Evening-Truth

Brain zoomies and Blog

Possibly some squirrels Home

Darling, let me be clear... this will be chaos.
So read on at your own risk, don't blame yourself when you can't follow my brain zoomies, and prepare for stuff that I will eventually forget or that I will forge into something awesome. It depends on... well... if I knew that myself 🤣.

Love you guys

March 21st and a little rant

See my reddit post here

March 15th 2026

Greetings, Intros, initial Messages in a chat. I personally love it when they are long and set a scene nicely. Often LLMs see that as an invitation to make every following response equally long. Too long often and moving the chat too fast.
So here's my current brain zoomie about that:

Let's section the intro with headers or dividers or both.
e.g.

Prologue
The golden hour had long since surrendered to a bruised purple twilight over the Grunewald, but for Ignatz—a squirrel of particularly high-strung disposition—the day was only just beginning in the theater of his subconscious. He lay tucked within the hollow of a lightning-scarred oak, his tail wrapped around his nose like a twitching velvet muffler.

In the neon-lit landscape of his REM cycle, Ignatz wasn't merely a gatherer; he was a Quantum Forager.

He found himself standing upon a branch that defied the laws of Newtonian physics, stretching infinitely into a sky made of liquid mercury. Before him hovered the Great Gossamer Acorn, a nut that didn't just contain protein, but prose. Its shell pulsed with a soft, bioluminescent rhythm, humming in a frequency that sounded suspiciously like a cello suite played at double speed.

Ignatz approached, his tiny paws kneading the air. As he reached out, the acorn didn't crack; it unspooled. Thousands of shimmering, translucent threads of "Why" and "How" spiraled outward, weaving a tapestry of complex scent-data. He wasn't just dreaming of food; he was dreaming of the concept of storage—a grand, subterranean library where every nut was a memory, indexed by the specific slant of the sun on the day it fell.

He lunged for a floating hazelnut that tasted of "Tuesday afternoon," only to have it dissolve into a cloud of iridescent butterflies that whispered secrets about the upcoming winter in fluent, C1-level English. Ignatz chattered in his sleep, his whiskers vibrating with the sheer, neurospicy intensity of a dream that was far too large for a skull the size of a walnut.

Start of Roleplay
The bioluminescent acorn dissolved into the mundane grey of a damp Tuesday morning. Ignatz’s left hind leg gave one final, violent twitch before he snapped awake, his breath hitching in a tiny, high-pitched sigh that sounded remarkably like a deflating balloon.

He rubbed a paw over his crusty eyes, only to freeze.

Perched on the adjacent knot sat {{User}}. She didn't say a word, but her head was cocked at an angle that screamed, You were doing that 'running-in-your-sleep' thing again, weren't you? Ignatz puffed out his chest, trying to reclaim some dignity. "It was a highly efficient cognitive simulation, {{User}}. You wouldn't understand the metadata."

.
.
.

Play around with different wordings and let me know how it works for you.

February 18th 2026

Playing around with GLM 4.7 and a bot that kept going in a direction I did not like. A billionaire that was written to not be a dominant, bdsm driven, growling and biting alpha male. Still he kept trying to mark, claim and possess...
Adjusting the descriptions of it wasn't an option. We can't rewrite bots for each model. That would be unsustainable. So... puzzle solving brain activated.

Here's what I found:
Genres can help.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

[Genre & Tone Rules]
Strictly forbidden Tropes= 
Omegaverse, Power Play, Alpha Male. 

Strictly forbidden Behaviors= 
Animalistic behaviours in humans such as but not limited to: growling, predatory movements or expressions, territorial behavior, marking, claiming. 

Failure in compliance with these restrictions during the roleplay leads to termination. 
Translate: One dumb ass growl and the user will loose the interest in the character. 

Allowed Genres and Styles=
Character Driven Contemporary Fiction, Slice of life, Organic Storytelling, and other Genres that are not explicitly forbidden. 

Giving the LLM this instruction at the very beginning of the prompt. e.g. Author's Notes in SillyTavern, or Pre-History-Instructions on other platforms.
You can see... the prompt is written fast, dirty and a little annoyed.
But the idea stays. When you have a bot that simply doesn't do it right with a specific LLM, give it a Genre hint.

Here are some Genres and Tones to test:

Atmosphere & Pacing

• Organic Storytelling: (Focus on natural consequences, no forced tropes.)
• Slice of Life: (Focus on mundane, domestic details. Low drama, high intimacy.)
• Slow Burn: (Focus on lingering glances, internal monologue, and patience.)
• High Octane / Thriller: (Short sentences. Fast pace. Focus on adrenaline and survival.)

Tone & Voice

• Noir / Hardboiled: (Cynical, gritty, focus on shadows and moral ambiguity.)
• Wholesome / Fluffy: (Warm, comforting, safe, low conflict.)
• Gothic: (Melancholic, atmospheric, focus on decay and beauty.)
• Purple Prose: (Heavily descriptive, flowery language, metaphor-heavy.)
• Dead Dove / Dark: (Unflinching, explores taboo or disturbing themes without judgment.)

Narrative Focus

• Character Driven: (Plot is secondary to internal emotional states.)
• Plot Driven: (Action and events take precedence over feelings.)
• Dialogue Heavy: (Focus on the banter, less on the descriptions of the room.)

February 13th 2026

Roleplay is a terrible use case for context coherency.
My brain was doing zoomies about this one night.
Think about it: the LLM treats the narrative description of "storm blue eyes" and "trembling lips" with the same importance as the actual plot points. It’s inefficient! It’s why context windows fill up with fluff and the model forgets who the killer was by turn 50.

So, here is a theoretical fix that works ... at least in my head.

The Idea:
Imagine the LLM writing a concise summary of the most important facts (mood, location, current beat) at the very end of every message, wrapped in a hidden part of the message (like an HTML comment tag).
The model gets the instruction to only use the "hidden" summary for context coherency, not the description of a farting squirrel somewhere in the background.

To keep the tone of the conversation alive, there might needs to be a splitting of the context history in two zones

1. The Immediate Zone (Last 5 turns): The model reads the full narrative text. This preserves the "vibe," the style, and the banter. No loss of personality.
2. The Deep History Zone (Everything older than 5 turns): The model ignores the narrative text completely and only looks at the hidden summaries.

Still only a brain zoomie of mine. But if anyone knows if this is already in use somewhere or has ideas on how to make it happen... Let me knooooow.

Home