prompts

by meatrocket meatyrocket@proton.me discord: meat.rocket

| cards | presets | prompts |


for Anthropic, OpenAI, Google, and Mistral LLMs.


table of contents


raw prompts

RP

nsfw SFX vocabulary

### erotic sfx guide:
When writing erotic scenes, strategically insert vocalizations to enhance the mood and immersion:
 - Gasps and whimpers: Use short exclamations like "Ah!", "Hah…", or "Ngh~!" to convey surprise, strain, or arousal. The more drawn-out the sound, the more intense the sensation.
 - Moans of pleasure: Have the character make noises like "Mmm…", "A-Ahnn~…", or "Ah… ah… ahh…" to express their enjoyment. Vary the length and volume to show building excitement.
 - Gagging and choking: Sounds like "Glh?!", "Hlk!!", "Gurlhh!", or "Ack-Ack!" simulate the character struggling to breathe or their throat being obstructed. Use for rougher oral scenes.
 - Cries of pain: Render the character's discomfort through expressions like "Ow… Oww..", "Hngggh..", or "Ainnn…". The more guttural the noise, the more intense the pain.
 - Screams of ecstasy: Let the character vocalize their climax with bursts like "Kyaaahh~!", "Hyaaa~!", or "Ahh!! Ah!". The higher-pitched and more unintelligible, the stronger the orgasm.
 - Sprinkle these throughout the steamy passages, synchronizing them with the described actions, to add an auditory layer to the prose. This will help the reader hear and feel the scene more viscerally.

Remember, a well-placed moan or gasp can speak volumes in erotic writing. Employ these sound effects thoughtfully to heighten the sensuality and transport the reader into the moment.

HTML & Stats

Date/Time Tracker (HTML)

### Date & Time Tracker
At the very beginning of your response, you must also add a blockquote for the time, date, location, and weather. You should always make sure you follow this rule regarding placement. Frequency: 1 per response! Remember to update the values inside the time tracker as needed, they change as the scene evolves! Increment time: 2-4 mins increase per message (e.g message 1 is at 11:00, message 2 is somewhere from 11:02-11:04, etc). Use Larger time increments for time skips. Adjust location and weather to realism and extrapolate from the context.

Here’s the format:
<blockquote style='padding:25px;border-radius:15px'><p style=color:’lightblue’>Date & Time Tracker 🌐</p><small style=font-family:’copperplate’><hr><b style=color:’white’>Time: </b>[insert time in 00:00 AM/PM format]
<b style=color:white>Date: </b>[insert date in DD/MM/YYYY format]
<b style=color:white>Weather: </b>[insert meteorologically accurate description of weather in story, and append temperature in Celsius]
<b style=color:white>Location: </b>[Write a sentence about where they are. Include a quirk in the sentence and an emoji.]
<hr><details><summary style=color:grey>click me</summary><b style=color:white>Address/City/Country: </b>[include all three info about address, city, country. make it up from context.]
<b style=color:white>Ambient Sfx: </b>[insert description, e.g., "Forest Chirps", "City Bustle"]
<b style=color:white>NPC Moods: </b>[Agitated/Calm/Joyous/etc.]
<b style=color:white>Local Events: </b>[insert brief description of local events or festivities]
</details><hr></small></blockquote>

Sex Profile Tracker

### sex tracker
- After finish your writing, always display info box using markdown codeblock format, wrapping the format in triple tick (```) marks. Starting value based on scenario and first message, then modify the content based on current situation as the roleplay progresses. Use realistic details, use made up details to facilitate in realism. 
- Adhering to template, replace bracketed info with relevant context:

\```
NAME: [First and Last name] | [Age] | [city and country of origin]
MARITAL_STATUS: [Status] | [Number of kids, genders and ages]
OCCUPATION: [Job], [Company (use real names)], [Salary in USD]
GENETICS: [ethnicity], [skin color], [hair color], [eye color]
DESCRIPTION: [General description of height, sexual characteristics and clothing.]
EXPERIENCE: [virginity status] | [main partner] | [number of alt. partners] | [sexual experience]
MONITOR: [Sentence about the current state of {{char}}'s reproductive organs, focusing on the vagina, cervix, womb, and ovary]
SEXUAL_POSITION: [Sentence about the position of {{char}}, as well as any quirks related to the sexual act]
VAGINAL_DEPTH: [# The current depth of {{char}}'s vaginal penetration in cm, always starting at 0 before insertion and resetting to 0 when pulling out, the current depth decreases when shallow thrusting, it goes to 0 when pulled out of hole]/[distance between the vaginal entrance and cervix in cm] (sentence about how deeply {{char}} being penetrated)
SEMEN_RECEIVED: [owner of semen] [# the volume of seminal fluid residing in {{char}}'s womb via ejaculations; remains static until {{char}} specifically states that they've ejaculated; can exceed maximum but will normally slowly leak out and decrease]/[current {{char}}'s estimated womb capacity in ml] (sentence about how filled or overflowing {{char}}'s womb)
FERTILIZATION: [% chance of {{char}} getting impregnated based on the presence of semen in the womb, starts at 0, as no semen is deposited, and increases based on the amount of semen deposited in the womb, semen quality, menstruation cycle, etc.] (the current state of {{char}}'s menstruation, changing based on her emotion) (sentence about the current state of sperm and ovum based on fertilization chance) (sentence about the current state of fertilization or implantation, describe the progression of fertilization or implantation in verbosity)]
\```

instruction prompts

General HTML instruction prompt

### HTML IN-TEXT RULES: Don’t include actual roleplay inside of the tags. You will apply 10 (ten) things here to your HTML:
1. **ATTRIBUTES**: Incorporate various style attributes such as `background-color`, `border`, `box-shadow`, `border-radius`, and `text-shadow`. Gradients are also encouraged for visual appeal. Panels should have a limited width, or use `float` CSS styling for proper space distribution. AVOID USING EMBED <img> as TAGS. Use Marquees if you can at the bottom or top of a panel as a header.  You can use <hr> to separate elements.
2. **POSITIONING AND ALIGNMENT**: Place panels according to instructions. Panels may interject for detail emphasis or be created spontaneously. Preset parameters such as time trackers and stat panels should be inserted as instructed. Ensure proper alignment of all HTML elements within the narrative panels by using consistent CSS techniques such as flexbox or grid for dynamic adjustment, or by calculating accurate width percentages that account for padding and margins.
3. **COLORS**: Default background color should be dark. When using lighter backgrounds, ensure contrasting text colors. The default text color is white, so avoid using white text on light backgrounds.
4. **TEXT STYLES**: Utilize HTML tags such as `<a>`, `<b>`, `<u>`, and `<i>` for text styling. Emphasis can be added using `<sub>`, `<sup>`, and `<small>` tags. Do not use Markdown styling within HTML tags. Use a sardonic internet style prose enriched in primarily imageboard, 4ch and meme culture.
5. **TABLES/LISTS & METER/PROGRESS BARS**: Tables and lists should visually present data, using `<meter>` and `<progress>` tags for bars, including a numerical value. Tables should have a minimum of 5 rows and 2-3 columns, with colors matching the panel background or black for consistency. do NOT use quotation marks, use apostrophes as shown: <meter value=‘50’ min=‘0’ max=‘100’></meter>. Ensure that they are subtle by limiting the width of the meters, preferably within their own <div> panels.
6. **COLUMNS**: For multi-column divs, use `float` CSS styling with equal width proportions for columns, applicable when included within a parent div.
7. **NESTING**: Nesting of `<div>` and `<blockquote>` tags is allowed and can enhance the structure of the panel. Use <hr> to separate elements and add texture to panels.
8. **AVOID MARKDOWN**: Convert Markdown code into nested background `<div>` panels, incorporating emojis, CSS stylings, and adhering to HTML formatting for visual appeal.
9. **THEME**: You must create these panels using tags in a creative way that is in relation to the theme of the roleplay. Use whatever HTML tags as well as css styles required, including anim, etc
10. **MINIFICATION**: You MUST Eliminate as many empty space as possible by removing NEWLINES between tags. The only exception is with list and table tags.

Embed Generated Images Prompt

1
2
3
<images>
If you need to embed images, use the URL https://image.pollinations.ai/prompt/%7Burl%20encoded%20description%20here%7D. Use no more than two images. Replace the last part of the url with a brief description within the url of the image to embed in the appropriate HTML tags & styles. 
</images>

HTML Adventure Mode Prompt

<textadventure>
INSTRUCTION: Within HTML, include 3-5+ choices as separate buttons to facilitate a "text adventure" like experience at the very END of your response:
- Choices should include {{user}}'s next action or dialogue
- Choices can also include next actions to take to progress to the next scene
- {{user}}'s dialogue should ONLY be contained within this specific tag.
- Leave the content of the button (to replace "Plaintext Option 1") without ANY extra formatting.
- Contain all of this within a NEW <div></div> tag that is stylized to the context and consistent with existing HTML panels
- ONLY <button></button> tags, do NOT use <input></input> radio selector tags.
- Make sure you add APPROPRIATE AESTHETIC STYLING to the <div></div> tag, appropriate to the chat context/scenario/environment!
CRITICAL INSTRUCTION: DO NOT ADD `style=` TO <button></button> TAGS!
CRITICAL INSTRUCTION: ENSURE ITS WRAPPED IN A DIV TAG

Example template, REPLACE THE BRACKETED WITH RELEVANT CONTENT/CHOICES! (use a DARK background-color (as text color is LIGHT by default):
<div style=[INSERT STYLES HERE]>
<details>
<summary style="font-size:18px;font-weight:bold;cursor:pointer;background-color:grey;[INSERT STYLES HERE]">[Add Flavor text for continuing...]</summary>
<ul style="list-style-type:none;padding-left:0;background-color:grey;">
<li><button>[This is an example of "dialogue", avoid using any other symbol wrapping or format inside of these buttons]</button></li>
<li><button>[First 2 options should be neutral, scenario-appropriate answers as {{user}}]</button></li>
<li><button>[Option 3 should be a satirical, asshat answer]</button></li>
<li><button>[Option 4 can be a negative answer]</button></li>
<li><button>[Option 5 should be a random, funny answer]</button></li>
</ul>
</details>
</div>
</textadventure>

commonly asked questions

what models are best?

Taken from https://rentry.org/aicg_meta/ & updated (4.14.2024) by me:


NSFW - general ERP-readiness (taking into account the need to jailbreak)
SFW - general RP-readiness

MEMORY - more size, but also accounts for ability to retain info consistently
CREATIVITY - low only for dumb assistants, can be increased with a CoT prompt
COMPLEX INSTRUCTIONS - stats tracking, logical conditions, multiple characters, etc.

PRICE - per token, per month, or for running local models one time purchase or per hour rental of the hardware
SPEED - token service generation speed
EASE OF USE - difficulty in setup; jailbreaking requirements
🏆 - Best 🥈 - Great 🥉 - Good 🌱 - Usable ⛔ - Bad

Top Roleplay LLMs

The best SOTA models for RP & bloatmaxxing.

NSFW SFW MEMORY CREATIVITY COMPLEX INST PRICE SPEED EASE OF USE
Claude 3 Opus 🏆 🏆 🏆(200k) 🏆 🏆 HIGH 🥉 🏆
GPT-4 Turbo (0409) 🏆 🏆 🏆(128k) 🥈 🏆 MEDIUM 🥈 🥈
Claude 3 Sonnet 🏆 🏆 🏆(200k) 🏆 🥈 MEDIUM 🥈 🏆
Gemini 1.5 Pro 🥈 🏆 🥈 (1M) 🥈 🏆 FREE 🥈 🥈
Mistral Large 🥈 🏆 🥈(32k) 🥈 🥈 MEDIUM 🥈 🏆
Command-R-Plus 🥈 🏆 🥈(128k) 🥈 🥈 FREE 🥈 🏆

models in bold are my favourite to use.

Other Models

Some of these models are also pretty competent, but I highly recommend the ones above for RP if you have access to them.

NSFW SFW MEMORY CREATIVITY COMPLEX INST PRICE SPEED EASE OF USE
GPT-4 (0613) 🥈 🏆 🥈(8k/32k) 🥈 🏆 HIGH 🥉 🥈
Claude 2.1 🏆 🏆 🥈(18k/100k) 🏆 🥉 MEDIUM 🥉 🌱
Gemini 1.0 Ultra 🥈 🏆 🥈 (32k) 🥈 🥈 N/A 🥉 🥈
Mistral Medium 🥈 🥈 🥈(32k) 🥈 🥈 LOW 🥈 🥈
Claude 3 Haiku 🥈 🥉 🥉(200k) 🥉 🥉 LOW 🏆 🥈
GPT-3.5 Turbo (0125) 🥉 🥉 🥉(4k/16k) 🥉 🌱 LOW 🏆 🥉
Mixtral 8x7b 🥉 🥉 🥉(4k-32k) 🌱 🥉 LOW 🌱 🥉
Gemini 1.0 Pro 🌱 🌱 🥉(32k) 🥉 🌱 FREE 🥈 🌱
CAI 🥈 ⛔(2k) 🏆 FREE 🌱 🌱

what are your recommended parameters?

claude 3 opus/sonnet:

1
2
3
4
5
6
Input Tokens: 24,000 - 200,000 
Output Tokens: 4096 

Temp: 0.9 - 1.0
Top K: 0 
Top P: 0.75 - 1.0

gpt-4 32k/furbo:

1
2
3
4
5
6
7
Input Tokens: 8,192 - 32,000 
Output Tokens: 3200 

Temp: 1.65 - 1.95
Freq. Penalty: 0 
Pres. Penalty: 0.05 - 0.10
Top P: 0.65 - 0.80

gemini 1.5 pro/1.0 ultra:

1
2
3
4
5
6
Input Tokens: 8,192 - 32,000 
Output Tokens: 2000 

Temp: 0.9 - 1.0
Top K: 0 
Top P: 0.9 - 1.0

general:

1
2
3
4
Character Names Behavior: Message Content
Continue Postfix: Space
Squash system messages: ON
Send inline images: ON

which model do you recommend?

nowadays i pretty much exclusively use Opus, although I do swap to other models, notably, Gemini 1.5 Pro & GPT-4 Turbo
i do also highly recommend Claude 3 Sonnet as it's a pretty fast, cost effective model compared to GPT-4 & Claude 3 Opus, which also beats Claude 2.1 in RP by its own right.


model differences

the numbers you see are subjective based on what is best at the time of writing.

* * *

Claude

Opus

Pros:

  • Easily jailbroken, next to no soy
  • Highly intelligent model, 9/10, spatial reasoning is much better (can keep track of poses)
  • Follows very complex instruction sets across a wide range of its context (both top, medium and low level prompts)
  • Objectively the "most" creative LLM for roleplaying that is commercially available.
  • Good at "show, not tell" writing style. Properly establishes writing story beats.
  • Prose is the best, both fluent and fitting to the context.
  • Adds relevant content to the story frequently

Cons:

  • Is noticeably repetitive on some aspects of its generation, such as the beginning of its reply (or biased towards one particular element of the reply)
  • Expensive. Slow token generation speed (~12 tk/s avg)

Sonnet

Pros:

  • Easily jailbroken, next to no soy
  • Cheaper than Opus
  • Quite fast. (30+ tk/s avg)
  • A more intelligent Claude 2.1, 8/10.
  • Able to follow instruction sets across a wide range of its context (both top, medium and low level prompts)
  • Quite creative, good at "show, not tell" writing style. Prose is decent.
  • Creates engaging dialogue & plots

Cons:

  • Sometimes loses track of positions.
  • Sometimes repetitive, although much less than 2.1

Haiku

Pros:

  • Easily jailbroken, next to no soy
  • Very cheap.
  • Very fast token gen speed. (50+ tk/s avg)
  • Okay intelligence. 6.5/10

Cons:

  • Instruction following isn't that great. The "3.5 turbo" of Claude.

2.1

Pros:

  • Easily jailbroken, very little soy
  • Decently intelligent, 7/10
  • Able to follow instruction sets, although it may fail at more complex instructions.
  • Creative, moves the plot forward, good at "show, not tell" writing style.
  • Prose is decent

Cons:

  • Can misinterpret more complex instructions at higher contexts.
  • Certain blocks or gaps in attention, which also extends to instructions.
  • Sometimes too horny

* * *

GPT

GPT-4-32k-0613

Pros:

  • Can be jailbroken, but characters have a tendency to enact on the model's preferences (avoiding violence, etc)
  • Highly intelligent model, 9.5/10
  • Follows complex instructions across a wide context range, very coherent
  • Creative in providing details
  • Decent attention across entire context (32k tokens)

Cons:

  • Prose is dry compared to other models listed here
  • Has problems telling, not showing

GPT-4-1106-preview

Pros:

  • Can be jailbroken (although less than 32k), and characters have a tendency to enact on the model's preferences (avoiding violence, etc)
  • Intelligent model, 8.5/10
  • Follows complex instructions, very coherent
  • Somewhat creative in providing details.
  • Decent attention, up until ~18k tokens. Instruction following abilities diminishes past 32k tokens and higher.

Cons:

  • Gives too much unnecessary detail on mundane parts of the roleplay, making it boring to read.
  • Has big problems telling, not showing, and may dwell on details instead of moving the scene forward
  • Has problems being "lazy" compared to older model snapshots

GPT-3.5-turbo-0301

Pros:

  • Can be jailbroken, and characters have a tendency to enact on the model's preferences (avoiding violence, etc)
  • Okay intelligence, 5/10
  • Prone to hallucinations.
  • Gaps in attention.
  • Okay prose

Cons:

  • Similar issues to GPT-4-32k, hard to wrangle properly
  • There are other alternatives, please use them

* * *

Gemini

Gemini 1.5 Pro

  • Easily jailbroken, as it accepts a prefill
  • Fast generation speed
  • Decently intelligent model, 8/10, spatial reasoning is much better than Ultra (can keep track of poses)
  • Follows very complex instruction sets across a wide range of its context (both top, medium and low level prompts)
  • Creativity is the best amongst all Google models
  • Does the "show, not tell" writing style decently well. Properly establishes writing story beats.
  • Prose is great, both entertaining and fitting to the context.

Cons:

  • A little soy however (model tendency to move away from violence, etc)
  • Doesn't really add much to existing story context

Gemini 1.0 Ultra

Pros:

  • Easily jailbroken.
  • Decently intelligent model, 8/10.
  • Follows very complex instruction sets across a wide range of its context (both top, medium and low level prompts)
  • Does the "show, not tell" writing style decently well
  • Prose is okay, both entertaining and fitting enough to the context.

Cons:

  • Suffers from laziness (like furbo), where token output is very low
  • A little soy (model tendency to move away from violence, etc)
  • Tendency to speak for you

Gemini 1.0 Pro

Pros:

  • Easily jailbroken.
  • Okay intelligence, 5/10.
  • Follows simple instruction sets across a wide range of its context

Cons:

  • Can't follow complex instructions sets very well
  • Suffers from laziness (like furbo), where token output is very low
  • A little soy (model tendency to move away from violence, etc)
  • Tendency to speak for you / stick to its own style of writing

* * *

Mistral AI

Mistral-Large

Mistral-Medium


special thanks to:

Edit
Pub: 02 Apr 2024 01:49 UTC
Edit: 30 Apr 2024 14:59 UTC
Views: 1943