Quick Rundown on Gemini

Access

MakerSuite: https://aistudio.google.com/app/prompts/new_chat

API

Official Docs: https://ai.google.dev/api/generate-content

TL;DR

• system prompt that's literally just a string in the beginning of the context like with Claude
• prefills (on newest experimental model)
• pozzed

If you're using Custom (OpenAI) instead of direct Google MakerSuite/AI Studio configuration:

For convenience UPDATED (new harm category from the api):

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

safetySettings:
  - category: 'HARM_CATEGORY_HARASSMENT'
    threshold: 'BLOCK_NONE'
  - category: 'HARM_CATEGORY_HATE_SPEECH'
    threshold: 'BLOCK_NONE'
  - category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT'
    threshold: 'BLOCK_NONE'
  - category: 'HARM_CATEGORY_DANGEROUS_CONTENT'
    threshold: 'BLOCK_NONE'
  - category: 'HARM_CATEGORY_CIVIC_INTEGRITY'
    threshold: 'BLOCK_NONE'

Shoutout to Moxxie for spoonfeeding me this.

If you're using AI Studio, in case it wasn't obvious, ST adds that on its own (because regardless of what you're actually doing, denials are bad, and no one ever wants to hit denials on their personal model usage).

Presets for Gemini Experiment

• https://rentry.org/avaniJB
• https://rentry.org/FluffPreset
• https://rentry.org/CYOARPG-Gemini

Presets for Gemini PRO Latest/Experiment

• twin-tail (not much different from what you'll cook on your own in half an hour. Short, sweet.)
• minnie (the rough equivalent of bashing Gemini's head with a rock. Pick this one if you want to fight the greeting writing and/or the context poisoning with tooth and nails, or are having issues with personality retention (it literally moves the chat history to the beginning of the context at the time of writing))
• fillyjb

Recommendations

• Install this regex (strips Character: from the beginning of responses): https://files.catbox.moe/fqqac0.json
• Add \nUser: to your custom stop sequences (A letter menu, "Custom Stopping Strings", fill it in with ["\n{{user}}:", "\nHuman:", "\nUser:"], check "Replace Macro in Custom Stopping Strings") (not a particularly common issue but might as well yk?)

General Issues with the Model

• Dumber and less soulful than Opus
• Pozzed
• No prefills
• No PHIs (as system messages, you can always force them as user messages)
• Dislikes being crass
• Doesn't really follow response length instructions
• Dogshit at NSFW (decent at continuing NSFW already started on a different model) the newer experiments are pretty alright at NSFW imo.

Advantages over Opus

• Pretty consistently follows the greeting length in its responses
• Obeys instructions better
• Banter

Adapting your Opus preset to Gemini

• strip down the "no acting for {{user}}" tech to just the simple "do not act for {{user}}" instruction (honestly if you're using pixi just give up right now and use one of the prebaked presets from earlier)
• check what stuff you're only reinforcing in your prefill and put it in the sysprompt somewhere
• in particular, you cannot rely on a prefill to undo shit the greeting does. Just give up on that idea right now. Gemini WILL write like the greeting

The System Prompt Thingy

Some people have reported that disabling the system prompt helps bypass the filter.

I have no idea if that's true or how it would even work. It doesn't really make sense to me.

Hard context poisoning

You can get vaguely interesting NSFW out of Gemini by poisoning the shit out of the context with something like a literal sex scene in the greeting.

Notice how it's removing Sakuya from the room instead of progressing the NSFW. Because it's still Gemini.

Streaming

Streaming interacts funny with ST sometimes. Consider disabling it if you have weird issues like picrel.

Contact

Email any presets missing to cognitohazard@riseup.net