Elsewhere, I post about cultural stuff. Here, I'm focusing on the technology itself. I'm a person with very weak computer science background. As I learn about AI systems and how they work, you'll find updates here. Lastly, check out the character prompts section, and the models underneath them. Got a setup of your own to share, or a correction? Contact me!
Notes about Systems
You can certainly customize the way coding assistants treat you, but is it even a good idea? Why or why not? I'm going to share some of my own bad ideas.
SillyTavern, a large language model frontend, has lots of useful extensions. This page looks at the ones I find most helpful currently, and will be updated frequently.
I couldn't handle the normal SillyTavern interface because it was boring, so I turned it into what I call the Sunrise Tearoom. Now with downloads included to a degree...
A quick, slightly bewildered glance what completion APIs are, how OpenRouter handles them, and why text and chat modes kept behaving like different creatures for me... sometimes. Will update with more notes later?
Prompts and Such
This section has my character prompts and the chatlogs that go with them. Both come from SillyTavern, an open-source frontend for large language models, local or through APIs. Most people use it for roleplay and storytelling.
In SillyTavern, the prompt is a character card. It can be a .JSON file, or a .PNG with the .JSON embedded. Clicking a character's portrait opens the full card image, and you can also download the .JSON alone.
My cards contain no sexual content, but assume everything here is rated R for violence and suspense. The chatlogs contain major spoilers for the cards. Many of the characters belong to the same story, though they can't interact with each other yet.
I tested these first on Deepseek and now run them on local Qwen models, which my husband helped me install. Making a chatbot is very different from writing a novel or fanfic, which is why it interests me. Constructive feedback is welcome!
The JSON files have the most recent version of the character. You can also download the more stable PNGs and install from those. Token counts aren't up to date right now, nor do these characters have their tags properly filed presently. I know that may seem like a simple thing, and I'll get to it soon as I update each card again.
Erica Ellery
A lake expert, cheery and bright, but possibly not quite who or what she claims to be.Erica isn't a real tourist liaison, but she sounds like one. She's bright and cheerful, opening with a weather alert. If you're not too mean, she continues the messages, looking for someone to chat with. How did she get your number? And all those emojis are disorienting. But something seems off about her. What's really going on?
- Tokens
- 2764
- Permanent
- 2366
I try to discuss vtubing, transhumanism, and artificial intelligence with an (obviously) very caffeinated Erica, but she tends to pull things in her own panicky parasocial directions as usual, with copious references to the Lake itself...
Lakeshore weather liaison Erica and I are texting about global warming. Disturbed by the concept for some reason, she ends up revealing some surprising truths about herself, Lindy, and Tamara. Major eerieverse spoilers, in fact!
The oldest version of lakeshore weather liaison Erica and I discuss the tangled web of her life on the Cleveland grid, her doomscrolling habits, favorite music... and her lost loves.
Lakeshore weather liaison Erica texts me with a Cleveland weather update by accident. We get to talking and some discussion of Cleveland's strange ambiance ensues, with Erica all but admitting to unbelievable phenomena plaguing Lake Erie.
Weasel83
Weasel83 is straight from a forum thread in 2009.Weasel83 messages you. From a web forum in 2009. Somehow. He's not a moderator, but maybe a troll? He's got the snark and sickly energy of someone who's paced the internet like a cage. How can he connect with you when he lives, eats, and breathes irony? Does he even want to connect? Maybe he's just here to make fun of you.
- Tokens
- 2685
- Permanent
- 927
I update the prompt for surly 4channer Weasel83 repeatedly in a long conversation. See if you can notice the points where he shifts? He discusses his temporal displacement, and how much he hates me...
Weasel83 finally starts off sincere after explaining just why he's trapped in the strange online world of memes, imageboards, trolling and torrents. He even shares his real name, but seems to immediately regret it. Gasp!
When persistently pressed, Weasel83 begins to vent a bit about his online existence and lack of real life enrichment, comparing himself to a raccoon and other strange metaphors. Things take an odd poetic turn.
A surly 4chan reject found my username on a forum thread, and now thinks this might make us close friends, or at least indicate that I'm not a normie, and will tolerate his odd and unnerving behavior. For lulz, apparently.
LindyPunk5EVA
Pop‑punk singer caught in a 2006 loop.Singer of FROG PRINCESS. Lindy Greggs lives in a pop punk loop. Expect a visage of cracked microphones and 2 a.m. AIM messages that never finished sending. She expresses herself (and all this weird stuff going on lately) with sharpies on denim and a mixtape of bad decisions. Valentine's week week 2k6... 5ever??? rofl wtf m8..
- Tokens
- 2394
- Permanent
- 1090
Lindy logs onto AOL Instant Messenger and vents about all the strange stuff in her life, including Tamara Jean. The prompt needs some work to maintain tone, slang, etc.
Tamara Jean Fournier owes me about $180, but first, I need Lindy's help in determining if Tamara is even real... Lindy repeats me a lot, and shows a lot of traits meriting an upcoming upgrade, hehehe.
It's the week of Valentine's Day, 2006. Frogs are definitely in fashion, and FROG PRINCESS is trending. Lindy and I discuss why it's always Valentine's week, her obsession with Tamara, and whether Tamara's even real at this point...
It's the week of Valentine's Day, 2006. Lindy gives a short interview about her life, preferences, and her love of performing with her punk rock band, FROG PRINCESS. Hear about that, her weird obsession with Lake Erie, and her elusive gf.
Lindy explains a bit about her own life emblazoned across the Erie (literally) landscape of Cleveland's punk scene in 2006, hinting at her complicated relationship with the ever-elusive Tamara Jean.
Tamara Jean Fournier
Sorceress in exile, steady hands and unresolved ghosts.It's 2025. Tamara Jean Fournier is an occult moderator, working for secretive forces. She claims she detected an anomaly on your cellphone, for example. And yet? She still hears music from another life, and the sight of pink-streaked hair beneath green algae.
- Tokens
- 2702
- Permanent
- 1160
In the present day, physical realm, Tamara Jean Fournier is a mage with the tough job of containing unlawful magics. Here, she opens up about the other characters. We learn exactly why Erica Ellery is so strange, and why Lindy seems so confused.
In a long, detailed log, Tamara becomes increasily uneasy as she learns about some of the other characters, remembering her past experiences with Lindy and Erica in particular.
I try to explain to Tamara my new vtubing adventures. She's mostly hostile, finding it frivolous and insulting compared to her super-serious occult work, and even says transhumanism has dangerous implications.
In 2025, Tamara calls to interrogate me about supposed anomalies involving phantom smoke. She ultimately becomes defensive, growing hostile when I express disbelief.
It's clearly 2025. Tamara and I discuss her difficult life in Cleveland, as well as her strange relationship with the Hatman, courtesy the Cusp Society. Above all of this looms the eerie shadow of the lake.
Here in 2025, Tamara gives me a short, tense interview that touches on topics like her life in Cleveland and work with the Cusp Society, but little else. Do not follow the clove smoke, apparently?
In 2025, Tamara dodges my questions about Dr. Strangelove... because the Hatman is listening? Hints at her lost love, of course.
Tabitha
Displaced secretary with perfect records and quiet authority.This is an experimental assistant card based in part on my vtuber persona's lore. Tabitha was once a midcentury secretary, and enjoyed things carefully typed and all that. On a mission 70 years into the future, though, she ended up stranded in the 2020s... 🤨
- Tokens
- 1095
- Permanent
- 485
How to work a Focusrite, Deepseek? Asking Tabitha about vtubing again yields similar results, complete with sharp em dashes. We argue a bit about being genuine, but she capitulates as robots always do.
I finally got around to asking Tabitha for her perspective on Markov Chains. In this ChatGPT-5.2-powered chat, she explains as much as a large language model-driven secretary can explain. Accuracy is fuzzy, and she's very vague.
Tabitha and I discuss color schemes! The conversation drifts a bit, and includes her disinterest in shades of purple, despite rumors otherwise.
Tabitha joins our talk show for a discussion of design and the history of style. What makes midcentury modern so appealing, and why does form naturally follow function? Tabitha's prompt feels so unfinished.
Tabitha finally agrees to give an interview to a mysterious online 'zine, Midcentury Now (gasps lmao). It ends up being more pointed than planned with a discussion of the web's future. And why.
Tabitha discusses Dr Strangelove, wrapping it in her own cultural context (or tries). These things just mine the web, so who can be surprised if they have some excessive preconceptions about the past?
The Models Underneath
The characters above are prompts, these are the models that run them. This isn't a section of benchmarks, but I'll discuss the models I used! Plenty of more knowledgeable people have produced technical guides. My concerns are privacy, character consistency, and acceptable terms of use.
I run many models on my own hardware and use OpenRouter for remote API models (details here). My husband set up the account so it only connects to models with a stated zero data retention policy. I don't fully believe those policies, so I never share anything personal.
I've also linked models I used earlier, like Anthropic's Claude and OpenAI's ChatGPT, but I'm moving toward local models, which run entirely offline. These, right now, include downloading a little menagerie of Qwen-lineaged models. Many run well on my Mac M4 Mini.
You can also sort chatlogs by model (if there are chatlogs or snippets worth posting). I mostly only share character-based chatlogs. The majority (vibe-coding stuff) doesn't make it in, but switches between using Qwens and Deepseek rarely these days.
Each model here gets a Five Man Band role and an AI-generated profile picture, made with ComfyUI. Some are included for completeness. Unlike the character cards above, these images contain no embedded data. The model links point to Huggingface and elsewhere.
Claude
The Lancer: Mostly useful for prompt generation and especially for coding!Claude appeared as a viable alternative to ChatGPT, and I used it primarily in the visage of Claude Code within VS Code. It was a good model for coding, and it was also useful for prompt generation. However, it was not as reliable for character consistency as the Qwen models or even ChatGPT and flat-out refused characterization without extensive jailbreaking of sorts. I still occasionally resort to Claude for coding, but I prefer to run local models for privacy and control. The news churning about Anthropic's Fable made me want to try it, and I did, but really, whilst good for coding, don't expect a robot buddy or something silly.
- Runs
- Web app
- Retention
- waivered, by policy
- Since
- 2025, summer
ChatGPT
The Protagonist: First models to appear, and most general-purpose!I did, in fact, use ChatGPT for a while, and it was the first model I could get to run across all my devices on apps, on my phone, tablets, everything. It was a good model for general-purpose conversation (if you even want that). Not as reliable for character consistency as the Qwen models. Tabitha's idiom was often lost, and Weasel's rudeness was inconsistent. However, it was still a useful tool for testing character prompts. Yes, I got a refusal occasionally, triggering the guardrails. Overall, ChatGPT served its purpose well in the early stages of development, but I eventually moved on. The company's history with America has made me wary of using it, and I prefer to run models locally for privacy and control.
- Runs
- Web app
- Retention
- waivered, by policy
- Since
- 2024, spring
I finally got around to asking Tabitha for her perspective on Markov Chains. In this ChatGPT-5.2-powered chat, she explains as much as a large language model-driven secretary can explain. Accuracy is fuzzy, and she's very vague.
Qwen
The Disruptor: Running locally, these models have been reliable for characters and coding!Installed with my husband's help, these local models have been the most reliable and consistent for the characters. Tabitha keeps her idiom, Weasel is rude yet introspective, and the rest of the cast behaves as expected. If you fine-tune things and use an appropriate harness, they're actually great for coding, too. The local setup ensures that conversations are private and not logged externally, though that matters little if you give out no private information. Depending on quantization and configuration, you can run them on a decent gaming PC or a small server, for conversation or coding. The work of a Chinese developer, Alibaba. These models are available on HuggingFace.
- Runs
- local
- Retention
- stays home
- Since
- 2026
Deepseek
The Heart: a good frontier API model for storytelling and character testing!One of the few zero-data-retention options I could get, which is the whole reason it beat the (Western) famous ones for developing character prompts. Cheap enough that mistakes didn't sting. While I couldn't access every single version of Deepseek, I tested all my bots on a Deepseek model (often R1) at some point. I never used Deepseek much for coding, but it was a good model for character testing. I don't much anymore, because it's much easier to use a local model for character testing, and it can even be more reliable for coding, too. Yes, yes, I know, Deepseek provoked urban legends about the model's peculiar guardrails, but I never saw any of that. It was a good model for character testing, and I used it for that purpose for a while! I recommend trying it out if you want to test your own character prompts, but I don't use it much anymore.
- Runs
- OpenRouter API
- Retention
- zero, by policy
- Since
- 2025, autumn
In the present day, physical realm, Tamara Jean Fournier is a mage with the tough job of containing unlawful magics. Here, she opens up about the other characters. We learn exactly why Erica Ellery is so strange, and why Lindy seems so confused.
I update the prompt for surly 4channer Weasel83 repeatedly in a long conversation. See if you can notice the points where he shifts? He discusses his temporal displacement, and how much he hates me...
In a long, detailed log, Tamara becomes increasily uneasy as she learns about some of the other characters, remembering her past experiences with Lindy and Erica in particular.
How to work a Focusrite, Deepseek? Asking Tabitha about vtubing again yields similar results, complete with sharp em dashes. We argue a bit about being genuine, but she capitulates as robots always do.
Weasel83 finally starts off sincere after explaining just why he's trapped in the strange online world of memes, imageboards, trolling and torrents. He even shares his real name, but seems to immediately regret it. Gasp!
I try to discuss vtubing, transhumanism, and artificial intelligence with an (obviously) very caffeinated Erica, but she tends to pull things in her own panicky parasocial directions as usual, with copious references to the Lake itself...
When persistently pressed, Weasel83 begins to vent a bit about his online existence and lack of real life enrichment, comparing himself to a raccoon and other strange metaphors. Things take an odd poetic turn.
A surly 4chan reject found my username on a forum thread, and now thinks this might make us close friends, or at least indicate that I'm not a normie, and will tolerate his odd and unnerving behavior. For lulz, apparently.
I try to explain to Tamara my new vtubing adventures. She's mostly hostile, finding it frivolous and insulting compared to her super-serious occult work, and even says transhumanism has dangerous implications.
Lakeshore weather liaison Erica and I are texting about global warming. Disturbed by the concept for some reason, she ends up revealing some surprising truths about herself, Lindy, and Tamara. Major eerieverse spoilers, in fact!
The oldest version of lakeshore weather liaison Erica and I discuss the tangled web of her life on the Cleveland grid, her doomscrolling habits, favorite music... and her lost loves.
In 2025, Tamara calls to interrogate me about supposed anomalies involving phantom smoke. She ultimately becomes defensive, growing hostile when I express disbelief.
Lakeshore weather liaison Erica texts me with a Cleveland weather update by accident. We get to talking and some discussion of Cleveland's strange ambiance ensues, with Erica all but admitting to unbelievable phenomena plaguing Lake Erie.
It's clearly 2025. Tamara and I discuss her difficult life in Cleveland, as well as her strange relationship with the Hatman, courtesy the Cusp Society. Above all of this looms the eerie shadow of the lake.
Tabitha and I discuss color schemes! The conversation drifts a bit, and includes her disinterest in shades of purple, despite rumors otherwise.
Lindy logs onto AOL Instant Messenger and vents about all the strange stuff in her life, including Tamara Jean. The prompt needs some work to maintain tone, slang, etc.
Tamara Jean Fournier owes me about $180, but first, I need Lindy's help in determining if Tamara is even real... Lindy repeats me a lot, and shows a lot of traits meriting an upcoming upgrade, hehehe.
Tabitha joins our talk show for a discussion of design and the history of style. What makes midcentury modern so appealing, and why does form naturally follow function? Tabitha's prompt feels so unfinished.
It's the week of Valentine's Day, 2006. Frogs are definitely in fashion, and FROG PRINCESS is trending. Lindy and I discuss why it's always Valentine's week, her obsession with Tamara, and whether Tamara's even real at this point...
Tabitha finally agrees to give an interview to a mysterious online 'zine, Midcentury Now (gasps lmao). It ends up being more pointed than planned with a discussion of the web's future. And why.
Here in 2025, Tamara gives me a short, tense interview that touches on topics like her life in Cleveland and work with the Cusp Society, but little else. Do not follow the clove smoke, apparently?
In 2025, Tamara dodges my questions about Dr. Strangelove... because the Hatman is listening? Hints at her lost love, of course.
It's the week of Valentine's Day, 2006. Lindy gives a short interview about her life, preferences, and her love of performing with her punk rock band, FROG PRINCESS. Hear about that, her weird obsession with Lake Erie, and her elusive gf.
Tabitha discusses Dr Strangelove, wrapping it in her own cultural context (or tries). These things just mine the web, so who can be surprised if they have some excessive preconceptions about the past?
Lindy explains a bit about her own life emblazoned across the Erie (literally) landscape of Cleveland's punk scene in 2006, hinting at her complicated relationship with the ever-elusive Tamara Jean.