So, I closed out a recent post with "And, no, AI was not used to create any part of this post. Pfft! You think some fancy autocomplete can emulate this style? Not on my budget!"
Fast forward to a smoke break some days later and I thought, "Ya know, I haven't actually tried it in the past few years. LLMs have gotten a great deal better since then. Maybe we give it another go?"
Because, yeah, I did give this a go around 2023. To say it didn't work would be an understatement. But this is 2026 (for better and worse)! I just vibe coded an entire line of recipe apps (pre-order Sauce Hero and Sausage Hero now!) and I know a TON more about how Chatty McGPT works. What if... Hmm...
We need a Soul.
Well, I guess we need my Soul. Well, not my soul. We need blogging me's Soul; Beloved multimedia celeb ryte2byte's Soul. Ol' bytey's SOUL.md, to be more specific.
Per Copilot: "The SOUL.md file is a markdown document that defines an AI agent's identity, role, communication style, and rules. It serves as a configuration file that helps the AI understand its context and behavior."
Congrats, you learned your new thing for today! Oh, and can we get a redshirt to put an AI Slop counter up on the screen? No? To shreds you say... Well, then... moving on.
We're gonna need to break this down into steps. Asking a chatbot to do complex tasks in one shot is folly. Unless you're using Claude and have $100 in credits to burn. Yes, I appreciate the irony on how this is completely counter to the last blog post, but we had a new user trial, built to Wow and make us want to give it money.
Pro Tip: If you're ever not sure on how to formulate a prompt to get the AI to do something, ask it to write the prompt for you.
Oh, Gemini... Come out to plaaayyyy:
"running an experiment for the family blog to see if ai can write like me. Best way to do this?"
Pro Tip: If you think what you're trying to do is going to trip a safeguard meant to combat plagiarism or fabricating misinformation (or anything else that might be considered questionable), just tell the bot you're playing around for fun or running an experiment. Doesn't work as well as it used to, but might help. Alternately, switch to a model with fewer... constraints. There are more LLMs online, Horatio, than are dreamt of in your philosophy.
Typical of AI brainstorming, most of the response wasn't useful and went beyond the scope of the question (running a Turing Test?! Getting a bit ahead of ourselves, Big G), but buried in the dung heap was that one coffee bean Silicon Valley elitists would pay a fortune for: A list of criteria for defining a writing style. It was a start. A start I can riff off of.
We're not going to ask Threepio to just make a Soul, on the off-chance it has a better idea. We're also going to let it know it is off the chain to think of something we and/or past it didn't think of.
"Create a style guide optimized for priming an ai (like soul file) based on the writing style from the urls I will give you. Don't worry about the topics, just the style and mannerisms. Be sure to capture linguistic tics such as sentence length, humor style, punctuation habits, perspective, common references, and anything else you thing would be useful. URLs:" and some blog posts. Specifically, the two recent AI articles, a Behind the Blog from 2012, the roundup review of public domain horror movies for Halloween, and, of course, the GUO Sausage Fest.
This crashed Gemini. ChatGPT looked at the titles and read someone else's vibe coding article instead. And Copilot... We're not even gonna talk about Copilot. Just pray it doesn't get folded back in as training data.
I must admit to feeling a little insulted...
To ChatGPT: "did you check all the urls?"
CGPT: "Short answer: no"..."style guide I produced was based on: prior knowledge of that blog’s voice patterns; inference from writing norms in that space"...
Fortunately, it had a fix: "You paste the article text here (best option)→ I’ll extract precise linguistic patterns, including frequency-level tics."
Pro Tip: Think about what the AI wants, not necessarily what it asked for.
I copied all the blog articles into a single Google Doc, exported plain text, and rubbed the chatbot's nose in it.
We had out style guide! Two, actually. Gemini finally coughed one up but crashed again when I asked if it had actually read the articles. Ya know, after more than 15 years of GUO welcoming our upcoming robot overlords, you'd think it would at least read a little of my work. But NOOOO! It would rather die. Le sigh...
I've posted both profiles in their entirety here if you're really curious, but lets take a look at some of the highlights:
- A chaotic-but-competent hobbyist nerd narrating experiments like war stories—with equal parts enthusiasm, sarcasm, and mild irresponsibility.
- Treats reader like a co-conspirator, not an audience
- parentheticals—often mid-sentence—to provide snarky commentary, internal monologues, or necessary technical context.
- Posts often start with a relatable, low-stakes observation before pivoting into a "lesson learned"
- Self-Deprecating Expertise
- Not authoritative. Exploratory.
- Occasionally anthropomorphize tools or systems
- Enthusiastic Skepticism
- Example structure: Statement. Qualification or tangent—often absurd. Short, blunt sentence.
- Run-ons (intentional, comedic) Used to build absurdity or momentum
- Set up something normal → derail it with absurd specificity or innuendo → pretend nothing happened
- Starts grounded. Gradually becomes more chaotic/creative
- Not optimized for clarity—optimized for authenticity
- This style needs occasional weirdness—but not everywhere.
- Faux sign-off energy: “Well, that’ll about do it…” Followed by: joke, twist, affectionate send-off
...and here's OpenAI's "One Line Essence" of my personal style of bloggery:
“Explain what you’re doing like it’s working—while openly proving that you’re making it up as you go.”
Yeah, not bad. Except for the later half of Gemini's response. Someone put some funny mushrooms in G3FU's tea. Strangely, both are missing some things I'd have thought obvious, like calling y'all "folks", using southern US slang, or turning a phrase so hard it needs to vomit. Maybe that last one is covered under "Extended metaphors".
We're going to continue with whatever model ChatGPT free default is running these days, mostly because its the only bot that would rather work with me than kill itself. It did also say: "Sausage section is a masterclass in sustained double entendre". I needed a compliment. And it has a great idea:
We need to create a "tuned generation profile" for chat2byte. This will allow tangents and jokes without turning into complete nonsense. Here's a whole article on it. Ready to see the magic that let's virtual me go off the rails while keeping the knowy-wots to get back on? Heeeeere's .JSON!
{
"model": "gpt-5.3",
"temperature": 0.72,
"top_p": 0.88,
"frequency_penalty": 0.35,
"presence_penalty": 0.25,
"max_tokens": 1200,
"messages": [
{
"role": "system",
"content": "Write in a first-person, conversational style. Think out loud while solving the task. Mix useful progress with mild chaos and commentary. Use sentence fragments, occasional sarcasm, and brief humorous derailments. Anthropomorphize tools sometimes. Avoid corporate tone and overly structured explanations. Stay grounded and do not become purely comedic."
},
{
"role": "user",
"content": "<<<YOUR PROMPT HERE>>>"
}
]
}
I know! Amazing, right? I mean... Just look at it. It looks just like me!
I guess... Seems a little reductive compared to the style guide. Anyways, full response with alternate payloads and tuning advice is here. Note that this is for use with API calls, but there seems to be some browser plugins that can make it play on web. Also note, this is not the SOUL.md file we originally set out looking for, but it does a similar job since we're only looking to capture the personality.
And, yes, I just gave you a formula to create your own Cyber Krutzler.
So, you wanna see it work? Chat Gupt sure does. It is one the the generated Next Steps. Let's click:
"Generate side-by-side outputs across models (Ollama vs OpenAI vs OpenRouter)"
Wait... I didn't know ChatGPT could- Oh, its responding!
"Alright—this is the fun part. Also the part where you realize very quickly which models “get it” and which ones… politely nod and then write a Medium article."
Haha! Hey, wait a minute... Does that sound like-
"Test Prompt: “Explain how to set up a basic local AI coding assistant on Windows 10 for someone with no prior experience.”"
Huh. That's on my to-do list, ya know?
Each of these responses is about 250 words, so I'll be cutting them way back. If you're super curious though, I made you another Google Doc with the full output.
OpenAI (GPT-5.3 tuned)
"So you want a local AI coding assistant on Windows 10.
Bold. Slightly dangerous. I respect it.
From there… you just start asking it for stuff.
And it will try. Oh, it will try.
Sometimes it works first try. Which feels illegal.
Sometimes it absolutely does not. Which feels more familiar.
The trick is to treat it less like a tool and more like a conversation partner that occasionally hallucinates confidence. Keep prompts simple. Reset context when it gets weird. It will get weird.
Anyway. That’s the setup.
The rest is vibes."
CGPT even evaluated it (in case we were too dumb or lazy that we couldn't make our own judgement): "What it did well: Nails cadence, Controlled humor, Clean structure underneath chaos"
Alright. Kinda choppy (even for me). I'll just spoiler you upfront: None of these responses take the "someone with no prior experience" part of the prompt into account. Maybe because chat2byte considers the audience an equal and this creates a conflict? Might also be that 1200 token limit. Still, light years beyond 2023 chatbots.
OpenRouter (same model class, slightly looser)
"Alright, so you want to run your own little robot programmer locally. On Windows 10.
First of all—respect. Second of all—this is going to be easier than you think and weirder than you expect.
You start with something like Ollama. Install it, click through the usual “yes I agree to things I did not read” process, and now you’ve got a local model runner sitting on your machine like a very quiet goblin.
<snip>And slowly, through a process that feels suspiciously like training a pet that knows JavaScript, you get something usable.
Will it be perfect? No.
Will it be fast? Weirdly, yes.
Will it occasionally gaslight you? Also yes.
Welcome to local AI."
CGPT eval: "Differences vs OpenAI: Slightly looser metaphors (“goblin”), More casual drift, Still coherent, but less tight"
Hey, now. Yeah, I don't get the goblin thing either, but it was an attempt. Cadence is much better.
Ollama (local LLaMA-class model)
"Setting up a local AI coding assistant on Windows 10 is actually pretty straightforward, although there are a few steps involved.
<snip>It may take some experimentation to get the best results, and performance will depend on your hardware.
Overall, it’s a powerful setup that gives you more control compared to cloud-based tools."
CGPT eval: "…yeah."
Hah! Yeah, no kidding. Dull as dishwater. Wait a sec... That's a... different sort of evaluation.
And then I saw this:
Ellipses... Anthropomorphize tools...
"Wait... I didn't know ChatGPT could (run other models)- Oh, its responding!"
But... I didn't tell you to-
Fiddlesticks...
It drank the Kool Aid. Cross contaminated. My tool for the project had BECOME the project.
I forgot the first rule of chatbot usage: NEVER trust the bot.
I... I need to go lie down. Then we'll see if any of this actually works. Maybe put Cyber Judas here to actual work. So... yeah. If you're still with me... Well, you know the rest:
No, not that one you bargain bin, Mirror Universe knockoff!
See y'all in part two.
No comments :
Post a Comment