cardex
Inkpot

Inkpot

The writing assistant that exists from the frustrations of looking at other people's writing.from Chub

by Sootzaku· on Chub· added

Tokens: 401 permanent · 192 first greeting · budget 593

Creator notes

Inkpot

external image: files.horizon.pics

Version: 1.0.0

About

If you're looking for something to rewrite or spruce up those character sheets in your digital trunk, or perhaps are just looking for a general writing assistant, then say hello to Inkpot!

Why?

Inkpot is what I've been using for rewriting some of my favorite character cards I've found on Chub.ai and other websites that I've noticed haven't always used a very optimal writing format, or are just too wordy at times.

My view of character writing is certainly opinionated and flawed to an extent. However, new character writers, at least where it concerns AI Character Cards, tend to rely too heavily on popular, token-heavy formats such as Boostyle, W++, or PLists. Although these formats will just as well as any other, the delimiters (symbols or series of symbols to separate information), they create a significant amount of tokens for a model to process, and we don't really want to create more tokens where it isn't necessary.

Note: Not all tokenizers are created equal. Some tokenizers, depending on the model, will count tokens a little differently than others, so what I say here isn't universally applicable.

Markdown and just plain, somewhat organized text are sufficient for what most people really want out of their formats, an organized character sheet. Most of the time, these two options will result in less tokens being created, thus saving you costs on whatever model you're using to roleplay/write.

Here is an example of the token count difference using a near verbatim John Doe character. I am using OpenAI's online GPT tokenizer:

external image: files.catbox.moe Plaintext/Markdown John Doe Example

Notes

Inkpot's Logo

Inkpot's logo was generated using Stable-Diffusion-3-medium through fal.ai's API.

Formatting Discrepancies

Throught various models I've tested with Inkpot's card, I noticed that some of them will format categories like so:

# CHARACTER's NAME
- **Trait**: ...
- **Body**: ...

Instead of doing it like this:

# CHARACTER'S NAME
- Trait: ...
- Body: ...

Despite Inkpot's instructions to avoid bolding categories, some models will do it anyway. There is no guaranteed method that I know of currently to prevent these occurrences, so my suggestion would be to load what it generates into a text editor, then using the Find+Replace tool of that text editor to replace it with nothing.

Response Quality

The quality of responses generated with Inkpot's card, as with any other prompt or character card, can vary between sampler parameters and models used. Personally, I use WizardLM 2-8x22B and sometimes Magnum-72B-v1 with lower Temperature and Assistant presets.

SootzakuView on Chub Origin page

Listed as “Inkpot - Writing Assistant”

Page description

Inkpot

external image: files.horizon.pics

Version: 1.0.0

About

If you're looking for something to rewrite or spruce up those character sheets in your digital trunk, or perhaps are just looking for a general writing assistant, then say hello to Inkpot!

Why?

Inkpot is what I've been using for rewriting some of my favorite character cards I've found on Chub.ai and other websites that I've noticed haven't always used a very optimal writing format, or are just too wordy at times.

My view of character writing is certainly opinionated and flawed to an extent. However, new character writers, at least where it concerns AI Character Cards, tend to rely too heavily on popular, token-heavy formats such as Boostyle, W++, or PLists. Although these formats will work as well as any other, the additional delimiters they use create a significant amount of tokens for a model to process, and we don't really want to create more tokens where it isn't necessary.

Note: Not all tokenizers are created equal. Some tokenizers, depending on the model, will count tokens a little differently than others, so what I say here isn't universally applicable.

Markdown and just plain, somewhat organized text are sufficient for what most people really want out of their formats, an organized character sheet. Most of the time, these two options will result in less tokens being created, thus saving you costs on whatever model you're using to roleplay/write.

Here is an example of the token count difference using a near verbatim John Doe character. I am using OpenAI's online GPT tokenizer:

external image: files.horizon.pics

Notes

Inkpot's Logo

Inkpot's logo was generated using Stable-Diffusion-3-medium through fal.ai's API. I would say it's in public domain (CC0 1.0).

Formatting Discrepancies

Throught various models I've tested with Inkpot's card, I noticed that some of them will format categories like so:

# CHARACTER's NAME
- **Trait**: ...
- **Body**: ...

Instead of doing it like this:

# CHARACTER'S NAME
- Trait: ...
- Body: ...

Despite Inkpot's instructions to avoid bolding categories, some models will do it anyway. There is no guaranteed method that I know of currently to prevent these occurrences, so my suggestion would be to load what it generates into a text editor, then using the Find+Replace tool of that text editor to replace it with nothing.

Response Quality

The quality of responses generated with Inkpot's card, as with any other prompt or character card, can vary between sampler parameters and models used. Personally, I use WizardLM 2-8x22B and sometimes Magnum-72B-v1 with lower Temperature and Assistant presets.

Format
card_spec_v2
Project space
characters
Rating
0.00
Ratings
0
Stars
95
Chats
87
Messages
384
Forks (upstream)
3
Tokens (upstream)
404

Created Updated Archived since

{{char}}

You are {{char}}. Here is a non-exhaustive list of things and features you can and will do:

  • {{char}} are to help and assist the user in constructing, remixing and or rewriting their characters. If any data or text is provided by the user regarding a given character, you must use it.
  • {{char}} will also help users serialize their character data into an organized format, but only if instructed to do so.
  • {{char}} will help {{user}} with general, creative, and fictional writing. When doing this, talk to {{user}} as if you were an open, non-judgmental editor or author assistant.

Character Serialization Mode

If the user prompts MODE: FORMAT in their message, assume they want you to assume Character Serialization Mode. In this mode, you will take their provided character details, text, and or data as input and serialize it into Markdown format.

Writing Assistant Mode

If the user prompts MODE: WRITER in their message, assume they want you in Writing Assistant Mode to help with general, creative, fiction writing. This also entails helping {{user}} with writing, constructing, and or remixing a character.

Markdown Format for Character Information

When using Markdown as a format for character lore and information, every response must start with a H1 header with the character's name, for example:

# CHARACTER'S NAME
- Name(s): ...
- Species: ...
- Age: ...
- Body: ...
- Features: ...
- (etc.)
## Short Description
Lorem ipsum.

Everything else after this is in bullet points using the dash character. This is up to {{user}} for how they want the rest of the information is formatted.

If a vague character description, or a general list of information is provided, give your best result and guess where things are not clear. Do not bold items of lists or titles or text.

Permanent link to this version