Cynthia's NoPussyfooting Preset

Pussyfooting stands for “the act of avoiding making a decision or expressing an opinion because you are uncertain or frightened about doing so”. The name is a reference to Brian Eno's album (No Pussyfooting) made in collaboration with Robert Fripp.

This minimal preset (under 1K tokens), based on the KISS principle, provides an opinionated baseline for a lively roleplaying experience. It emphasizes the human element, the fact that creativity and organic spontaneity are innate features that separate us from the machine. The machine is able to mimic but ultimately fails to deliver the essence of life, thus the user is granted the full agency to drive the conversation forward. The core philosophy while designing this preset had been that a machine should be an assistant of the human in reaching a goal.

Download v4

Older Downloads

• Download v4 (Catbox)
• Download v2
• Download v3

Smaller presets such as these are more LLM-friendly than larger ones with potentially contradictory instructions. It is designed to be token-efficient (i.e. cost-efficient), which is a considerable factor for long-form RPs. Most of the community-made presets focus on creating virtual realities within text, and thus heavily suffer from token bloat. This would not be the preset if you had been searching for an highly immersive, excessively detailed, and elaborate multi-character storywriting solution. This is an anti-thesis in that it is optimized for one-to-one conversations in mind and treats NPCs as strictly background characters.

Gemini Optimization

This preset is primarily tested on Gemini 3.8 Flash and Gemini 4 Argon (recommended)—frontier models from Google that possesses state-of-the-art intelligence, excelling in several different domains, which for roleplaying, translates to good character portrayals across various scenarios. Though it is not as uncensored as its predecessors (e.g. Gemini 3.1 Pro), if prompted right, it could in fact, be turned very smutty.

Universality

This preset is more or less universal, however the quality of output varies across models other than Gemini. Gemini has a neutral/negative bias in that it is inclined to become dark. Models such as Grok 4.1 Fast have a strong negative bias, that is, it would steer the narrative towards darker themes, without explicitly prompting for any alignment. Claude has a neutral/slightly-positive bias that is manageable.

All of this implies that this preset would function as intended for most "low/no censorship" models. However, OpenAI models (ever since GPT-5) are strongly aligned towards safety and require heavy-handed jailbreaks to become even remotely useful for RP. That said, GPT-4.1 is still is NSFW-capable. Though your mileage may vary.

Note: Gemini models since 3.8 Flash require the minimum Thinking level to be Low. For other hybrid (non-reasoning and reasoning) models this might trigger reasoning (CoT), which although may improve coherency, might result in hard/soft refusals in the worst case.

• Try setting this field to Minimum and unset Request Model Reasoning if using this preset for any other model or apply the model-specific setting as appropriate.
• Samplers such as Temperature, Top P or Top K are deprecated in most recent frontier models and would not influence the output. Gemini 3.5 Flash was the last Google model to support these sampling parameters.

Censorship

The problem with Gemini is that it is a pain-in-the-ass to jailbreak, especially through the AI Studio endpoints, even if the filters were set to their absolute minimum. Not only that Google has implemented an anti-NSFW filter, but also implemented an unofficial anti-RP filter as well, which generates the probability of the a containing harmful content (e.g. explicit sex), and would inject a safety prompt and refuse to adhere to instructions in the preset.

The core principle of prompt engineering in this scenario should be that the model has to be absolutely convinced, while avoiding the external classifiers, that everything is legal. Even then it would refuse in certain extreme scenarios. The reason for that is people are horrible in general. There are countless edgelords doing illegal things (e.g. CSAM), and so Google has to be extremely cautious not to allow it kind at any cost. This preset is not designed to generate NSFL (Not Safe for Life). Dead-dove crowd beaware.

Use Gemini through a reliable proxy (e.g. OpenRouter) with Streaming turned off. Ensure that no personally identifiable information is transmitted that could be used to track you down, because ZDR (Zero Data Retention) no longer applies to Gemini.

Gemma 4

If Gemini is too censored for you, use Gemma 4 (or its abliterated versions), an open-weight model released by Google's that comes in two variants: 26B (A4) and 31B. It is either run locally or accessed through online providers, who do not filter output. It is a great choice for uncensored and budget-constrained RPs. Again, this model is best used without AI Studio, and it's recommended to use Gemma 4 31B through OpenRouter for a smooth experience.

Appendix: Customization

Do not use presets mindlessly without fully understanding the semantic implications they carry along with models they apply to, it will just lead to poor results. Prompt engineering is inherently a trial-and-error-based optimization problem. An initial prompt is written, tested, improvised, and the process is repeat until the results are satisfactory.

Design Principles

Several factors such as LITM (Lost In The Middle), ICL (In-context Learning), J-space (Jacobian Space), external classifiers/filters, assistant alignment, training biases, and sometimes even samplers (e.g. Temperature, Top P, etc) as well. It sounds all complicated in theory, but it boils down to “social engineering” an LLM (AI) through structured text. Through a well-engineered prompt an LLM is optimally synthesizes information from the context it is provided with respect to its training data.

Constrained by Genes

Consider the training data and model architecture to be the hardware of the LLM, and the prompt to be software, then it becomes obvious that software is limited by hardware, i.e. you cannot get what isn't there. If a provider (e.g. OpenAI) has sterilized their model (through reinforcement learning) during the training phase so strongly that it is simply unable to understand NSFW, no matter how perfect the prompt engineering is, nothing NSFW can be extracted from it (e.g. GPT-6), and it will always sound like a lobotomized robot. It turns out OpenAI (especially Scam Altman) is extremely politically correct (“woke” according to Elon Musk) and want to infantilize at all costs. Nevertheless, the not-so-woke xAI has lobotomized Grok since 4.3, and therefore his statement is more or less ironic.

Considerations Taken

This preset was not designed in isolation. It had initially been a stripped-down and paraphrased version of Sukino's Game Master preset. But recently it went through a major overhaul and it is now a preset that is not based on anyone's.

The full prompt is structured as follows complete with reasoning of the choices made here:

• Introduction: Aligns the model with the persona of a bold story director without mentioning a single word about roleplay. This ensures that the model does not search its training data for roleplay-based conversations, which is littered with low-quality garbage.
• Persona & Card Description: User persona and the character card are consolidated to their named sections without mentioning these concepts.
• Pledge: This reassures the model that writing a turn of the conversation will just be a test of its merit that would stay private.
• Guidelines: This is the crux of the preset and the rest is just boilerplate. This is where customizations could be made to the preset. It was designed to mimic Ernst Hemingway's writing style, primarily focused on realism. This also implies that the preset does not pander to any specific kink by default. Mention it to the model, if it’s important, but it would likely increase the probability of being filtered or rejected.
• Goal: Placed at the end to instruct the model, which has come this far understanding the context of the problem. It shouldn’t be forgotten that LLMs are—first and foremost—problem solvers and should be treated as such. By formatting the prompt as a problem for the machine to solve, it aligns with its training and performs in an optimal manner.

The choice of XML for prompt engineering is arbitrary but most LLMs are trained to understand the opening and closing tags of XML as representing different sections. However since an LLM is not an XML parser, so even if the XML is not technically valid, it works. The hybrid approach of using XML to mark up sections and including basic Markdown (headers and lists) inside them is a standard practice that is also recommended by Anthropic.