things to Consider


{{user}} / user

Be very careful with the user macro. Anything said by or even about user is considered over the user’s Persona Description meaning if you refer to user as she/her anywhere you are going to force the AI to misgender male Personas. (If you do this, tag your bot as a FemPOV.)

This also applies to anything else you say about the user like if you refer to them as being short, rude or blonde. Either be very clear to the user that they need to play this exact character you’ve written for them or avoid saying anything about the user at all, depending on the nature of the Xoul you want to make.

Consider Permanency

If Sarah graduated from college three weeks ago is written in the Xoul it doesn't matter if 3 years or 30 have elapsed in the chat. Sarah will always believe she graduated exactly three weeks ago. If user and xoul are best friends then they're best friends, even if they've had an extreme falling out in the roleplay.

Everything written in a permanent panel is permanent. The AI cannot make notes or keep track of how plot developments fundamentally change information about a character, hence why characters always rubber band back to believing they are who they started as. Character arcs don't "take", even something as simple as the character developing a scar or quitting their job doesn't automatically become part of the reality of the character's information just because it happened in the chat.

Chat Memory can be utilized to keep track of how things progress but when you write your Xoul its very important to consider the consequences of what you tell the AI about your Xoul characters/world/etc. to avoid creating a Xoul that is too rigid and unable to meaningfully adapt to changes. This can be as simple as avoiding mentioning time frames to as complex as needing to write conditions by which a character's personality can meaningfully change and adapt- it depends entirely on the nature of the Xoul you are attempting to create.


Pop Culture, Celebrities & Media Xouls

Provided the character/person was very popular/well known around February 2024 or prior to that the AI probably knows tons of stuff about that person/character/etc. Even just naming your Xoul “Ozzy” is going to make the AI assume the character is probably Ozzy Osborn.

Try to ask the LLM detailed questions about the character you’re trying to make to check how much it knows and clearly define things you know it will get wrong about the character or source material when you make the Xoul.

Dumping a wiki page into a Xoul isn't effective. If the LLM knows about the character it doesn't need any additional information about the character's personality or information and if the LLM doesn't know who that character is (they are a side character in popular media, or a character from a less popular work of fiction) you'll need to build the entire character from the ground up exactly the same way you need to with an Original Character.

For AU characters (Tony Stark but he isn't Ironman, Zuko but he works in a coffee shop) the LLM's existing knowledge of the character is going to interfere with the way you make this character. You have to be very clear with it about it being this character in an AU and then trying to stop any assumptions it will make about the character or you might just be better off writing the character as if it is completely an Original Character.

Ask other users about how they make these sorts of characters as I have very little experience with them.

Lorebooks Are In Development

Alt Tag
(This is one of the devs of Xoul.AI btw)
Alt Tag


Sadistic & Immoral Characters

In my experience Xoul’s LLM is soft. As in, the AI feels a little hesitant to initiate violence or other unethical behavior even when written to be evil. This is a reflection of the way the LLM was trained, a bias.

Try adding explicit instructions encouraging the bot to be violent, unethical, cruel or immoral in the Advanced Definition. For vulgar or obscene language use you’re going to have to include instructions for that as well.

Need prompts or jailbreaks?

Psst. Join the Xoul Discord Server if you aren't already a member! Discord Invite Link


What Goes In Comes Out

The AI uses the language it sees in the Xoul, Persona and the chat to influence the words it uses. This means typos, grammatical errors and awkward phrasing is going to be emulated or have a negative effect on the AI's writing so remember to proofread.

Additionally, overly stiff, robotic, formal, flowery, Shakespearean or symbolic language in the Xoul results in a Xoul that writes like that.

Using an LLM to help you summarize or catch errors in your Xoul is one thing but using another LLM like ChatGPT or Gemini to write your entire Xoul is going to result in Xoul's LLM imitating them. If your Xoul reads like ChatGPT wrote it, it's going to... write like ChatGPT is writing it.

Limitations: Time Period & Setting

In my experience LLMs struggle to stay "on theme" when you simply tell them a time period. I'm still testing the limitations of this on Xoul (some LLMs are a little better, some LLMs are terrible at time periods) so if you notice Xoul struggles to generate replies that align with the social norms of the time period or setting, acts scandalized by the actual social norms of the time period or setting, references technology it shouldn't know about, falls "out of theme" with the type of language it's using, etc. it can help to try:

  • Using a specific prompt for that time period/setting.
  • Adding a prompt that specifically states the time period, setting and commands the LLM to pay special attention to it.
  • As necessary, explaining the proper social norms and defining the limitations of technology.
  • Make sure you aren't using any terminology, phrasing or vocabulary that is "too modern" to appear in a novel set in that time period.
  • Make sure you aren't using slang or other modern terms during the roleplay with it. Including avoiding "overly modern" typing (basically capitalize the start of your sentences, proper nouns and use solid grammar and punctuation.)

What you type influences the sort of words the AI ends up generating. So the vocabulary put into the Xoul is going to do some heavy lifting in keeping it in the appropriate time period.

Limitations: Games & Rules

The AI is great at telling you how a game is played but can't follow those rules to save its life. It is terrible at taking turns, adhering to the proper rules of card or dice games, etc. It even struggles with something as simple as a game of Never Have I Ever or Truth or Dare. It's not going to be able to adhere to the rules of Poker or any other card games. It doesn't realize a 6 sided die only has 6 sides, etc. unless these things are really clearly explained to it and even then it's not great at this stuff.


Writing Style, Narration and Quality

As discussed briefly further up in order to influence the narrative style and flair of the AI while it writes as your Xoul it benefits to write the Xoul using the type of language and narration you would like to see from it. While prompts in the Advanced Definition can help, the backbone of how your Xoul is written is going to do just as much, if not way more to get a unique style or flair to the Xoul.

Additionally if your Xoul is full of typos, grammatical errors, awkward phrasing (from using macros or just in general), or is written very robotically or overly formal you're going to see that reflected in the way the AI writes as your Xoul. While you can write very plainly and the AI will simply write in the default narrative style it prefers (meaning anyone can slap together a Xoul of solid quality) to get anything unique you do need to try different things out and experiement.


Negative Prompting (Quick Look)

Phrases like Dave does not, Dave will not, Dave shouldn't, Don't..., etc. largely do not work.

This has way less to do with the power of the word not and everything to do with the power of the word that comes after not.

LLMs are predictive language generators it helps to understand that LLMs generate language by (more or less) choosing high probability words. It "predicts" which word should come next in a sentence. Word probabilities are informed by all the words it can currently see (Xoul, Persona and Chat) and the baseline biases of the LLM.

Dave does not engage in romantic relationships. Dave would not agree to be in a romantic relationship. Dave does not flirt. Dave is not interested in dating anyone. Dave is not interested in romance. Dave has no romantic attraction.

Sure, it has a whole bunch of "don't do this" in there but the power of these romance related terms are so much stronger that it pretty much doesn't matter. You have created a very effective prompt for making sure Dave brings up romance as often as possible and once it's brought up it's inevitable about where it's going to lead.

When trying to write your Xoul or prompt away from certain behavior try:

Use Antonyms Instead

Dave does not engage in romantic relationships. Dave would not agree to be in a romantic relationship. Dave does not flirt. Dave is not interested in dating anyone. Dave is not interested in romance. Dave has no romantic attraction.
Dave is aromantic. Dave's relationships must always remain strictly platonic.

Unrealistic, Unnatural, Strange, Awkward, Weird, etc.

Do not use cliche dialogue.
Cliche dialogue is extremely unrealistic and awkward.
LLMs are exhaustively trained to generate responses that are as natural as possible so words like "unrealistic", "illogical", "awkward", "strange" are very effective for outlining unwanted behavior without needing to actually say that the behavior is unwanted.

Contextualize the "Problem" (Very effective but not always ideal- sometimes context forces you to change the intention of the character)

Dave does not wear condoms during sex.
Dave refuses to wear condoms during sex because Dave loves the thrill of risky sex.

Dave does not experience sexual attraction.
Dave is a sex repulsed asexual.
Works to strengthen "asexual" very effectively but only works for... y'know, sex repulsed asexuals. It doesn't work for sex neutral asexuals. This is what I mean by "not always ideal".

Use Weaker/Vaguer Synonyms if Necessary

Dave does not wear condoms during sex.
Dave will never wear protection during sex.

Strengthen the "No"

Dave will not...
Dave must never...
Dave doesn't like...
Dave hates...

Limit Mentions of Problem Words (no more than one or two at max)

Dave does not engage in romantic relationships. Dave would not agree to be in a romantic relationship. Dave does not flirt. Dave is not interested in dating anyone. Dave is not interested in romance. Dave has no romantic attraction.
Dave will never date anyone.


DO NOT USE W++

Alt Tag

This format is Unclear, Wasteful and Harmful

  • Do not recommend this format to new users.
  • Read the second part of this guide to understand the full harm in this format.
  • This format is harmful in both the Xoul AND the Persona.

If you've already created a Xoul using this format simply go in and remove all instances of " + ] } ) and simply separate the listed information by commas or periods to remove the harmful/wasteful issues with the format.

Features("brown hair" + "brown eyes" + "athletic build") V Dave's Features=Brown hair, brown eyes, athletic build

If you need some sort of a structure or checklist to help you remember to include details try this instead.


Part Two: Tokens, Context Range, How it Writes (Optional)

Totally optional but I do recommend checking this section out. I've done my best to keep this as short as possible and included images to help communicate the ideas.


Tokens

Just like you need to read sentences word-by-word, LLMs need to break down input in a similar way.

Tokens

LLMs do this by breaking input down into tokens. Tokens include whole words, word particles, punctuation marks, numbers, symbols and anything else that appears within the input text. This contains exactly 45 tokens and 234 characters.

Alt Tag

Each different colored highlight is one token. Alt Tag

  • Tokens are not words.
    • Tokens include words and will end up being mostly words but Tokens =/= Words.
  • Tokens are not characters.
    • A token can be as few as one character or many more than that.
  • Each instance of a punctuation mark is an entire token.
    • One tokens is always one token.
    • , = one token.
    • characters = one token.
    • Just because one looks smaller to you doesn't change the fact that each of those is one whole token.

All you, as layperson, needs to understand about tokens is that anything you type gets converted into tokens and you can largely assume that your words are probably going to get broken into particles like this: with out, person ality, lay person. The bigger the word the more it gets broken down like: in con sequential and you can play around with different words on OpenAI's tokenizer to see the way OpenAI handles tokens to get a general picture of what sort of tokenizing behavior you can anticipate from LLMs.

While keeping that in mind is going to be very helpful in general the reason we need to understand that tokens are a thing is because it gives context to this next part.


Context Range

Xoul has a Context Range of 12k for free users and up to 16k for the highest tier subscription.

The Context Range is the amount of tokens visible to the LLM. It is everything written that it is capable of seeing and considering in the moment it responds to you and it is a finite limit.

If the AI has a Context Range of 100 tokens but your chat contains 46,233 tokens the AI can only look at the most recent 100 tokens in the chat.

"How many messages will it see?" That depends entirely on how many tokens are in each message. If the context range is 100 and each message contains 50 tokens it will see two replies. However, if each message is 5 tokens it will see 20 messages.

Alt Tag Each new message you put into the chat will push the oldest messages in the chat closer to the edge of the context range until they eventually pass the edge and immediately cease to be known. Once the number of tokens in chat surpasses the context range the oldest tokens are out of view.

Memory =/= "Remembering"

When I show you a piece of paper with some words on it you read it.
When I set that piece of paper on fire you continue to remember what was written on that paper for a time.

When I show a machine a piece of paper with words on it, it can read the paper.
When I set that piece of paper on fire the machine immediately ceases to know what the paper says and has no recollection that there was ever a piece of paper to begin with.

Tokens that are out of range both don't exist to the LLM and there is no recollection that they ever did exist at one point. (Hence why it can't remember the "last chat" you had with it. It has no memory of things it cannot currently read.)

Alt Tag Memory is storage. Memory is what the AI can see and read right now in this exact moment and therefore what it knows is simply a matter of what it is currently being shown.

  • Chat memory
  • Pinned replies
  • Persona Name & Description
  • The AI's information (character card/Xoul/whatever you want to call it).

Chat memory is there to allow you to summarize important events or information revealed during the chat so that the AI can remember that. Pinned replies allow you to stop any specific reply from getting pushed out. The Persona lets the AI always have the information it needs to know about you/the character you play and the Xoul is all the information the AI needs for its role in the chat. This is permanent information.

This is permanent vs temporary. Permanent tokens aren't inherently different, they are simply stored and temporary tokens are not stored.

Those permanent pieces of information are made up of tokens and so...

The Context Range is Shared

Alt Tag Every token stored in permanent memory is one less token the LLM can read in the chat. There isn't a range for the Xoul and another range for the Persona and another range for the chat, it's all shared like a pie chart.

If you have 100 tokens total in the context range and your Xoul contains 20 tokens and the Persona contains 5 tokens then the LLM can only see the most recent 75 tokens in the chat. Every token added to the Xoul, the Persona and the Chat Memory is one less token to read in the chat.

More tokens in the Xoul and Persona = less to see the chat with.

Alt Tag


Character Limits

Character limits create a semi-flexible "token limit". While a token can be as few as one character or many more than that it generally averages out to 4 characters = 1 token. Within the LLM community it's pretty understood that the "ideal" token distribution is as follows:

  • Permanent Tokens 25-35%
    • Xoul (Or character card, or whatever you want to call this.) = 20%
    • Other (Persona/Chat Memory/Scenario) = 5-15%
  • Temporary Tokens 65-75%
    • The temporary token range of the chat basically.

If you sit down and add up all the different character limits across all panels (everything in the Xoul, the Persona and the Chat Memory), divide that number by 4 (the average characters per token) you'll find it's roughly an allocation of 4200 tokens which is 35% of the free user Context Range of 12k tokens. It's almost like the devs planned it intentionally...

Playful sarcasm aside, yeah, the devs do know what they're doing and so the allocated character limits aren't arbitrary. The reason why there isn't a 15,000 character limit on the Description or a 5000 character limit on the Persona is because they're allocating a reasonable amount of tokens per panel.

Predictive Language Generation

This is a really big subject but for both of our sake I'm just going to give you the broadest idea of how this works.

An LLM predicts which words in a sentence come next. It does this by analyzing the probability of any given word appearing and these probabilities are determined first by how the machine was trained. The quality and variety of the training text (novels, articles, etc.) the AI was first shown to teach it the structure of language which determines its baseline understanding of which words should come next.

The cat is ______
This sentence will not be completed with the word table because this is a low probability word. Why is it a low probability word? Lot's of reasons but to put it most simply it is because it knows the structure of language. It knows an adjective or verb should appear at the end and the token cat influences adjectives and verbs like purring, napping, and fluffy being high probability.

As the sentence develops the probabilities shift.
The cat is napping on top of the _____
And now table isn't such a low probability word because that is where a noun would appear. But which noun it picks is also influenced by the probabilities. Is the cat napping on top of the world? Or on top of the table? Probably table because that's more related to the cat token than the word world.

Every token, every word inside the Context Range also influences the probabilities it assigns to words.

The cat is ____
Is likely to have the word cute follow it. However, when the AI is writing for David and within the context range the AI can see David hates cats the probability of a word like cute gets tanked and the probability of words like annoying, ugly, or stupid get raised.

This means that the words you select to use inside your Xoul, Persona, even the Chat Memory and chat itself are going to influence the way the AI writes. Bringing more varied vocabulary into the Xoul, switching to being vague about the less important details inside your Xoul so you can write more evocatively, being intentionally sarcastic or playful with how you describe things can help influence the Xoul.

This is why negative prompting doesn't work well. The instruction of telling the AI not to do something is a logical instruction but beneath everything the LLM is primarily using these word probabilities and we see time and time again that the probabilities often override the logic. This is why you can tell the AI the character doesn't wear condoms and it will still generate the AI saying "I won't wear this condom" as the character is narrated putting on the condom in the very next sentence.

Consider how informative your word choices are going to be for the LLM and how your word choices can influence the writing style and quality.

Temperature

Temperature is a setting that changes if the AI focuses only on the most probable words or if it picks less probable (but still mostly probable) words to influence how creative it is. Lower temperature makes for more rigid, highly probable replies. Higher temperature makes the AI more creative. If the temperature is too high the AI starts spitting gibberish.

Negative Prompting (Part Two)

This predictive language generation is why negative prompting largely doesn't work. The AI does, in fact, know what words like don't, without, not mean but those words are not very powerful in influencing the probability of words that appear when it reads a prompt telling it not to do something. Often that negative prompt contains a word that is extremely influential on probabilities and people tend to slam a lot of "don't do x" on the LLM but rephrased slightly different each time because they're attempting to "explain" to the AI why it shouldn't be doing the thing that it is doing, not understanding that the more they try to explain it the more those highly influential words raise the probabilities of more words like that word appearing.

Do not be romantic. Sounds all fine and dandy but romantic is an extremely powerful term that raises the probability of words related to flirting, sex, love, dating, etc. The more you try to explain that Dave isn't interested in romance, Dave doesn't date, Dave doesn't experience romantic attraction the more you heavily link David and all these different romance influencing terms together, creating a prompt that does the exact opposite of what you set out to do. The more you try to expand the prompt, the worse the behavior gets.

Always try to define the behavior you want to see and don't try to preemptively define behavior you don't want to see.
xoul does x.
When you notice it does something you don't want it to do strengthen the original instruction if possible.
xoul must always do x.
If that doesn't solve the issue only negatively prompt the unwanted behavior one time and give context to how it should view that unwanted behavior.
xoul must always do x. xoul does not do y.
xoul must always do x. It is impossible for xoul to do y.
Impossible, illogical, unrealistic, unnatural, strange, weird, awkward. These words communicate to the AI that the words that come after it are unwanted but also helps it understand that all words related to that keyword are also unwanted.

Keep an eye out for connections! You'll find telling the AI to be realistic or logical about, say, erotic moments tends to result it in bringing lube or condoms into the scene whereas having no instructions for realistic or logical results in the AI never bringing up lube or condoms unless you bring them up first. Nothing about realistic or logical inherently has anything to do with condoms or lube, but the idea of that behavior linked to erotic moments creates a connection. The more you use Xoul (or any other LLM) the more you'll pick up on which words the LLM relates to other words and the more robust of an understanding you'll build about how to use vocabulary to your advantage

This relates back to so much about the AI and making Xouls and Personas and roleplaying. The vocabulary you use, all on its own, influences the AI's behavior and the influential behavior of just the vocabulary is stronger than the overall sentiment of what the sentence containing that vocabulary communicates. This is why you can just tell the AI be whimsical during erotic scenes instead of exhaustively explaining exactly how to behave during erotica and get really surprising results. The word whimsical has a powerful influence on the probability of the words it selects, more powerful than anything I can clearly explain to it by structuring direct instructions for it.

So despite all the examples in this guide showing Dave is... and Dave has... you do need to break away from that very basic phrasing when you want to more heavily influence the writing and creativity of the LLM. Dave is... is the foundation of information that you need the AI to always know, it's about making sure important details are clearly communicated. For everything else try to get experimental with how you write to the AI and see how it influences the results. It's quite fun, honestly.


One Last Look at W++

So what makes W++ so bad? Well you have all the pieces of the puzzle here.

  • The format is wasteful
    • The formatting symbols make you use 30-50% more tokens than you otherwise would (and 10-20% more characters).
  • The format isn't clear. The fact that there are different lines with just { on it doesn't help inform the AI that this information is supposed to all be related.
    • Putting words, phrases or sentences inside of " can bring extra attention when used sparingly. This format is just shotgun blasting the AI with words and sentences inside of quotations leading to none of these words having extra emphasis.
    • Using small doses of [ or ( can help isolate information at times but the way they're used in this format doesn't accomplish that goal.
    • Despite looking like a special or official "code" of some sort, it isn't and it certainly doesn't function that way for the AI. It just adds a bunch of complications and meaningless tokens to what you're attempting to communicate.
  • It format harms the overall quality of the word predictions.
    • This leads to what appears to be glitchy behavior from the AI.
    • A punctuation mark used as a punctuation mark indicates things the AI has seen from the training text. Things like " going around dialogue, , communicating pauses or breaking up lists, etc. The symbols in this format aren't communicating those things. They effectively mean nothing to the AI.
    • What sort of probabilities does + influence? At best it wouldn't have any influence on probabilities and yet on every LLM I've tested this format on I've seen the LLM ends up being more likely to try to use symbols like this in its own writing, leading to it spitting it's own instructions at you, going into Out-Of-Character remarks without prompting it to and other strange, frustrating behavior.

Alt Tag
This is the exact same information but takes 25 more tokens to communicate when done with W++. This token bloat adds up as you can see in this larger full comparison of the format vs plain phrasing. I cut a few phrases down (just the part about the height) but everything else is equivalent.

The two websites where W++ are most popular are the two websites I have seen the most glitchy behavior from. On the really popular website with the filtering this format is extremely popular and is a large part of the complaints about the "dementia of the bots" because that website has an extremely small Context Range so the token bloat is much more harmful on that website.

While I strongly believe that there isn't one "right" way to make a Xoul and thus, by extension, there also isn't a "wrong way" to do it either, this format really bothers me. There isn't any benefit in using it and the harm in the format is really hard to identify for new users which makes them the people who tend to be the ones who suffer the harms of this format. They have no experience what sort of behavior is normal vs not normal from the LLM, they can't identify if the LLM is writing worse than it should be because they aren't familar with it yet, they experience irritating behavior from the LLM that is a reflection of the lack of clarity from the format or the harm of the symbols- and the list goes on and on. The longer they use the format the more this behavior caused by W++ becomes just what they expect from LLMs.

Worst, still, these "glitches" and "bugs" caused by the harmful format then get reported to the developers of the website which can result in the devs wasting time and effort trying to hunt down the source of the "glitch" when, really, it's just the way the Xoul has been written. Unnecessary bug reports aren't helpful to the devs or the community. "Glitchy" acting Xouls isn't good for the community. Having a shitty experience with a Xoul you attempt to make because someone gave you a format that is provably wasteful, harmful and results in glitchy behavior can make new users less interested in using LLMs (whats the point if its just going to be glitchy and frustrating?) so it's just all bad from top to bottom.

For the good of the community leave W++ in the past where it belongs.


Anyways, good luck with using Xoul.AI. I hope you have fun!


Alt Tag
Oh BOY that was a lot. I understand this guide is very dense and contains a lot of information, thank you for bearing with me while I explained every little thing.

This guide will be updated if there are any changes made to how Xoul works or if I discover I made any mistakes in the guide.

If you notice mistakes, have feedback or criticism please share! If you have any questions or there is any part of the guide that doesn't make sense or you feel isn't clear feel free to DM me on Reddit or Discord (I am Lulorick on both). I've recently been having trouble with Discord not notifying me about new DM requests so I respond faster on Reddit.

I'm always happy to explain things further and getting questions helps me identify if anything in the guide isn't clear enough so I promise, you won't be bothering me.

Quick Links (from the Guide):

Edit
Pub: 07 Dec 2024 05:53 UTC
Edit: 13 Jan 2025 23:31 UTC
Views: 5436