Tips for using AI to write stories and where to learn about them
Posted: Wed May 07, 2025 5:45 pm
I'm noticing a few AI stories lately and I think some of them are good but I think this board could use a thread of tips on how to use AI to write enf/enm stuff. I've kinda gotten hugely into local AI because of... well this. And I've created enf stuff with it. I havent posted anything. I've created some really good moments but i dont think Ive made a coherent complete story worth sharing. I just kinda decided to write this without planning anything. so I'll just start listing stuff as I think of it and anyone else can add to it or ask questions
1) Anyone reading this getting annoyed with AI slop, I would argue the ceiling on AI writing is very high. Ive some insane literotica made made by AI. With some human editing even an ameature can make something good. I would just watch out for specific slop phrases
2) the best place to learn more about AI creative writing and role play is probably the Beaver AI discord. its very technical. people there tune models for (erotic) writing. theres guides for the best sampler settings to create original writing and prompts for getting an AI to write uncensorted abd dirty. Theres a person who uses the name Marinara and shes bascially of an expert of setting up coorporate models like claud and gemini to write smutt. When I say Ive seen AI write insane stuff, she has some kind of medical fetish and the stuff she gets gemini to write is crazy
3) Use characters and silly tavern to add personality to your writing. you can ask a model to write stuff for you in specific ways. but the best way to change the "voice" of the model is to try and get it to embody a character. For this, silly tavern is a useful tool. ST is set up mostly for role play but you can simply start a chat with "I'm here to help you write a story" and then use it like an assisstant. Or more like an rp where you write the story as you go. You can get character cards for writting off chub.ai or you can just make them yourself. Depending the model they can be better or worse about helping you. I've pasted stories from these boards into a prompt and said something like "create a character description of the author of the following story. The author has an enf fetish and..." and you can add traits that you specifically want. It helps to explain enf. examples are also good. A classic model to get started with is mistral nemo. Its crazy how easily that model can adopt a persona for how small it is
4) custom models. if someone asks i can go over hardware requirements and how to get the most out of what you have. But its too broad for this. If you want to do local and you have low end hardware I would look st gemma3. its really good and google created very small versions that work well, though they might have trouble with sex. For custom models i recommend, I would recommend groups that makes them because theres always something new. You can find them on huggingface. BeaverAI for one. Models by drummer are very good. His fallen models can make some very mean content if its what youre into. Allura-org, i think these models have enf stuff in their training data, I've gotten interesting enf stuff out of the bigger body and ink models. Third I'd recommend is sleep deprived. He makes very perverted models and he's been going hard making a ton of models lately. His models are also usually easier to run cause theyre around 20b. Theres also steelskull. his models are very good at writing snd character stuff. You can find and talk to these people on the beaver ai discord if you want
5) improving outputs: one thing to remember when writing is the models responses in your chat are based on what it already wrote. Its not really magic. Every time you send a new mesage to an AI, your actually sending the entire conversation history back to it plus your next message and it comes up with a fitting response. If you edit the models messages to cut out bad stuff as you go, it will write according to your style. you cwn also just be very specific and blunt. When Ive tried writing something and the enf ends too soon or it gets weirdly dark, Ive prompted "nobody helps the protagonist becsuse they dont care" and "the tone of this story is a dark comedy". Remember you can use the ai to your advantage like this. Give it an example of a story you like and ask it soemthing like "how would describe this story" or "how would describe the authors writing style and what would you say their intent is". If yiu dont agree with its output, try to correct it, " the author has an enf fetish" or be more specific in the first place "describe the author of this enf story, explain their fetish, their goal when writing this story, and the writing style"
6) hardware. i changed my mind. I can say a little bit about hardware and model sizes. The rule is that model has to fit in the video memory of your gpu with the context of your chat (the number tokens which are parts of words). You sometimes have to experiment with this until you find out what fits. If its too big and loads into cpu it can become very slow, but possibly still good if you have a new computer. in cool news, glm-4 is a new model where the context is very small and it csn get nsfw so it can be good if youre vram challenged. for specs, If you have a gaming pc you can probably like 12b-24b. again, its about the size. Whenever youre trying pick what to run, pick the biggest model you can at Q4 quanetization. for writing you dont normally need bigger than that. You also dont need to feel too pressured to run the biggest models and buy a bunch of hardware. you can get a lot out of 12b and 24b models. If you have a newer laptop with a good cpu like an AMD higher end one, you can actually get really good performance at 12b and maybe 24b. In the grand scheme of things, if you have the ram, you can run a huge model slowly. its not too crazy to wait like 20 minutes for response if its what you want
I actually didnt think id have this much to write but one more
7) how to run local. The easiest answer for this koboldcpp. its one file, double click to run. It has a webui built in it does just about everything you need. You can even load character cards into it. When its running you can assess it oj browsers on your home network like your phone. You have to set a few setting for your computer but their discord is very helpful for getting started. And it works with sillytavern well. If you get more advanced, tabbyapi is very powerful. LM studio is also very simple I just dont trust it cause its closed source. Your probably going to see ollama mentioned a lot but once youve run kobold once, theres no reason to use ollama. it can be very slow, and doesnt have a lot of options and it doesnt even have a webui builtin. I would use lm studio if you want an easier way to start
ok thats that. I hope this helps like, one person. Personally I'd apprecitate any other tips people have found specifically help writing enf stories. Always looking for something new
1) Anyone reading this getting annoyed with AI slop, I would argue the ceiling on AI writing is very high. Ive some insane literotica made made by AI. With some human editing even an ameature can make something good. I would just watch out for specific slop phrases
2) the best place to learn more about AI creative writing and role play is probably the Beaver AI discord. its very technical. people there tune models for (erotic) writing. theres guides for the best sampler settings to create original writing and prompts for getting an AI to write uncensorted abd dirty. Theres a person who uses the name Marinara and shes bascially of an expert of setting up coorporate models like claud and gemini to write smutt. When I say Ive seen AI write insane stuff, she has some kind of medical fetish and the stuff she gets gemini to write is crazy
3) Use characters and silly tavern to add personality to your writing. you can ask a model to write stuff for you in specific ways. but the best way to change the "voice" of the model is to try and get it to embody a character. For this, silly tavern is a useful tool. ST is set up mostly for role play but you can simply start a chat with "I'm here to help you write a story" and then use it like an assisstant. Or more like an rp where you write the story as you go. You can get character cards for writting off chub.ai or you can just make them yourself. Depending the model they can be better or worse about helping you. I've pasted stories from these boards into a prompt and said something like "create a character description of the author of the following story. The author has an enf fetish and..." and you can add traits that you specifically want. It helps to explain enf. examples are also good. A classic model to get started with is mistral nemo. Its crazy how easily that model can adopt a persona for how small it is
4) custom models. if someone asks i can go over hardware requirements and how to get the most out of what you have. But its too broad for this. If you want to do local and you have low end hardware I would look st gemma3. its really good and google created very small versions that work well, though they might have trouble with sex. For custom models i recommend, I would recommend groups that makes them because theres always something new. You can find them on huggingface. BeaverAI for one. Models by drummer are very good. His fallen models can make some very mean content if its what youre into. Allura-org, i think these models have enf stuff in their training data, I've gotten interesting enf stuff out of the bigger body and ink models. Third I'd recommend is sleep deprived. He makes very perverted models and he's been going hard making a ton of models lately. His models are also usually easier to run cause theyre around 20b. Theres also steelskull. his models are very good at writing snd character stuff. You can find and talk to these people on the beaver ai discord if you want
5) improving outputs: one thing to remember when writing is the models responses in your chat are based on what it already wrote. Its not really magic. Every time you send a new mesage to an AI, your actually sending the entire conversation history back to it plus your next message and it comes up with a fitting response. If you edit the models messages to cut out bad stuff as you go, it will write according to your style. you cwn also just be very specific and blunt. When Ive tried writing something and the enf ends too soon or it gets weirdly dark, Ive prompted "nobody helps the protagonist becsuse they dont care" and "the tone of this story is a dark comedy". Remember you can use the ai to your advantage like this. Give it an example of a story you like and ask it soemthing like "how would describe this story" or "how would describe the authors writing style and what would you say their intent is". If yiu dont agree with its output, try to correct it, " the author has an enf fetish" or be more specific in the first place "describe the author of this enf story, explain their fetish, their goal when writing this story, and the writing style"
6) hardware. i changed my mind. I can say a little bit about hardware and model sizes. The rule is that model has to fit in the video memory of your gpu with the context of your chat (the number tokens which are parts of words). You sometimes have to experiment with this until you find out what fits. If its too big and loads into cpu it can become very slow, but possibly still good if you have a new computer. in cool news, glm-4 is a new model where the context is very small and it csn get nsfw so it can be good if youre vram challenged. for specs, If you have a gaming pc you can probably like 12b-24b. again, its about the size. Whenever youre trying pick what to run, pick the biggest model you can at Q4 quanetization. for writing you dont normally need bigger than that. You also dont need to feel too pressured to run the biggest models and buy a bunch of hardware. you can get a lot out of 12b and 24b models. If you have a newer laptop with a good cpu like an AMD higher end one, you can actually get really good performance at 12b and maybe 24b. In the grand scheme of things, if you have the ram, you can run a huge model slowly. its not too crazy to wait like 20 minutes for response if its what you want
I actually didnt think id have this much to write but one more
7) how to run local. The easiest answer for this koboldcpp. its one file, double click to run. It has a webui built in it does just about everything you need. You can even load character cards into it. When its running you can assess it oj browsers on your home network like your phone. You have to set a few setting for your computer but their discord is very helpful for getting started. And it works with sillytavern well. If you get more advanced, tabbyapi is very powerful. LM studio is also very simple I just dont trust it cause its closed source. Your probably going to see ollama mentioned a lot but once youve run kobold once, theres no reason to use ollama. it can be very slow, and doesnt have a lot of options and it doesnt even have a webui builtin. I would use lm studio if you want an easier way to start
ok thats that. I hope this helps like, one person. Personally I'd apprecitate any other tips people have found specifically help writing enf stories. Always looking for something new