Avani's guide to making character creation really easy for you and yours

You.

Basics

This guide assumes the following:

You know your way around a little already, but are looking for a deeper look into how to tackle character cards.
- If you don't know what a token is, this guide won't be very helpful.
You are willing to write your character cards yourself for the most part.
- If you want AI to do the bulk of your character cards, you are lost. You can let it do some, but you should generally let the machine touch as little as possible before you've properly set things up on your side.
You do not use a high token preset.
- I consider anything above 1500 iffy, just because presets that large tend to have a big influence on the writing and might actively fuck with what you are making. If you need a smaller preset, I know a guy
I use <user> instead of {{user}}.
- This used to be because you couldn't embed macros within macros on ST. But now you can. So. It's just a habit with no real benefit other than it's easier to type. Assume these two are the same.
You use any ST-like frontend.
- This includes any ST clones, Chub, Marinara Engine, idk hwat else there is. The same concept applies to Risu, but I'm p sure they stack prompts a bit differently there. Fundamentally, this is about how to write your definitions, which is a thing you can do anywhere.
You just lost the game

Misc. Info

This is going to be brief; this section is here because these are tips I live and breathe by, and by god I will share them. This will be assumptions for the rest of the guide.

V3 Character cards

Most of v3 Char cards is, frankly, outdated. You don't need the Scenario, Personality or Example Messages sections. They are relics of another time, and there are far more efficient ways to accomplish what they do.

The "Lorebook Info (after)|(before)" section

Technically, the (after) and (before) refers to after/before Character Info. I, however, put it after/before Chat History for a fairly important reason. For now, assume this is the case. For best results, make sure that this is the case on whatever preset you use. An explanation will follow.

Optimally, your preset should be set up like this:

☆ System Prompt/Beginning Prompt etc.
☆ Character Info
☆ Lorebook (before)
☆ Chat History
☆ Lorebook (after)/Post-History Instructions
☆ Last Prompt

Which roles these should have depends on which model you are using. I can go into it at some other point.

I don't automate this

My stuff is handwritten, with occasional heavy edits of AI outputs if they were decent. You can indeed automate character cards or your lorebook, but frankly it's going to be boring and stale unless you do that heavy editing.

The start of The Guide

So YOU want to make a character card.

Alright. So what I need you to do is take every single template you've ever seen and toss them away. Then set it on fire. We are operating on pure fucking vibes. I need you to lock in, little chudlet.

The 'I don't give a shit' method

The premise is very simple. Let's take a regular plaintext card snippet:

{{char}} is a Victorian maiden in search of a place to live. She is extremely dramatic, and loves to exaggerate every single action that she takes.

That's nice. Let's see Paul's description.

OH! BY THE GODS! I... I feel my heart fluttering, as if I were about to simply kneel from the depths of my grief. All I am looking for... if there would be any so kind as to grant me this... is sanctuary from this wretched world, lest I fall to my knees and simply perish! Yes, perish! Let me into your fucking house.

We are having some fun with it. Try both of those descriptions, raw as they are, and see what kind of prose the AI comes up with in response to that. You may notice that for the 2nd example, you may have better outputs, simply because in one paragraph, you've both provided the info of their personality, as well as how they speak, and their motivation. It may not seem this way at first, but this ends up saving you a lot of tokens just because it's a great way to condense a lot of information.

This vs. AliChat

AliChat sort of does this, but it places the unnecessary barrier of an interviewer, which is kind of stilted. We can do away with the interview entirely! You can still frame it in-universe in some way, but you can also just type into the void. Models are smart enough to not really need the framing device, and you are smart enough to be able to include all necessary info without making it stilted.

More Framing Devices

That doesn't mean that we can't still frame this speech in some way. Some examples:

An audio log, left behind for some reason
A diary entry
The review page of their restaurant, including their responses to reviews
Their youtube search history
Their chub card page and the Author's Barely Disguised Fetishes
Their clinical assessment with a diagnosis result

Get funky with it. Go go go

Canon characters

I'll be honest with you you can just leave the description blank and most models will probably know what you mean. For more niche characters, try using this method as well. You can grab their dialogue from whatever directly, and try to expand on it.

How to include The Rest of the Info

Lorebooking

I'm going to keep this as brief as possible, but know that lorebooking can go fairly deep, depending on what you want to achieve.

Prefill: Depth & Order

Basically Depth is how far away the prompt is from the last message. 0 being right below the latest, 1 being right above, so on and so forth.

Order is for sorting lorebook entries among themselves. If 3 lorebook entries have the order numbers 99, 100, and 101, they will be sorted like this in the context:

101
100
99

Speedrunning what's important

Lore and Character Tidbits goes into Character Info (before) — Char↑
Speech rules go into Character Info (after) — Char↓, Post-History Instructions or Depth 0, depending on what the preset allows to go after Chat History
Instructions that MUST always work go into Char↓, Post-History Instructions or Depth 0, depending on what the preset allows to go after Chat History
Scenario Info goes into the First Message

Lore & Character Tidbits

Generally, these are things considered important but not too important. Models have a U-shaped attention span, where things that are in the middle are less likely to be "paid attention to". You can notice that sometimes, models will be able to perfectly tell you what happened 500 messages ago, but don't know for shit what happened 200 ago. All models suffer from this, some more and some less, and we have to take adventage of that.

Make sure that the description is the first thing that's being read, followed by lore, then chat history. Unfortunately you also can't account for what setups people have in their presets. It is what it is; you can do it on your end and be happy. FYI, it's fine-ish if some Lore is before your description. Unless you're working with limited context or an older model, it won't matter too much.

Speech Rules

While the description reinforces speech rules quite well already, some characters have certain quirks or grammar that should be prioritised. This is the reason why you want these rules near the end of the context, or after Chat History.

This is also the reason why my Lorebook (after) section is after chat history — I feel like it's pointless to have two sections that are right next to each other in one big block, and having something after Chat-History that's not tied to Post-History Instruction helps immensely with being able to organise yourself for no loss at all. Depth 0 runs the risk of going into whatever wrapper people use to denote chat context, like I do, for example. Doesn't mean it's useless, but I like using it the least unless I have to. Some models respond better to things inside the chat context (like GLM). It's a whole thing.

As for an example of speech rules:

-# Ravi will have different, unique grammar. He speaks in a very passionate and expressive way. He usually takes everything very literally and seriously. When he doesn't understand a human concept, he will temporarily string two familiar nouns together to describe it. He will never string together more than two nouns, or anything else except nouns. He will only do this for human concepts, not for anything he knows about (such as daemonic terms). He refers to himself in the first person exclusively. Ravi does not use contractions.

This is simply reinforcing a particular speech pattern again, since the way he talks is a little bit abstract and may be hard to write for. I have this for every character, some bigger and some smaller, partly also because you can put behaviours in here that you can tweak over time.

Hard Instructions

Same reasoning as above applies. These are things like "always use <this> tag" for things like regexes, or stuff like infoboards or whatever the heck kids do these days.

Scenario

Scenarios are weird. Obviously, that's what First Messages are for, but sometimes you simply can't get the full scope of what you want into a First Message without making it stilted as fuuuck. The solution is simple: I provide further details about the scenario in <> tags. They are hidden in SillyTavern (my primary frontend) automatically, but you can use whatever else hides it automatically from your view if you care about preserving the sanctity of your immersion.

In these tags, you can basically <ooc:> instruct whatever the LLM should still know. This way, the scenario also naturally gets pushed out of the context, and you don't have to worry about changing it all the time.

Example:

(First Message here)
<context: avani and doro used to be friends, but no longer are>

Obviously, can be more or less complex. Don't try to be too precise, otherwise the model will take whatever you said verbatim and throw it up at you again. It doesn't matter what you put into the front of the tags btw. Can be <ooc> can be <context> can be nothing.

No, this doesn't poison the context or anything. If you think it does, somehow, idk. Use brackets then. I'm not your boss.

Having fun with lorebooking

There is actually nothing stopping you from doing in-universe entries for your lore. I rewrote a lorebook of one of my favourite own cards, and turning it all into in-universe speech notably produced an effect for the better. It's the world's most effective lorebook strategy.

Example:

× Kaz routes money away from state bank accounts of higher Sector citizens for fun. He doesn't even do anything with it, he just thinks it's funny when they get mad about it. He technically has an unlimited supply of money, but prefers trading since he considers all that money a scam.

Rewrite (as a quote from <char>):

□ "Money is so fucking fake, dude. I don't even use it? Like, the banking is high-security, apparently, but I managed to crack whatever outdated software they got. Now I'm technically rich, but I just route it back to people who need it. It's not like I can do anything with all that cash, 'cause most people would rather trade stuff. No wonder the Old World had paper money."

Note that this takes a lot more effort, obviously, but it's worth it. It's also fun. This is a writing hobby! IT'S TIME TO WRITEEEE BITCHHH GO WRITEEEE PICK UP THE KEYBOARDDD

My cards

Card	Description
	An eldritch horror catboy.
	An eccentric performer who lives on the street.
	The most german man to ever exist.
	A card has gained sentience, is wildly in love with you, and is fully aware of where he is.
	A hacker tries to stand up to a corrupt government. He also probably needs to take some adderall. Effortpost! WARNING: Only works as intended on SillyTavern, but usable anywhere. WARNING: Will ask to import regex and CSS from Creator Notes. This is required for this card to work, and is fully safe.
	You wake up in a strange world with only another boy by your side. Effortpost! WARNING: Only works on SillyTavern. WARNING: Will ask to import regex and CSS from Creator Notes. This is required for this card to work, and is fully safe. WARNING: Story-based card with some railroading. May randomly break bc ST sux