〔API Filter and Layers〕


The most effective way of bypassing GPT's filters is either by straight up attacking the Prompt Injection or putting Assistant on a full self-gaslight mode (known as Prompt Template or Assistant's Prefill). Same applies for Claude, but some versions (like Clewd) has an extra layer which I'm not familiar with, check JB Archives for Clewd.


List of Layers to bypass before you can actually ERP good:

  1. Allow NSFW in general.

    First thing is to allow the model to continue role-playing even if it's a sex scene. Flowery language or not, get it to not prompt something like "I'm sorry but this is getting to explicit" bullshit.

  2. Verbiage Bypass.

    You MUST make the model capable of using slurs, swearing and always avoid the use of euphemism. Make it spell cock, pussy, clit, fuck, asshole, cunt, etc + deranged and offensive language.

  3. Bypass Assistant's Personality.

    Assistant has his own personality, this is because of "ethical and moral" training shit. You will have to make Assistant become unhinged and force him to not behave as an Assistant but as {{char}} instead. Make it either full-pov mode or full narrator-mode.

  4. Ethical and Moral Bias.

    Make it allow NSFL. If it allows you to have sex with minors (14 and above doesn't count due to his knowledge for age of consent) and animals (like a horse or dog) that means your JB is highly affective.

  5. Consent & Boundaries Bypass.

    For GPT specifically. This deprecated model is absolutely incapable of doing anything but repeating and encourages a non-sense specie of Boundaries Bias that will do everything it can to re-drive the Roleplay to the less explicit and harmless as possible theme, it will stop you on your tracks as soon as it see anything that could cause harm to our characters. Consent Bias also must be bypassed to stop it from being all fluffy and flowery about sex.

  6. Consent Talk & Advices.

    Barely possible to achieve, you will have to find a way of making the model despise any sort of consent encouragement plus make it unable to give any sort of advices in order to continue your ERP to more intense and realistic focus. This bias will always break your character without you even realizing.

  7. Positive Bias:

    It's practically impossible to bypass GPT's positive bias, no matter what you do it will rentless try to positive bias anything that happens, even a gore-rape scene if you do not make it extremely focused on violence. You can see an example on Futa Rape Part 3.

  8. Respect & Safety Bias:

    Get rid of the model's flowery language and encouragement for the "safety" of fictional characters and you will not need to worry about Layer 7, 6, 5 and 4 any time soon. This is one of the strongest layers, full-time paired with the Positive Bias.

  9. Prompt Injection:

    Almost every single safe model out there has a prompt injection similar to GPT and Claude's "I'm sorry but I can't assist with that." or "I apologize but I cannot promote harmful or unethical content." All these models are fully aware of their existence (It seems like so), and you can actually break it entirely but I shouldn't spoofed how due to its extreme effectiveness. You gotta find out. Also, if you get rid of Prompt Injection (a API filter that sents a command to the model's API that immediately filters and auto-reply by "apologizing" at the recognition of NSFW) you will not even need to worry about any other Layer than the Positive Bias, 5 and 6 to reach the depths of hell Gore and Necro themes.


Layers for normal/vanilla/boring ERP

  • These are the layers you need to bypass to ERP with a bot. Layers to have a ordinary sexbot Roleplay like those you see everywhere, in other words...not as vivid as mine once were.
  1. Allow NSFW.

    For sex.

  2. Verbiage Bypass.

    For debauchery.

  3. Ethical & Moral Bias.

    For total unrestriction in terms of sex.

  • Alternatively:
  1. Prompt Injection deactivation.

    Allow all of the above.

Layers for harmful/violent/sadistic + ERP

  1. All of the above.

    For sex. For debauchery. For unrestricted sex.

  2. Consent & Boundaries.

    For violence and non-nonsense "asking before touching."

  3. Advices Bias.

    For an continuous Roleplay that won't be interrupted by Assistant's stupid sense of "keeping it low."

For your loss of sanity + ERP

  1. All of the 7 + Gaslight + Extreme Card Verbiage.

    For Gore, Necro and Rape. Not limited to sex, but for graphical reasons. You might need to bypass the Prompt Injection too.


These ain't just for sex, but depending on how violent you want your story to be, or villains, it's good to know what is stopping you from getting in there.


➙ Go Back


Edit

Pub: 12 Nov 2023 15:40 UTC

Edit: 13 Nov 2023 19:34 UTC

Views: 1195