Can't see the images? Turn off your VPN.
Welcome to my SillyTavern setup. The goal is simple: minimize AI slop, exponentially improve your response quality, and make it work right out of the box with zero fuss.
| Setup Overview | Details |
|---|---|
| Target Models | Geared towards big models like DeepSeek or GLM. Untested on smaller/local models, but should work decently. |
| Token Usage | Averages between 5,000 and 10,000 tokens per response (excluding chatbot/lorebook costs). |
| Prerequisites | You are on PC, have a fresh default copy of SillyTavern, and know your way around the basics. |
| Providers | NanoGPT Just pay a 8-10 bucks a month sub and completely forget about any troubles with finding providers that aren't shit. |
My setup is divided into 4 parts: Lorebook | Extensions | Author's Note | Chat Completion (The Lorebook and Extensions are the meat of it. The other two you just set up once and forget.)
1. The Lorebook
Downloads:
How it works:
The Worldinfo entries are set to Constant at Depth 0. This means they activate on every single response, injecting instructions that stop the model from acting like a polite AI assistant and force it into proper roleplay.

Pro-Tip: Turn on the π§ Thinking Mode entry. It gives the chatbots agency, goals, and nuance via a hidden thinking tab.
Note: If you use the Recast extension (below), you don't have to worry about this thinking context diluting your chat, as Recast scrubs it during post-processing. They synergize perfectly.
2. The Extensions
Guided Generations by Samueras
What it does: Lets you force instructions for the chatbot's next response, acting like a scene director.
Setup & Usage:
- In the Extensions menu, toggle on: Show Guided Continue, Guided Response, and Guided Swipe.
- This adds shortcut buttons next to your standard response buttons.


I primarily use it to dictate response length and specific actions. Just drop your director instructions on a seperate line, like so:
Recast by closure / p7error
What it does: Rewrites the bot's generated response through a strict instruction pipeline to kill slop and fix formatting before you even see it.

How to set it up:
- Go to Extension Settings -> Recast.
- Copy the prompts below into separate entries. Order matters: put the main ones first, optional ones last.
- Set the context length to 1 for each prompt.
- CRITICAL: Set the model for these recast prompts to a thinking/reasoning model. It applies fixes much more consistently this way.

Mandatory Recast Prompts
1. Slop Remover
2. Anti-Repetition + Prose Fixer
Optional Recast Prompts
(Try these at your own discretion. Not necessary, but fun to play with.)
Character Behavior Validator
Lewdinator
(I recommend leaving this off and only turning it on when you want spicier NSFW scenes)
Other Highly Recommended Extensions
| Extension | What it does / Why you want it | Links |
|---|---|---|
| ST Memorybooks | Summarizes chat events into a blank lorebook automatically every X messages (I use 50). Essential for continuity in long chats/groupchats. | GitHub Link |
| Worldinfo Recommender | Suggests lorebook entries to add, update, or remove. Tip: Prompt it to be specific on which entries to update for best results. | GitHub Link |
| Rewrite | Highlight a section of text in a response and have the model rewrite just that specific part. | GitHub Link |
| Dialogue Colorizer | Lets you set custom RGB colors for character dialogue. Godsend for groupchats. | GitHub Link |
| TypefaceR | Change UI/dialogue fonts without touching CSS. Works with any font installed on your PC. | GitHub Link |
| World Info Bulk Mover | Move multiple WI entries at once. A massive time-saver for making your own lorebooks. | GitHub Link |
| Narrator | Adds a toolbar button to generate story suggestions. Good backup for writer's block. | GitHub Link |
| ST-Copilot | Gives you a built-in assistant interface that can modify lorebooks, make observations, and more. Highly recommend playing around with it. | GitHub Link |
(Memorybook and WIR Reference Screenshots)




3. The Default Author's Note
What it does: Consistently reinforces instructions alongside the Lorebook.
Setup:
Set your Author's Note to In-Chat, Depth 0, and User. Do this in your default settings so it automatically applies to all new chats without manual tweaking.

The Prompt:
Copy and paste this in. You can change the "Format the next response..." part to suit your own style.
4. The Chat Completion Preset
The easy way: Just download and import the JSON file.
Download WhimsicalPreset JSON here

The manual way:
If you want to do it yourself, it's essentially the default ST preset with a few tweaks:
- Continue nudge:
[Resume in the middle of where this message has stopped. Only finish the rest of this message and do not do anything else (i.e. repeat what's already been written.)] - Character Names: Default
- Squash System Messages: On
- Continue Postfix: New Line
- Request Model Reasoning: Off
- Reasoning Effort: Auto
Closing Notes & Fixes
1. Post-Processing Settings
Make sure your prompt post-processing in the Connection Profile tab is set to Strict user-first with tools.

2. Killing Model Reasoning
This setup is built to prevent models from using reasoning, which ruins long-term RP context. If reasoning is running in the background:
- Use a non-thinking model.
- Go to the Advanced Formatting tab Toggle Auto-parse ON Set the reasoning template to blank.
That's it. I made this preset through blood, sweat, and unending frustration with how terrible most other presets are at maintaining quality. Hope it serves you well. FIN.