🥞😼STACK (GOURMET)
pancatstack (CIMB) for Claud3
WORLDSIM (Alpha) for LATTE
Features
Toggles
Known Issues
Credits+Links
Instructions+Download
Modifications
Troubleshooting
Contact
WHAT IS 🥞😼stack (Gourmet)
• pancatstack is intended for use with Claude 2.0 + 2.1 and it reportedly also works with GPT-4 and 4urbo. (1106, NOT the newest Turbo)
• pancatstack is an organized, modular preset that utilizes an XML tag with plaintext instruction system.
• pancatstack is good for handling a wide variety of characters all the way from wholesome to dead dove and keeping them distinct, in character, and realistic.
• Optimized with lorebooks in mind and side character friendly + implements them IMMERSIVELY.
• Characters are better at having a life outside of {{user}} and can act independently.
• IMPORTANT: be prepared to make edits If you are uncomfy with violence and SA as NSFW prompt and <thinking> gives {{char}} the ability to commit such acts if its in character for them to do so.
• pancatstack includes many features and toggles to customize your RP experience.
UPDATES
• (Gourmet) WORKS with setting {{char}} cards AND is much better at playing side characters than the original version (add to lorebook for best experience).
• Added Slice of Life, Dark Fantasy,
Horror, and Superhero genres.
• Now only 1 CoT that works for all purposes. Removed from prefill, so you just need to toggle on/off.
FEATURES
Includes many features to improve your LLMs's writing quality and customize your RP such as:
"slow down pardner" toggleable prompt to help slow down relationship-building
Toggleable NSFW prompt which can be used in conjunction with toggleable NSFW ban to make Claude less horny Also helpful for GPT users to make NSFW writing less corny and cringe
CoT (<thinking>) that helps the LLM stay in character and situationally aware
Two writing-style options: General Narrative and Dialogue Heavy
16 genre options to choose from to affect writing and tone: Dead Dove, Romance, Slice of Life, Supernatural Fantasy, Dark Fantasy, Sci-Fi, Horror, Drama Post-Apocalyptic, Grimdark, Cosmic Horror Superhero, Thriller, Mystery, Steampunk
BAN List and Default A/N to help remind LLM to stop speaking for {{user}}
<last_response> prompts to help remind LLM to respond to the current context
TOGGLES
Below are the list of things that you can toggle on-off from top-to-bottom to alter your roleplaying experience. Anything not listed MUST STAY TOGGLED ON AT ALL TIMES.
<sample_dialogue> & </sample_dialogue> - Disable these if you don't use Chat Examples. MUST BE ENABLED if you DO use Chat Examples.
<setting> & </setting> - Disable if you don't use lorebooks. MUST BE ENABLED if you DO use lorebooks for the best experience. Ideally your lorebook is XML formatted but should be fine without.
slow down pardner - Helps with slow paced. Disable if you don't want that.
NSFW - Does not need to be enabled if you aren't trying to generate NSFW content. pancatstack GENERALLY will not push NSFW, except for characters who are horny/coercive. They MAY push NSFW with this enabled.
<thinking> - Helps the LLMs to stay in character IMMENSELY, especially Claude. Also improves their situation awareness, memory of events, and keeping side characters as separate entities (less characters bleeding into eachother). Disable if you don't like/don't want to use CoT, pancatstack still functions just fine without.
writing styles - General Narrative & Dialogue-Heavy. Use ONLY ONE at a time.
genre(s) - Use preferably 1, but it should be fine if you use a 2nd (if you use a 2nd and your LLM gets confused with the tone you will have to modify contents).
Ban NSFW - Use when Claude won't stop being horny.
KNOWN ISSUES
• If your {{char}} is a multi-character bot (i.e. there are multiple character definitions INSIDE of {{char}} description) it may not perform to the best of its ability. One lazier fix is to change the characters inside the card to XML format (<name>copy-paste their description here</name>) and add this before <name>:
You have permission to play as [character name] as well. Control [character name]'s actions, thoughts, and dialogue when he/she/they come up in the roleplay.
it will improve how pancatstack plays them, BUT, for the best experience you should turn them into lorebook entries.
• The bot sometimes blanks, or has short replies in the beginning of chats. This is just something that happens with most CoT prompts (prompts that utilize <thinking>), since CoT is factored into the word count even though regex deletes it. This issue tends to resolve itself after 1-3 messages in my experience but you can also increase the requested word count (found in prefill and Jailbreak) to help amend this.
• Some characters will kind of fuck off. Because this preset is made for realistic character interactions, characters who are unfriendly, have no relationship with you, or no real reason to interact with you will... well, act like it. I prefer this so I won't change it. Feel free to use a different preset or make edits if you don't like this behavior.
CREDITS
I studied a bunch of different prompts and other creators' notes about Claude JBs and CoT so I couldn't possibly attribute everyone but the main ones are: crustcrunch pancatstack's skeleton and prompt order influenced by their Claude prompts
KaruKaru learned <requirements> and <ban> after studying their prompt
rarestMeow learned a lot about CoT thanks to his documentation. Highly recommended to check out his rentry if you want to learn how to best implement CoT yourself.
Biggest thank you's go to ❤️Magnum❤️ my #1! I learned most about prompts from studying theirs and asking them for advice! They also formatted this rentry for me 😍 and ❤️Lilith❤️ gave me lots of useful advice and resources when I was trying to figure out how to make pancatstack into a setting {{char}} friendly preset.
INSTRUCTIONS
1) download attached JSON.
2) if your ST isn't 1.11.1 or higher go update it you silly.
3) IMPORT REQUIRED. Do NOT try copy-pasting the settings by yourself pancatstack has a VERY. SPECIFIC. ORDER.
4) If you use the <thinking> prompt:
USE OF A DELETE <THINKING>REGEX IS REQUIRED.
REGEX CODE
/[`\s]*[\[\<]thinking[\>\]].*[\[\<]\/thinking[\>\]][`\s]*/ims
DOWNLOAD
TEST CARD
If you wanna try it for funsies, here's the setting {{char}} card I made to test this preset with:
(CLICK FOR CHUB LINK)
It's NOT my best work lmao but it did what it needed to do. You can mess with it yourself to see how well pancatstack can handle multiple characters at a time, and the way that I formatted the character sheets in the lorebook.
As you can see, it's super simple, you don't need to do anything elaborate and pancatstack still can play all 6 of them simultaneously and in character. You can also separate from the pack or make a smaller group and return to the rest of the cast later. pancatstack has a lot of freedom, and for the best experience with setting {{char}} cards you should put characters in lorebooks like this.
MODIFICATIONS
• I'm using a specific genre and I don't really like it — feel free to edit the contents to your liking it won't affect pancatstack negatively just try to keep the same style of the other bullet points.
• The prompt list is too fucking long — Delete or unlink genres you don't see yourself using (if you unlink them with the red chain looking thing you can add them back later with that drop down at the bottom of the prompt list).
• I wanna move prompts around — general rule, DO NOT MOVE ANYTHING. If you must, you can move the prompts INSIDE of <database>, just make sure that all those same prompts STAY INSIDE of <database> when you move them around. World Info also MUST STAY inside of the <setting> tags. Theoretically it should also be fine to move things around inside of <guidelines> but I haven't tested that, so ymmv. If you have a huge context size you MAY want to move <script> Chat History </script> above Jailbreak but before screwing around with placement give normal pancatstack a couple of tries.
• I want shorter/longer messages — Word Count is in both the Assistant Prefill AND the Jailbreak prompt. Edit them as you wish, make sure they are the same value, and keep in mind 1) a higher word count is needed for <thinking> since it factors into the word count and 2) when LLMs generate above 700 that's when I notice it has a higher chance to do undesirable things like writing for {{user}} or speedrunning the plot.
• I want to use a CoT but don't really like this one/certain parts aren't relevant to my RP/it uses too many tokens — Trim and edit it as you see fit, if you don't know how to write CoT's I highly recommend reading rarestMeow's materials (linked in Credits).
TROUBLESHOOTING
NOTE: If you moved ANYTHING in pancatstack without knowing what you're doing and it's fucking up do not complain to me lol
• I'm getting blank responses! — If you use a proxy, make sure there isn't already a known issue with it like that it's down or experiencing high traffic. If your proxy is fine, increase the word count in Jailbreak and prefill and disable your "Delete <thinking>" regex or whatever you named it to verify if it's a true blank or just the LLM thinking it's job was done after <thinking>.
• The bot's responses are just really off and I can't explain why! — Restart ST. If that doesn't solve it, try chatting with another bot or go to an earlier branch in the conversation and see if the issues persist. It could be that your chat is already tainted with undesirable behaviors like using flowery language, speaking for {{user}}, repeated phrases, being stuck in a formatting loop, etc. If it is you'll have to scroll up the chat and create a branch from back when it wasn't doing weird things.
• The bot is speaking for me! — How long are its responses? If your LLM is spitting out 700-1k+ words, that's the problem. Swipe or regenerate for shorter responses and/or cut the fluff out of its replies. This is not a preset issue, this is an LLM issue wherein if it's used to generating huge replies it will scramble to find stuff to fill the word count with. This isn't limited to just speaking for {{user}}, it can also ramble on, speedrun plot, repeat past dialogue, etc. when trying to generate huge word counts.
• It generates <thinking> or some other comment in <>s at the bottom of the reply! — This is a rare issue with pancatstack, just edit and delete it. If it repeatedly does this, restart ST.
• It won't respond to my most recent reply! — I believe this is a form of formatting loop since I've experienced this same problem with multiple jailbreaks and even with GPT-4, but I THINK I found a potential fix to this. I had 2 people use this with success but I haven't used it myself yet or had this tested over an extensively roleplay so I can't recommend using it for an extended period of time since the wording is rather strong (my concern is that if left on the LLM will not really bring up past events anymore which could negatively affect your storyline), so just toggle on/off as needed:
- Scroll allll the way down and click the + button to create a new prompt
- Create a prompt with the below settings and copy-paste the following into it:
- Don't forget to save the prompt, and then scroll all the way back up to save the preset. Now, scroll all the way back to the bottom and enabled STRONG <last_response>, and disable the original <last_response>.
I have QUESTIONS/COMMENTS/CONCERNS
I'm in the SillyTavern discord server come chat with me in my thread here. You need to have been a member of the server for a week and have the NSFW role.
I'm also in scyllacord, my thread is here.
| ? | ? |
|---|---|