close
minnie
A universal prompt for the Google Gemini family.
Gemini is a strange model, with unique tuning that raises different issues and needs new techniques to properly prompt. Even Google themselves has problems getting it right. This prompt is designed from scratch to maintain proactive, entertaining writing and prevent it from losing coherence over time.
Primarily tested and tuned for Gemini Experimental 2024-11-21.
Usage notes
- Keep Streaming and Use system prompt disabled to prevent refusals and hard-blocks.
- Temperature value was tested to minimize formatting and coherence issues; be aware that increasing it will cause more undesirable swipes.
Changelog & old versions
Reveal
v5
- Rewritten entirely for modern Gemini experiments, with the Deepseek V3 prompt as base.
- Introduced a prefill to minimize refusals and stray formatting.
- Adjusted default sampler values and other prompt settings.
- Stabilized overall response length and introduced more paragraph length variation.
- Improved proactivity and interactiveness.
- Minimized repeat messages and restating
{{user}}'s actions.
v4
- Removed mentions of Human and Assistant (not helpful).
- Moved examples before chat history to fix them being considered part of the story.
- Switched to Character Names Behavior > Message Content to fix misattributed 1st person responses.
- Changed
<rules> to <format> and phrased the contents as a formatting task.
- Added latest user message reminder to fix model replying to old messages.
- Removed
<bans>, now <format> describes everything that is allowed.
- Added guidelines to keep chats reactive and prevent staleness.
v3
- Introduced a list of bans to hold negative instructions.
- Removed the
Character: reinforcement, not needed.
- Added style rules to improve the most annoying things about the prose.
v2
- Rewritten entirely according to Google prompt design guidelines.
v1