petNyan® AI Studios presents...

The Schizokino petNyan/sillyCat™ Preset

Which consists mainly of a little trick: Sending the character definitions as AI Assistant to poison the context right from the start!

Disclaimer/Notes

This is mainly an experimental, somewhat unstable preset for people who want to play around with prose styles and are tired of default LLM outputs.

If you are happy with the replies you're already getting, you shouldn't be using this. You'll have better luck with a more standard, lightweight preset that you can easily find around (e.g., by searching something like model name + json).

• IMPORTANT: I highly recommend starting a new chat.
• Mainly tested on DeepSeek-v4-Pro and GLM-5.2 intended to make their writing better. This isn't an end-all, be-all preset, but it will definitely help in achieving more unique outputs, however.
• Some cards work better than others. I recommend bots which have a unique/schizo writing style. The more stylish the definitions, the more you'll squeeze out of the gimmick. Experiment. Or even write out a schizo bot yourself! (And share it!)
• You won't get much out of 'patient file' type of cards, or "-{{char}} is x" spam (unless it's stylish). Works best on 'longprose', particularly with divergent sentence structures and lots of grammar inflections, varied punctuation, character thoughts, etc. (Check the logs for some example bots).
• Can be used on all models, with varying performance depending on whether or not the LLM allows you to kill the CoT, general model intelligence, creativity, swipe RNG, etc.
• Kind of pointless on scenario and mechanics/rule-heavy bots with little focus on writing style (which isn't these cards' purpose to start with), but confirmed to work on them as well—you just get more 'instructional' type of prose, and it may adhere to instructions better (these things like listening to themselves more, it seems).
• This preset does turn the output more schizo (i.e., slightly more "retarded"), but it can and will make some fun stuff happen, plus make the prose style much more "fresh."
• If it's getting too crazy or confusing, consider disabling style injections and/or turning on reasoning for a few messages.
• Though it's plug-and-play, you'll have to tweak and experiment to achieve your desired result; it can and should serve as a base for you to experiment with your own.
• Works best at the start of a new roleplay. Don't expect it to do much (if any) lifting on a 100+ message-deep RP when context poisoning has already established a steady-state.

Download

OK shut up where's the thing >:3

Latest sillyCat™ version: v1.4 (Mirror)
Older: v1.3 | v1.2 | v1.1 | v1.0

Injections

• Last-Minute Guidelines:
  
  To be used mainly with DeepSneed to reinforce the formatting, with which this model can sometimes struggle.
• Schizo Style (Strength):
  
  Use one of these to make the sentence structure wilder. Edit them as you see fit, Strong can get spammy with semicolons and such.
  Feel free to turn them off after a couple of messages and once the prose style has been stabilized.
  There is a random option in case punctuation starts getting too ubiquitous (consider editing them out from messages, too).
  Also, you can just disable them entirely.
• Word Count (Amount):
  
  Enable one for the desired message length.
  Disable to let the LLM do its thing (they usually get pretty yappy if left to their own devices).
• [Introduction] (NEW):
  
  Added an optional intro toggle in case prose style needs stabilization; it's a bit purple-y. You can write more stuff in here or change it in case you want a specific style. The more words/stylish, the better. Off by default.
• CoT Templates:
  
  Switch one of these on to make the model think harder about a certain scene or pattern, if killing the CoT isn't possible, or to 'kill' some models' such as GLM's. Do not combine with the one below.
• Thoonk Killer (Prefill):
  
  Activate to kill the CoT. Precludes the injection above ofc.

Parameters

• I've left most of these default since they can have a massive effect on the output. It's up to you to find the sweet spot for each.
• Streaming: I've set it to on because it seems to be popular, but feel free to change it.
• Context Size: Mostly personal preference. DeepSeek documentation says after 60k you get latency problems, but it can handle up to 1M just fine. For RP'ing purposes I set it to 32k because I usually summarize after that point, plus going overboard causes issues with context poisoning, instructions, etc.
• Response Length: personal preference. No effect on reply length generation (except truncation), it's just to control your billing. Set it lower (like 1k) if you use something expensive like Opus so you don't have a heart attack if the model decides to ramble.
• Temperature: I recommend 0.9-1.1, since the preset is already kind of schizo-inclined. ~0.7-1.3 is the range you should be playing in, with 0.05 increments.
• Frequence Penalty: 0.0-2.0, in 0.1/0.2 increments. Mostly for prosekekery shenanigans. 1.0 is what a lot of people use, I left it at 0.0 since sometimes multiple repeating words or sentence constructions can be fun (eg., tap-tap-tap, "and x and y and z," etc.).
• Presence Penalty: same as above.
• Top P: 0.9-1.0 is what I recommend. Usually something like 0.9-0.95 should keep you safe from a message cascading out of control, but it does kill sudden bursts of creativity: token calculation is cumulative and used on every single step of the way, so in a 200 word reply if you set it at 0.9 those are 20 'esoteric' words you're probably killing (statistics don't work that way but you get the idea). If you adjust temperature you should play with this value as well.
• Top K: 0-100 range. At higher values (~50-100) increases vocabulary breadth on some models. I've set it at 50.

Logs, comparisons, and examples

• DeepSeek v4 (preset v1.2)
  
  Yuki by @raingazer
• DeepSeek v4 / GLM 5.2 (preset v1)
  
  Ae-Ra, draft/unreleased bot by @Astroturf
• More DeepSeek v4, GLM 5.1 (preset beta)
  
  
  
  Paisha by @Astroturf
  Momo by @raingazer
• DeepSeek v4 (preset alpha, very early)
  
  Uroe by @taora
  Akutaro by @Astroturf
  Helvetia by @snombler
  Seija by me, @petNyan :3

FAQ

That's it...?

Yes.

This preset release is mainly to share my findings experimenting with context poisoning, weird sentence structures and writing styles inside definitions. System instructions are based off momoura's.

DeepSeek/X model isn't listening, particularly to word count...

This is a somewhat unstable preset and can get quite wild at the start of an RP.
You'll have to reinforce instructions until you establish a good rhythm/structure. Swipes are your friends.
All the usual rules from using chatbots still apply: if you don't put in some work, you'll get the usual stuff; don't expect different results doing the same thing.

What happens if I grab this preset and turn all the sent as 'AI Assistant' roles into 'User/System' and remove the AI Assistant text?

You'll get a very standard, lightweight preset that will let each model do its thing and write its usual fare of slop standard outputs. Consult the v1 comparison log, which was the only difference between presets.

Is it better to kill the CoT to achieve better prose stylization?

During extensive testing I've found that the answer is a resounding yes, reasoning has a decisively bad effect on prose quality. It almost always ends up leading the output toward a more 'stochastic,' standard result, usually full of very stale phrasing that you start skipping over. These things remain next-token predictors, after all.
If anything, I think it's always worth it to experiment with reasoning turned off when testing out models. It can always be turned on for specially complex scenes, or when you need a specific CoT template.

Cot Templates?

CoT templates. They rule the day in today's day and age of forced reasoning models.
They can have a moderate effect to reinforce prose style, and can help you get out of sticky situations such as very complex logistical scenes involving different locations, size differences, etc.
I highly recommend crafting a template of your own to suit your needs, and trying to make it stick to a writing style so that the next-token weighting doesn't turn the scene too robotic or analytical.

Updates when?

More soon! Probably reformatting with some stuff I've learned, plus improvements some anons suggested.

Updates

• 28/06/26: v1.4 released. Changelog:
  • Added lines to reduce echo/parroting (starting a new chat is recommended).
  • Added a CoT for echo cancellation.
  • Some formatting so it looks nicer.
  • Re-enabled Streaming by default.
• 27/06/26: v1.3 released. Changelog:
  • Parameter changes: Context @32k, Streaming @off, Top K @50, Web Search @off.
  • Schizo Style (Light) is now, well, lighter.
  • Some wording changes to make models like DSv4 more proactive (e.g., using the word 'plot' more).
• 23/06/26: v1.2 released. Changelog:
  • Added CoT Template to 'kill' GLM's reasoning (with help from @Anonymous).
  • Added 'Word Count' toggles.

• 22/06/26: v1.1 released. Added bot links for the logs. Changelog:

  • Cleaned up a bunch of stuff, fixed a few XML tags and semantics.
  • Increased context size to 64k tokens.
  • Fixed 'Chat Examples' wording since it always gets sent as User; I recommend disabling these prompts if the card doesn't have any.
  • Added [Introduction] that can be enabled for added stabilization at the start (feel free to edit it out to your liking; the more words/stylish, the better).
  • Added 'Schizo style (random)' injection.
  • Added preliminary CoT for Sex Anti-Passivity.

  Thank you @Scriptanon for the feedback!

• 20/06/26: v1.0 released.

About

More stuff (bots)
https://rentry.org/petNyan

Contact
petNyan@proton.me