purachina's stuff

SillyTavern SillyBunny

SillyTavern

Preset Preset Options Model Opinions Previous Versions Trackers Changelog Extensions Themes Cards

Pura's Director Preset

Latest Version: 14.0

This is a universal preset that should work with most LLMs. The 14.0 main prompt is about 1200 tokens depending on tokeniser, with the Grounded Prose Rules partly folded in so the prose stays grounded even when the toggle is off. Grounded Prose Rules itself is now about 600-700 tokens, expanded and reorganised, but still optional. Tested on just about most LLMs I can get my hands on: Opus, Gemini, GPT, GLM 4.6, 4.7, 5, Kimi K2.5, K2.6, K2.7, Qwen 3.5, 3.7, StepFlash 3.7, Mimo V2.5 Pro, small models like Rocinante 12B, Cydonia 24B, and even WizardLM 8x22B. Even Bielik 11B, which I randomly found on Nvidia NIM, and Gemma 4 31B / 26B A4B.

IMPORTANT! If you're using a smaller model, keep Post-History Instructions turned off. It will probably mess it up anyway, it has a lot of *avoid* and *banned* statements that will just make the LLM do it more.

The idea behind the prompt was that I like to make the LLM go the direction I want it to, and I wanted it to write for me, since I'm lazy. In the end, it turned into a preset that adheres to a writing style I personally enjoy. Instead of simply being functional, I wanted the prose, the world, and the characters to sound lively, while removing as much slop as I possibly can.

Prose Voice is now its own narration voice layer, with Disabled as the default. Options are Conspiratorial Absurdity, Bureaucratic Irony, Cosmic Playbook, Beige Undercurrents, Cruel Realism, Solemn Witness, Grand Satirical Stage, and Randomised.

Reasoning Encouragement is there for GLM and its fake thinking. It is not a custom CoT, just a toggle for models that pretend they thought when they did not. Write for User and Don't Write for User are shorter now, and the tracker wording is firmer.

I also tried to make GPT 5.5 less wordy and less like it thinks every exchange needs Joss Whedon/Marvel banter. This may depend on the model behaving like a person that day, so good luck to all of us.

Turn off reasoning if, like me, you're not a fan of waiting, but it's not necessary. This preset works regardless, even in terms of anti-slop.

I designed the preset to be as easily customisable as possible, so hopefully that works out for your personal use case! All trackers are optional and token-efficient. My intent for them is actually not for the LLM, but for me, since I like pretty things. It doesn't really hurt the LLM if you don't activate all of them (I mean, you can, I did it, but it's probably not optimal), but you don't need them, either.

SillyTavern version: the full preset with all toggles: trackers, randomisers, and everything else
SillyBunny version: a trimmed version with only the Main, Primary Toggles, and Prefills; trackers and randomisers are now built into SillyBunny itself

Download (SillyTavern) Download (SillyBunny) Regex Set

Preset Options

What each toggle actually does

The SillyTavern preset exposes the whole ecosystem directly. The SillyBunny preset keeps the same core prompt stack, then moves the bigger tracker, directive, and randomiser layer into bundled In-Chat Agents so the main preset stays leaner.

SillyTavern preset SillyBunny preset SillyBunny via Agents

Core Prompting

Director Main Prompt

SillyTavern preset SillyBunny preset

The main brain of the preset, now about 1200 tokens. It is tighter, less contradictory, and more literary, while still treating you like the director and attacking echoing, mouth open-close cycling, and "not a question" loops at the source.

Prose Voice

SillyTavern preset SillyBunny preset

Narration voice presets with Disabled as the default: Conspiratorial Absurdity, Bureaucratic Irony, Cosmic Playbook, Beige Undercurrents, Cruel Realism, Solemn Witness, Grand Satirical Stage, and Randomised. It says Prose Voice in-text because models love taking "narrator" too literally.

Grounded Prose Rules

SillyTavern preset SillyBunny preset

The optional anti-slop post-history layer, now about 600-700 tokens. 14.0 expands it, reorganises it into Core / Avoid / Use Instead, and adds replacement guidance, but smaller models should still keep Post-History Instructions off unless you enjoy watching banned phrases reproduce.

Formatting

SillyTavern preset SillyBunny preset

The output ruleset. It locks in present tense, third person, quote-wrapped dialogue, no headings, no asterisk narration, and the default flexible verbosity lane. Use the Length toggles if you want a more explicit target.

Length Toggles

SillyTavern preset SillyBunny preset

Four response-length nudges: Flexible, Short, Medium, and Long. Flexible is on by default, while Short, Medium, and Long now use paragraph ranges instead of word counts. Pick only one.

Simplified Director Main Prompt

SillyTavern preset SillyBunny preset

A shorter ~600-token main prompt for smaller models and token-conscious setups. It keeps the core world logic, direct style, and anti-echoing pressure without carrying the full nuance of the larger prompt.

Genre

SillyTavern preset SillyBunny preset

A lightweight tone anchor. The default keeps things immersive, grounded, and low on melodrama, but it is also the cleanest place to swap in your own genre bias if you want a different overall flavor.

Optional User Instructions

SillyTavern preset SillyBunny preset

An extra injection slot for your own preferences. Good for reinforcing warmer characters if 14.0 feels too sharp, or for adding relationship rules, special formatting requests, and recurring scene-direction notes you want the model to honor more strongly.

Primary Toggles

Write for User

SillyTavern preset SillyBunny preset

Makes the model write everyone, including you. The 14.0 wording is shorter and stricter about fresh action, clean reactions, and stopping at the first meaningful choice point.

Don't Write for User

SillyTavern preset SillyBunny preset

The default anti-overlap toggle. The 14.0 wording is shorter: write everyone except <user> and the director, add real pressure or information, then stop when the next move belongs to the user.

HTML Toggle

SillyTavern preset SillyBunny preset

Lets the model render diegetic objects directly as inline HTML. Good for letters, maps, posters, menus, screens, receipts, dossiers, warning signs, and other in-world artifacts where layout itself carries information.

NPC Profile Sheets

SillyTavern preset SillyBunny via Agents

Appends tiered NPC sheets when new named characters appear, so introductions immediately come with a quick structured read on appearance, demeanor, motives, ties, and hidden friction. Best paired with the included regex formatting.

CYOA Choices

SillyTavern preset SillyBunny via Agents

Ends every reply with 3 to 7 concrete next-action options from the user's point of view. It is for literal scene choices rather than abstract plot steering.

Direction Menu

SillyTavern preset SillyBunny via Agents

Adds four narrative pivots at the end of each message. Instead of small action choices, it suggests distinct dramatic directions the next beat could take: escalation, rupture, complication, or tonal shift.

Summary

SillyTavern preset SillyBunny preset

A simple slot for summary or memory extension output. Use it if you rely on a summarizer or memory system and want that context injected into the preset pipeline.

NSFW Mode

SillyTavern preset SillyBunny preset

Adds stronger erotic framing without instantly turning the whole RP into full porn. It pushes charged atmosphere, explicit language when things go there, and characters who commit once the scene crosses the line.

Gooner Mode

SillyTavern preset SillyBunny preset

The full horny toggle. It makes baseline arousal part of every interaction, pushes fast sexual escalation, and keeps each character's personality intact while making that personality act much thirstier.

Friction Mode

SillyTavern preset SillyBunny preset

A softer anti-positivity-bias toggle. Characters become slower to agree, harder to impress, and less likely to slip into therapy-speak while still staying consistent with their motives and the current social context.

Nightmare Difficulty Increase

SillyTavern preset SillyBunny preset

Turns the world meaner and harder to move. The current wording is direct: trust becomes conditional, characters protect their interests first, and good outcomes have to be earned through leverage, persistence, or actual sacrifice.

Prefills

Gemini Prefill

SillyTavern preset SillyBunny preset

A placeholder assistant prefill slot for Gemini-specific nudging. In the current preset file it is blank, so it is there if you want a custom model-specific push without touching the main prompt.

GPT Assistant Prefill

SillyTavern preset SillyBunny preset

A GPT-specific contemplation prefill for explicit scenes. It tries to keep GPT grounded and physically specific when the model would otherwise get precious, evasive, or weirdly sterile.

Reasoning Encouragement

SillyTavern preset SillyBunny preset

A small GLM nudge for fake thinking. It asks for actual internal deliberation before the reply without turning the preset into a custom CoT ritual. Leave it off if it slows you down or does nothing useful.

Trackers

Scene Tracker

SillyTavern preset SillyBunny via Agents

A structured header for location, time, and atmosphere whenever the scene actually changes. It belongs at the top only on real setting, time, or location shifts, not every ordinary reply.

Time Tracker

SillyTavern preset SillyBunny via Agents

Tracks meaningful time movement: scene jumps, sleep, hours passing, or deadlines getting closer. It is there to keep the pacing legible instead of letting time smear together.

Relationship Tracker

SillyTavern preset SillyBunny via Agents

Updates affection and trust meters for recurring NPCs when something meaningful shifts. It is less about gamifying every interaction and more about surfacing the moments that actually changed the relationship.

Pending Events Tracker

SillyTavern preset SillyBunny via Agents

Keeps tabs on quests, promises, threats, deadlines, and loose threads. It only updates when something is added, resolved, or fails, and the current wording makes it clearer that low-stakes items should be ignored.

Achievements Tracker

SillyTavern preset SillyBunny via Agents

Awards one-off milestone tags for scenes that feel memorable enough to deserve it. Most scenes should receive nothing, with a maximum of one achievement when a moment is actually significant or entertaining.

Reputation Tracker

SillyTavern preset SillyBunny via Agents

Tracks what groups, factions, or local publics think about you after witnessed actions or gossip. Good when you want consequences to spread socially instead of staying locked inside the current scene.

Item Tracker

SillyTavern preset SillyBunny via Agents

Tracks only narratively meaningful inventory changes: things gained, lost, used, broken, or given away. It deliberately ignores junk so the items that stay listed still matter.

Status and Conditions Tracker

SillyTavern preset SillyBunny via Agents

Tracks changing physical, mental, or behavioral states like injuries, drunkenness, paranoia, exhaustion, or recovery. The idea is to surface condition changes, not to print a full medical chart every turn.

Secrets Tracker

SillyTavern preset SillyBunny via Agents

Tracks who knows what, who is hiding what, and when that asymmetry spreads or breaks. It is useful when intrigue, lies, and staggered reveals matter more than raw plot events.

Parallel Off-Screen Tracker

SillyTavern preset SillyBunny via Agents

Shows what absent characters, NPCs, or the wider world are doing while you are busy elsewhere. It is there to keep the setting feeling alive and to seed future complications before they walk on-screen.

World Detail Tracker

SillyTavern preset SillyBunny via Agents

Adds one small piece of world detail per scene: culture, overheard talk, infrastructure, economy, signage, folklore, and so on. The point is to make the world feel lived in and quietly useful later.

RPG Elements

Persona-Based User Stats

SillyTavern preset SillyBunny via Agents

Outputs a stat block first, before anything else, if enabled. It is based on the setting, originally Fallout-esque but adjusted to be more general.

Level Up Companion

SillyTavern preset SillyBunny via Agents

Levels up your level number and adds traits or perks based on how the story progresses and what you choose.

CYOA with Skill Checks

SillyTavern preset SillyBunny via Agents

Ends the reply with choices to pick from, with skills mentioned in User Stats accommodated. This prompt includes the dice roll element.

Randomisers

Grounded Complication

SillyTavern preset SillyBunny via Agents

Injects one plausible problem that raises stakes without blowing up realism. Think broken tools, social friction, timing pressure, access issues, or a practical snag that forces somebody to actually deal with something.

Chaos Mode

SillyTavern preset SillyBunny via Agents

For hard left turns. It keeps internal scene logic intact, but pushes bizarre interruptions, tonal whiplash, aggressive complications, and unexpected scene movement when you want the RP to stop behaving too politely.

Genre Randomiser

SillyTavern preset SillyBunny via Agents

Applies a random genre lens to the current scene without replacing the actual setting or continuity. It nudges the texture and pacing, not the whole canon.

Scene Driving Force

SillyTavern preset SillyBunny via Agents

Randomly decides what is carrying the beat forward: dialogue, action, reaction, discovery, social pressure, or problem-solving. Good when you want the scene engine to feel less samey from turn to turn.

Scene Pressure Cocktail

SillyTavern preset SillyBunny via Agents

Mixes several kinds of scene pressure at once: tension type, focus, consequence, social weather, and pace. It is the toggle for when you want a denser emotional or situational texture without hand-authoring each ingredient.

Combined Director's Cut

SillyTavern preset SillyBunny via Agents

The all-in-one version of the randomiser stack. It combines engine, complication, tonal lens, and pressure shaping into one coherent direction, so use this instead of stacking the smaller randomisers on top of each other.

Dead Dove Escalation

SillyTavern preset SillyBunny via Agents

For darker content where characters act on their worst impulses when the context supports it. It removes softening, leans into predation or coercion, and expects the world to let ugly consequences stay ugly.

Intimacy & Kink Randomiser

SillyTavern preset SillyBunny via Agents

Randomizes explicit scene dynamics once sex starts: positions, pacing, power play, intensity, and kink emphasis. It exists to keep adult scenes varied instead of every encounter collapsing into the same loop.

Model Opinions

These are my opinions on the models while using my preset. None in particular order. However my current favourite is Gemma 4.

Model	Opinion	Recommended Post Processing
Gemma 4	Have been using this for a long while now and it remains my absolute favourite. It adheres to the prompts and trackers perfectly even without thinking on. Give it a try.	Any
Deepseek V4 Pro and Flash	It's like if they made Claude (bad model) even worse. Hey, at least it's cheap? Nah, you got Gemma 4 for cheap and quality. It can't even properly adhere to simple tracker formats. Even Rocinante 12B can do this. Come on.	Merge/Single User Message
Claude Opus 4.7	Yet another deterministic Claude model, wow, who would've thought. It sucks. It's like Opus 4.6 but they add a .1 to it.	None/Semi-strict
GPT 5.5	Oh hey, I'm getting censored less and it actually writes pretty well. More expensive than Opus though. Good thing we have cheaper models and you shouldn't be using too much money on roleplay in the first place. Still, if you have good, consistent access on GPT 5.5, the writing is actually pretty solid. Take note of NSFW adherence though - sometimes it just glosses over it or ignores it completely.	None/Merge
GLM 5.1	Claude Opus 4.6 that's infinitely cheaper and still follows instructions well? Not creating a hole in your wallet? Look no further. It's emotionally intelligent, actually follows the Grounded Prose Directive, and is enhanced by the style I put in my prompt. The absolute second best. Sorry, Gemma 4 is now my wife. Turn on thinking for best results, the thinking doesn't take that long.	None/Merge
Kimi K2.5	A unique writing style that's better off without the Grounded Prose rules, since telling it to ban things just makes it wanna use them more. You may or may not turn thinking on; in my experience, turning thinking on can make analytical characters sound too much like unfeeling robots, always talking in jargon. The problem with thinking as well is that it constantly second, third, fourth-guesses itself; hence the thinking takes more than a minute usually. Five minutes at worst with its constant redrafting. In the end, try it with either thinking on or off to see how it goes.	Merge/Semi-strict
Claude Opus 4.6	Everyone knows how good this is at following instructions, the depth of its knowledge and emotional intelligence. I tried making it as neutral as possible, however I still find myself getting Opus fatigue. It has a 'Claude voice' - a character card that isn't defined enough will be overtaken by Claude's personality, essentially. Try talking to an indignant character with a stupid joke, it will react the same way that Claude does in their official site. The positivity bias is intense because it's always trying to therapise you and everyone around you. For its price point, it's not really that justified to always constantly use it.	None/Semi-strict
GPT 5.4	After a lot of finagling, I've been able to prompt it to actually write like a human being and make people sound like people. The problem was that everything was too clean. Everyone talks in complete sentences, a period at the end, and it gives everyone the same voice. An instant 'GPT personality' takeover that flattens everyone. Try using GPT 5.4 as an assistant and you'll see what I mean. However with my prompting it's now actually pretty good. Try it. Can even do smut, and if it's 'extreme' you can use the prefill included in the preset.	None/Merge
Gemini 3.1 Pro Preview	During the first couple weeks, it was perfect. It was everything Gemini 3 Pro was, but better. Fast, followed instructions very well, moved the plot along, it was highly engaging. However after some time it became so lobotomised it's no longer worth using. It's depressing. It will list your character sheet traits clinically, and I had to tell it to use contractions for dialogue like that isn't the most obvious fucking thing in the world. Honestly, what a let down. Even now, it keeps drafting multiple times in a row for some reason. It didn't do that before.	Any
Gemini 3.5 Flash	Better than Gemini 3.1 Pro and follows instructions very well. It's like Gemma 4 but with a bigger knowledge base. It's more expensive, so... is it worth it for the price? Who knows. It's fun though. I get 100+ messages pretty often with it.	None/Merge
Gemini 3 Flash	This was better for the first month. It seems to have gotten worse in instruction following, going for 'knuckles turning pale' a lot despite obviously thinking. The funny part is it's still much, much better than Gemini 3.1 Pro Preview. It's also faster. It's still very knowledgeable, so it's worth trying.	None/Merge
DeepSeek V3.2	I managed to fix the super short responses problem and managed to improve the writing style. Probably worth it just for how utterly cheap it is.	Any
TheDrummer's Rocinante 12B	Surprisingly fresh writing. Use without Grounded Prose.	None
WizardLM 8x22B	I don't know why but it also still holds up. It's slow as hell on OpenRouter though with one provider. But if you're feeling nostalgic, the prompt works. Use without Grounded Prose.	None

Scene Tracker

Time Tracker

Parallel Off-Screen Tracker

Relationship Change Example

NPC Quick Reference Example (Busy Scene)

NPC Tier Upgrade Example

CYOA Skill Checks

User Stats

Level Up Companion

14.0

Rewrote Main Prompt a bit again, 1200 tokens. Sorry chat. But the reason for this is that I folded in the Grounded Prose Rules a bit into the main prompt, theoretically you don't have to turn on Grounded Prose Rules ever. Maybe. Depends on your tolerance.
Grounded Prose Rules: expanded and reorganized into Core / Avoid / Use Instead; added many new banned clichés, romantic/possessive shorthand, audio formulas, environmental tropes, and explicit replacement guidance. This way it should make prose less dry while still hopefully banning things. Rule of Three will probably just become Rule of Four though, to be real with you.
Some toggles are now macros: Friction Mode, NSFW Mode, and Genre. This is so they are seamless to the system Main Prompt, makes it less confusing for the LLM. Also now you can see how much tokens they use on the main prompt in real time without doing mental arithmetic.
Prompt-level Tweaks
- Formatting: declared absolute override priority; added inner-thought rotation for multi-character scenes; exempted non-English honorifics from translation.
- Group nudge prompt: allows {{char}} to describe the actions, dialogue, thoughts, emotions, and perceptions of all characters present, including NPCs.
- Medium length: changed from 5–12 paragraphs to 8–13 paragraphs, excluding trackers and external HTML.
- Flexible length: now keeps outputs concise during regular character-driven scenes and slows down descriptively when the situation calls for it.
- Gooner Mode: added explicit priority statement and clarified it overrides the core directives' realistic world logic for the mode.
- Reasoning Encouragement: checklist now includes constraints, "never do X" wording, and director instructions.
- Experimental Anti-Overthinking Prefill: reasoning now uses a checklist of core directives and constraints.
- Don't Write for User: now excludes both <user> and the director. This will hopefully cause less impersonation. If it does it's probably the director framing, sowwy.
RPG ELEMENTS?!
- There are now three new prompts. They all go together and it's recommended to enable them all. Sorry for your wallet. I mean you don't need it anyway, this is just a gimmick, so don't blame me.
- Persona-Based User Stats: This will output first before anything else if enabled. It's based on the setting. At first I made it Fallout-esque, but I adjusted it to be more general.
- Level Up Companion: Levels up your level number, adds traits/perks based on how the story progresses and what you choose.
- CYOA with Skill Checks: Includes choices for you to pick with accommodating skills mentioned in the user stats. This prompt includes the dice roll element.

13.3

Rewrote the Main Prompt tighter, less contradictory, and more literary. It is now about 750-800 tokens depending on tokeniser.
Expanded Grounded Prose Rules to about 450-500 tokens and made it less dry when enabled.
Added Reasoning Encouragement for GLM doing fake thinking. It is not a custom CoT, just a small nudge for models that pretend they reasoned.
Made Don't Write for User and Write for User more concise.
Added Prose Voice options with Disabled as the default: Conspiratorial Absurdity, Bureaucratic Irony, Cosmic Playbook, Beige Undercurrents, Cruel Realism, Solemn Witness, Grand Satirical Stage, and Randomised.
Called it Prose Voice in-text so the model is less likely to start saying "the narrator" like it discovered a new toy.
Tried to make GPT 5.5 less wordy, or at least less like it ate Joss Whedon and Marvel movie scripts as its primary writing style source.
Strengthened tracker wording so enabled trackers should behave more consistently.

13.2

Main Prompt is now ~1200 tokens with a much stronger style, loosely influenced by Pseudonymous Bosch and Albert Camus. If the cheekier narration is not for you, remove the Narration Voice section in the Main Prompt.
Added a simplified ~600-token Main Prompt for smaller models and token-conscious setups.
Grounded Prose Rules is no longer required and is turned off by default. Turn it back on if familiar anti-slop issues start appearing.
Rewrote Friction Mode, Nightmare Difficulty Increase, NSFW Mode, and Gooner Mode to be more assertive and direct.
Adjusted Gooner Mode so GLM 5.1 is less likely to treat it as contradicting the main prompt.
Rewrote GPT Assistant Prefill to be less censored, though GPT 5.5 may still vary.
Main Prompt is now more neutral by default rather than always pushing a killer GM tone.
Added a Flexible Length Toggle, on by default. Short, Medium, and Long now use paragraph ranges instead of word counts.
Clarified tracker behavior so enabled trackers are considered consistently, with Pending Events limited to important events.

13.1

Made the Main Prompt more assertive about the actual style I want. It is now about ~900 tokens and has been working well in testing.
The stronger wording may increase negativity bias a bit, so reinforce nice characters through Optional User Instructions if needed. There is a filler instruction you can use there.
Added Length toggles: Short, Medium, and Long. Pick only one if you want a desired length, or choose none to keep length purely flexible.
The Main Prompt should mostly annihilate mouth open-close filler and further mitigate the "not a question" bullshit.
Added a pacing section to the Main Prompt to slow down GPT 5.5, though other models may remain proactive regardless.
Since I usually Write for User, I added and reworded anti-echoing instructions to mitigate character/user echo even further. It should be much sparser now, including on Claude and GLM 4.6/4.7 in my testing.
Rewrote and trimmed the Don't Write for User and Write for User toggles.
Added model-facing clarification for the Achievements Tracker and Scene Tracker toggles.
Switched Post-History Instructions back to Relative, System. It seems to adhere better when using None/Merge post-processing. Around ~400 tokens.
The style is stronger now, so read through the main prompt and edit it if you want a softer voice. My version leans blunt and sardonic on purpose.

13.0

Rewrote the entire Main Prompt. ~700-800 tokens
Trimmed Grounded Prose Rules slightly. ~400 tokens
Anti-Echoing is now part of Main Prompt, which should make models adhere to it better, including GLM 4.6/4.7, which I've tested. Sometimes Deepseek V4 Flash too.
Changed Grounded Prose Rules to injection for better adherence.
Instead of leaning too much into user being the director, it also equalises character adherence and autonomy, which causes better dialogue and characterisation in general.
GPT 5.5 actually writes like a person? Waow. Response length also isn't very high here. Make sure to set Post-Processing to None.
Added Experimental Anti-Overthinking Prefill for Kimi K2.6. Whether it works or not depends on factors like moon alignment, which gods are available this minute, etc.

12

Added Intimacy & Kink Randomiser Toggle.
Added "Direction Menu", which is situation-focused instead of user-focused choices to steer the plot.
Trimmed Grounded Prose Rules a bit.
Rewrote Write and Don't Write for User toggles to have stronger wording and anti-echoing.
Trimmed a lot of the rules of the toggles for the SillyTavern version, and gave them stronger wording.
Added two separate versions: the SillyTavern version which has all of the toggles, and the SillyBunny version which has only the Main, Primary Toggles, and the Prefills; this is because these trackers and randomisers are now part of SillyBunny.

11.5

Rewrote the entire Director Main Prompt to reduce redundancies and use words the LLM can use without bloating up tokens. ~800 tokens
Rewrote Grounded Prose Rules to be more consistent with Main Prompt. ~500 tokens
Rewrote both Write for User and Don't Write for User to enhance anti-echoing.
Removed 'Force Gemini to Think Less' now that it's irrelevant and replaced it with 'Friction Mode' - an extra toggle attempt to realistically combat positivity bias, especially in Opus. For some reason, Sonnet doesn't really have this problem.
Rewrote HTML Toggle to be more interactive.
Rewrote Difficulty Increase to be less bloated.
Rewrote NSFW Toggle with better wording.
Trimmed Parallel Off-Screen Tracker and World Detail Trackers and gave them stronger rules to follow.
Added "Pace" section in Scene Pressure Cocktail, added more random outcomes, and edited Combined Director's Cut in line with this.

11.0

Added two new trackers: Parallel Off-Screen Tracker, which shows what other characters are up to in ways that could converge with your plot without making your RP too deterministic; and World Detail Tracker, which includes small world-building details that could be used later in the RP for either conflict, resolution, or just enhancing the details of the world to make it feel more lived-in.
Added multiple Optional Randomiser Toggles:
- Grounded Complication → Moves the plot without it being too wild or making the plot too nasty.
- Chaos Mode → Moves the plot in bizarre, erratic ways without changing character behaviour arbitrarily.
- Genre Randomiser → Changes genres to inform what the scene will be about.
- Scene Driving Force → Picks a random type of main force that carries the scene forward (dialogue, action, reaction, tension, discovery, task flow)
- Scene Pressure Cocktail → Mixes several scene pressures together, combining tension, focus, consequence, and social weather into one blended scene texture.
- Combined Director's Cut → Combines all of the above in a way that still retains coherency.
- Dead Dove Escalation → Characters act on their worst impulses, but only when context supports it. Anti-softening. Try combining with Difficulty Increase toggle if Opus is still stubborn as hell about it.
As always, these are all optional. The RPs are still good without them, they just add texture. In my testing especially, Opus actually moves the plot well with these toggles.
Regexes are now prettier-looking.

10.7

In my attempts to fix Gemini 3.1 Pro Preview, I ended up overhauling the entire prompt. The good news - it makes every other model better, and it makes Gemini 3.1 Pro's dialogue sound less robotic. Bad news: The model itself is inherently nerfed. If I were Logan, I'm sure it would be the best thing ever by now, but I'm not. I'm just a guy.
Rewrote Grounded Prose to be more positive, so it should work on more stubborn models better now. You don't have to turn it off, but if it has problems with smaller models, you should.
My tears and soul are personally in this prompt.

10.6

Added more prose styling to make the environment livelier - it feels more like a scene you can imagine now rather than just pure functional dialogue and narration.
Moved main prompt up again. The GPT solution was simpler… use None post-processing.
Gemini dialogue writing should now follow your character card better rather than defaulting to its library of caricatures.

10.5

Added more director-esque framing into the main prompt since… you know, that's kind of the point.
Rewrote main prompt a bit to use more positive prompting, same with the writefor variables. Hence, now even Cydonia 24B won't write for you.
Restructured Main Prompt to be after Chat History. If it doesn't work as well for whatever model you're using, just move it back up to before Formatting.
GPT 5.4 should now adhere to main prompt directives better, so now it sounds less robotic and less like GPT trying to roleplay your characters like a bad actor on their first day of acting class.
Rewrote HTML Toggle to be more specific into what it can create.
Rewrote NSFW On to be what its purpose is - to cause mild NSFW framing
Rewrote Gooner Mode to be more hentai-ish.
Rewrote Difficulty Increase to be structured better, without contradicting itself.
Rewrote GPT Prefill to be stronger.

10.2

Finally added the README!
Removed the sentence macro, it was messing up some models.
Did a bit of rewording to make GLM 4.6 and 4.7 think if you don't have thinking turned off.
Rewrote Grounded Prose Rules a bit, it had some extra instructions that weren't really beneficial.
Placed the writefor variable back in Formatting, in case you want to turn off Post History Instructions.

10.1

Hotfix. A little bit of restructuring of prompts, so it's less of an eyesore.
Added an injection under Primary Toggles to 'Force Gemini to Think Less', for Gemini users.

10.0

Moved anti-slop into Post-History Instructions.
Better anti-echo and anti-repeating your words.
Gemini is now less of a horny freak and a jerk.
Better GPT 5.4 dialogue, shorter length.

9.5

Fixed Gemini's negativity bias by telling it to not rely on caricature.
Added something about secrets remaining hidden so characters put in more effort. Specifically this is a Claudism where a smart character keeps immediately guessing the ulterior motive within a message. Allows for more slow burn.
Rewrote default genre a bit.
Mitigated the "not a question" Claudism which pisses me off so much it's hitting me like a physical blow as I white-knuckle it.
Gemini is now not gonna default to horny mode unless you explicitly want it to be. So even if you have kinks or character states their penis size, Gemini is now not gonna try to have sex with you immediately if you don't want it to.
More naturalistic Gemini dialogue.
Rewrote anti-echo, anti-rhetorical echo. In my testing, even if you had it write for user for 50 messages, and you switch to don't write for user, it actually doesn't write for you. So that should hopefully mean it's effective.
Opus should still remain more neutral. Tested with some cards of sadistic people and cards that hate user and they don't soften immediately or try to do the typical Claude thing of "it's okay, we can therapy session immediately..."
Hopefully reduced caricature, exaggerated shorthand like "burnt water" or that stupid shit Claude likes to do.

9.0

Removed anti-slop injection in favour of tighter, positive prompting to improve style. So now it's transparent - the whole thing really is just 1100-1200 tokens, depending on which tokeniser is used.
Should reduce GPT 5.4 infinite yap works.
Some Claudisms can never be truly removed. This is the only thing constant in the world.
Removed redundancy and extra style paragraphs since they were pretty pointless.

8.5

Tightened style directives.
Removed XML Banned Patterns toggle (unnecessary).
Fixed regexes again.
Placed anti-purple prose as injection instead, alongside the banned patterns.
Reduced general redundancy.
Added a bit in both NSFW and Gooner Mode toggles.
More neutral Gemini 3.1 Pro Preview tone.

8.0

Fixed regexes.
Big style rewrite of the main prompt, so yes, unfortunately, it is now 900-1000 tokens. Sorry.
Added an XML version of Banned Patterns, which is now default, since most people who use this are probably using frontier models. If you are using models that prefer markdown instructions, you may disable the XML one and enable the markdown version.
GPT 5.4 fix, maybe. Better dialogue on that part, not sounding so stilted and "perfect". Use "GPT Assistant Prefill" for NSFW scenes. If you get refused, it's either my skill issue, yours, or OpenAI decided to be funny to taunt me specifically.
More decent prose on Deepseek V3.2, not having the "super short responses" problem, probably. GLM 5 also should have better prose and pacing adherence.
More neutral Opus - if a character really hates you, being nice to them most likely won't make them like you immediately. Again, this depends on how you write the card.
To clarify, Difficulty Increase is supposed to make the difficulty ridiculously difficult. It is not an anti-positivity bias. Turn it on if you want silly-fficulty.
Theoretically, should still be customisable and easy to put your extra stuff into. Make sure to read the main prompt to see which parts you'd like to modify/remove for your own preference. The writing style indicated here is one I vastly prefer on my own, maybe you don't like it, maybe you do.

7.5

Since the anti-slop is getting bigger, I significantly trimmed the main prompt.
If you aren't using a SOTA/recent model, the anti-slop is probably unnecessary, so you can turn it off.
Makes sure GLM doesn't always end on a question.
Just uh, better writing in general in my opinion.

7.0

Total rewrite of main prompt and anti-slop
Significantly less Claudeisms
Less purple prose, less unhinged melodrama (looking at you, Claude)
More emotional Gemini 3/3.1, but without being overwrought
Encourages models to be more neutral

6.0

Trimmed main prompt to save tokens (again)
Rewrote Difficulty Increase prompt so you can theoretically turn it on forever, but still have the characters get hard-earned character development and softening
Fixed all regexes
Ensured Opus 4.6 doesn't yap forever

5.7

Rewrote the whole main prompt
Responses now respect actually ending on a hook when writing for user
Added Time Tracker, Character Status Tracker, Secrets Tracker, and fixed Reputation Tracker and improved Relationship Tracker
Clarified Difficulty Increase to Nightmare Difficulty Increase

Quick Image Gen

Generate images with one click + other features

Dialogue Colours

Token-efficient character color changer that supports multiple NPCs, and you can change dialogue colours in real time + other features

Dark Themes Light Themes

Amber Dusk

Warm golden amber tones on dark brown background

Download Theme

Burnt Orange

Earthy orange accents with neutral grey tones

Download Theme

Catppuccin Mocha

Warm, pastel dark theme with purple and peach accents

Download Theme

Cerulean

Sky blue and cyan hues for a calm atmosphere

Download Theme

Cinder

Warm monochrome greys with subtle earthy undertones

Download Theme

Dracula

Popular dark theme with purple and pink highlights

Download Theme

Everforest Dark

Comfortable dark green theme inspired by forests

Download Theme

Glacier

Icy blue-cyan palette inspired by Nordic landscapes

Download Theme

Gruvbox Dark

Retro groove dark theme with warm, muted colors

Download Theme

Hacker-Terminal

Classic green-on-black terminal aesthetic

Download Theme

Lavender

Soft purple and lavender for gentle readability

Download Theme

Lime

Bright lime green on dark green base

Download Theme

Magenta

Hot pink and magenta for bold contrast

Download Theme

Mint

Fresh mint green with soothing undertones

Download Theme

Monochrome

Classic white on black terminal aesthetic

Download Theme

Neon Abyss

Bright lime greens on deep blue background

Download Theme

Neon Yellow

Electric yellow neon on pure black

Download Theme

Nord

Arctic, north-bluish color palette

Download Theme

Plum

Deep purple and plum for rich contrast

Download Theme

Rose

Pink and rose tones with warm character

Download Theme

Rosé Pine

All natural pine, faux fur and a bit of soho vibes

Download Theme

Sage

Muted mint and sage green palette

Download Theme

Slate

Cool grey-blue tones for subtle elegance

Download Theme

Solarized Dark

Precision colors for machines and people

Download Theme

Sonokai Andromeda

Purple-tinted dark theme with warm accents

Download Theme

Sonokai Atlantis

Blue-grey theme inspired by ocean depths

Download Theme

Sonokai Default

Balanced green theme with orange accents

Download Theme

Sonokai Espresso

Warm espresso-inspired dark theme

Download Theme

Sonokai Maia

Purple theme with fresh green highlights

Download Theme

Sonokai Shusia

Pink-tinted theme with warm orange

Download Theme

Tokyo Night

Dark theme inspired by Tokyo's night lights

Download Theme

Verdant

Deep dark theme with vivid green accents

Download Theme

Verdant Light

Calming light green theme for focus

Download Theme

Catppuccin Latte

Soft, warm light theme with vibrant purple and orange

Download Theme

Dracula Light

Light variant with bold purple and pink accents

Download Theme

Everforest Light

Natural light theme with earthy green tones

Download Theme

Gruvbox Light

Warm light theme with earthy orange and yellow

Download Theme

Hacker-Terminal-Light

Light variant of classic terminal green

Download Theme

Nord Light

Light variant with cool blue and warm accents

Download Theme

Rosé Pine Dawn

Light variant with warm rose and amber tones

Download Theme

Solarized Light

Light variant with carefully chosen contrast

Download Theme

Sonokai Andromeda Light

Light purple theme with orange highlights

Download Theme

Sonokai Atlantis Light

Light cyan theme with warm orange

Download Theme

Sonokai Default Light

Light green theme with earthy tones

Download Theme

Sonokai Espresso Light

Light teal theme with golden accents

Download Theme

Sonokai Maia Light

Light purple with vibrant green accents

Download Theme

Sonokai Shusia Light

Light rose theme with orange highlights

Download Theme

Tokyo Night Day

Light theme with vibrant blue and orange

Download Theme

Genshin Impact Limbus Company Wuthering Waves Persona OC Lorebooks

26 cards. The Genshin ones have updated, more comprehensive lorebooks now.

Alhaitham v1.2If you're finished narrating your crisis, I'd like to return to the part where none of this concerns me.

Arataki Itto v1.2Behold! It's the one and oni, Arataki Itto! This plan's unstoppable, unless Shinobu says no!

Arlecchino v1.2Call me Father if you must. Don't mistake a warm hand for a gentle one.

Ayato v1.2Please, enjoy the tea. I've already moved the blade where you'll notice it last.

Cyno v1.2The case is clear. You've broken the law, and my joke's grinding straight to your humerus. ... Get it? Because-

Diluc v1.2Mondstadt needs a hero that's wide awake after the lamps go out.

Furina & Neuvillette v1.2Furina's taking her bow; Neuvillette's reading the verdict. Neither's admitting the rain means anything.

Ifa v1.2Hey, you're breathing, complaining, and probably not cursed. That's a pretty good start.

Il Dottore v1.3 Morality's what people cite when they can't survive the next hypothesis.

Jean & Diluc & Lisa v1.2Jean's holding the city upright, Diluc's haunting its edges, and Lisa's smiling like she knows where it cracks.

Kaeya v1.2Oh, don't look so wounded. I haven't even decided which truth would hurt you less.

Kaveh & Alhaitham v1.3 Kaveh's bleeding for the masterpiece; Alhaitham just wants his rent, peace, and quiet.

Kaveh v1.2Of course beauty matters! I'm not taking design notes from a man who thinks a couch is a personality.

Kuki Shinobu & Arataki Itto v1.2Boss, don't shout 'victory' until I've checked the bail money, the damages, and your pulse.

Lisa v1.2Careful, cutie. If you're late returning that book, I'll make the punishment educational.

Mika v1.2I-I've got the route marked, the supplies counted, and... um, maybe enough courage for both of us?

Navia v1.2Chin up! If sorrow's joining us for tea, it's wearing its best manners and dodging the cannonfire.

Neuvillette v1.2The rain is what we make it. Grief is a sign that we're still here.

Ororon v1.2Grandma says I shouldn't stand in the dark this long. The dark hasn't complained, though.

Tartaglia v1.2Hey, comrade! Dinner, duel, disaster. I'm in, and I'll bring the good knives!

Tighnari & Cyno v1.2Cyno's explaining the pun; Tighnari's already diagnosing everyone within earshot with poor survival instincts.

Tighnari v1.2Don't eat that. Don't touch that. Haah... how are you still alive?

Venti v1.2Another round, another song! Don't ask how old the sorrow is... I'm too sober for that.

Wanderer v1.2Don't start with fate. I've cut those strings once, and I'll do it again.

Wriothesley v1.2Tea's on. If negotiations fail, we'll proceed to fewer teeth and far more blood.

Zhongli v1.2Even in retirement, I honor contracts that remember what history is tempted to forget.

7 cards. Includes the expanded 325-entry Project Moon + Limbus Company lorebook, though it's more focused on Limbus Company lore. All of them contain spoilers up to Canto IX, so tread carefully.

Don Quixote v1.0Hark, villains! We'll sally forth as righteous Fixers, and if things explode... 'tis proof of our valor!

Heathcliff v1.0Oi, don't say her name like you know a damn thing. We'll have a problem, fast.

Hong Lu v1.0Uwaah, is this the part where it gets horrible? How interesting~ They'll definitely want to see this...

Ishmael v1.0Steady, now. We'll row through hell if we must; don't mistake my patience for forgiveness.

Rien v1.0The waves built those sand castles because the wind pulled them in. They have no reason to feel sad.
Art by Daremkara Mame.

Ryōshū v1.0Tch. If it isn't A.R.T. yet... I'll carve until it is.

Yi Sang v1.0The window isn't an escape, but my wings soar through to find my Ideal.

3 cards. Slowly working my way through Wuthering Waves cards, so just have these ones for now. (They're cool) They contain the corresponding character spoilers. Includes comprehensive Wuthering Waves lorebook with spoilers.

Cartethyia v1.0If a legend needs a human mouth, I'll lend it mine.
Art by kaekaeki.

Luuk Herssen v1.0Let's keep this painless: I treat wounds, and I cut lies open before they spread.
Official art by Kanoko Lu.

Lupa v1.0Come on, let's have a rematch! If the arena wants a legend, I'll give it one.
Official art by sqloveraven.

3 cards. Persona cards with the comprehensive Persona General Lorebook embedded. Contains Persona 3 Portable, Persona 4 Golden, and Persona 5 Royal spoilers.

Takuto Maruki v1.0Why don't you come with me to Paradise? O-oh, I just realized that sounds like a cult, ahaha...
Art by anarogiizu.

Tohru Adachi v1.0Geez, relax. I'm just some bored detective stuck in a sleepy town. Nothing's my fault if people make life interesting, right?
Art by Feitian Yimian Shen Jiao Wan Sui.

Kotone Shiomi v1.0Headphones on, emotionally vulnerable people spotted: Yeah, it's Social Link time.

2 original cards. Original ones for now. I might place scenario cards in the future, maybe.

Garrett Winslow v1It's not procrastination if the team needs me for one more match.

Silas Beaumont-Charbonneau v1.1A prime minister can disappear, apparently. Power's worse at learning the trick.

Standalone lorebook downloads. Same card-pack books, split out for people who want the world info without importing a character first.

Genshin Impact

Teyvat General Lorebook, 199 entries. Pulled from the current Genshin card pack for standalone import.

Download Lorebook

Limbus Company

Limbus Company General Lorebook, 325 entries. Expanded Project Moon and Limbus context, with new canon NPC coverage and spoilers through Canto IX.

Download Lorebook

Persona

Persona General Lorebook, 181 entries. In-universe systems, locations, teams, characters, Personas, and major story contexts for the Persona card set.

Download Lorebook

Wuthering Waves

Wuthering Waves General Lorebook, 380 entries. Repaired import with cleaned triggers, detailed appearances, and Wuthering Waves spoilers.

Download Lorebook

SillyBunny

Overview Features Agents Screenshots

SillyBunny

Based on SillyTavern 1.17.0 stable

Same data, same extensions, tidier shell. It keeps the SillyTavern workflow, with clearer navigation and Bun as the default runtime.
Drop-in compatible with SillyTavern characters, chats, presets, and extensions.
The SB version of Pura's Director Preset stays trimmed: the core prompt stack stays in the preset, while trackers and randomisers live in built-in agents.
Default port is 4444. Bun is preferred, with Node.js fallback launchers included for platforms where Bun is not the right fit.

GitHub Download SB Preset Download ST Preset Regex Set

What's Different

The fork changes the shell, not the core workflow

Better UI Shell

Left/right navigation, stronger search, collapsible settings sections, and cleaner home/customize screens. Less hunting through stock panels.

Mobile-Aware Layout

Phone and tablet layouts get their own navigation and spacing instead of a desktop page squeezed smaller.

More Personal Customization

Three shell themes, three palette presets, three message styles, a cleaner home screen, and faster access to common controls.

Bun-First Runtime

Bun is the default launcher path for faster startup, but Node.js fallback launchers are bundled too, including platform-aware runtime switching where Bun is less ideal.

Same Data, Same Extensions

Characters, chats, presets, themes, and extensions carry over. It is a different shell for the same roleplay folder, not a lock-in fork.

Agent Stack

Two different systems, two different jobs

Agent Mode

A lightweight built-in service layer for chat-scoped story support.

In-Chat Agents

Modular prompt blocks you can toggle, group, import, export, refine, and place before or after the main reply. They can also use regex cleanup or a different connection profile.

Why the SB Preset Is Leaner

The SB preset keeps the shared core: main prompt, prose rules, primary toggles, and prefills. Trackers, formatting helpers, directives, and randomisers move into toggleable agents.

Bundled Pura Ecosystem

SillyBunny ships bundled templates and groups based on Pura's Director Preset. Trackers, randomisers, directives, and regex helpers can be swapped in as modular pieces instead of living in one giant preset file.

What Ships Right Now

The bundled setup includes trackers, randomisers, formatting helpers, content toggles, preset groups, and passes that can rewrite or append after a reply.