Sup HN! I'm Mike Cramblett, a CS grad (Harvey Mudd '02) and developer building for the Build with Gemini XPrize. Let me start with a sincere compliment of my own to you: thanks for your interest in HypeBESTIE(HB)!! I also really appreciate that the community appreciates human writing. I thought I had to have the LLM rewrite my comment for me so I don't sound like a dumb-dumb. Turns you y'all don't mind! Thanx!
Okie dokie! Lemme tell you how I made this thing. So you know in how 2026, you person end up with file or files of random prompts to use on your hard drive? One of the fun things you can do when you're bored is scroll through and look for inspiration or just for laughs. So a few weeks ago, I found a "zoo animal complimenter prompt", that I'd apparently written in this past spring. I can't remember what inspired me to write that prompt, but it was instructions for a coding LLM to make an HB-like agent, that would run on an automated scraper on zoo animal instagram feeds. Mind you, this was pre-Jimothy(that's a term now), do any of you remember who the viral animal was in like, February 2026? I guess I could ask Google. I also guess I could still build my zoo animal complimenter agent. I don't know why I wanted one in the first place though.
I realized that the underlying sincere complimenter prompt could be used for humans too, fairly obviously. I could tell it worked when I uploaded some janky selfies of me sitting at my computer screen in the middle of the night, not wanting to be photographed. But I did it for testing, and science, anyway. But its sincere, but goofy compliments, still made me feel good.
So HB is basically a one step pipeline, driven by a single persona prompt for Gemini 3.6 Flash. I vibe code the React framework, but I had to debug it a bunch. I guess that's part of the story too, but not what I'm writing now. I'm going to focus on my speciality, persona prompting. So, the prompt has four important components, the latter three are all to minimize boring RLHF: regular "you are HB, a sincere complimenter app..." and blah blah instructions including the json schema, a gamified human-ness quality called Turing Grade and Personality Juice (TG/PJ, see my Kaggle research below), a set of philosophical notions about the mathematical nature of love, and a pithy instruction to remove em-dashes. I hope some or all of that surprised you, none of it is random, and it all came out of ad-hoc research I've done.
The first, "normal", part of the HB prompt is pretty much what you'd get if you gave my Show HN title as a prompt to a coding LLM. You can go ahead and try it if you want. Full disclosure: some of you actually will like it better than HB. If you try it, you'll get much shorter, much more generic responses. But they'll also be "reined in" in a way that some of you may like.
The second part is where I start to pull away from RLHF, using my own eccentric AI research. I've got an executive summary version of my TG/PJ notebook on Kaggle, in addition to the full thing, if you're curious about that. But, I'll explain a bit here. Generally, my plan is get the model to get out of the boring RLHF space by measuring the human-like quality of its writing as the reverse of a simulated AI detection algorithm, but with a simulated explanation. With that, it can turn it into a game to optimize the response on, that's not the "RLHF game". Which is about as boring as an improv warm up game.
I also included philosophical notions about love to breathe color and life into the response. Mathematically speaking, a representative sample of everything humanity has ever known or will every know about love has already been tokenized into any well trained LLM. My goal was to give the model a "rubric for sincerity", but without literally telling it how I wanted it to be sincere. So instead I instructed it about everything it already knows about sincerity, and love.
Finally, I told the model to use emojis instead of em-dash. Before that, the whole text was full of em-dashes, it was nuts. I find that if you give the model a reason not to use them or something to use instead, you'll actually get what you want.
Fun fact and science: Originally, HB was going to be called "HypeBFF". But, since I put "you're the user's best friend..." in the prompt, HB said "BESTIE" every time, because it was an obvious attractor state. So I decided that instead of fighting it, I'd just call the it "HypeBESTIE" instead. And thus HB basically named itself. Also HypeBFF sounds even more like a tween TikTok micro-drama...
mcramblett•2h ago
Okie dokie! Lemme tell you how I made this thing. So you know in how 2026, you person end up with file or files of random prompts to use on your hard drive? One of the fun things you can do when you're bored is scroll through and look for inspiration or just for laughs. So a few weeks ago, I found a "zoo animal complimenter prompt", that I'd apparently written in this past spring. I can't remember what inspired me to write that prompt, but it was instructions for a coding LLM to make an HB-like agent, that would run on an automated scraper on zoo animal instagram feeds. Mind you, this was pre-Jimothy(that's a term now), do any of you remember who the viral animal was in like, February 2026? I guess I could ask Google. I also guess I could still build my zoo animal complimenter agent. I don't know why I wanted one in the first place though.
I realized that the underlying sincere complimenter prompt could be used for humans too, fairly obviously. I could tell it worked when I uploaded some janky selfies of me sitting at my computer screen in the middle of the night, not wanting to be photographed. But I did it for testing, and science, anyway. But its sincere, but goofy compliments, still made me feel good.
So HB is basically a one step pipeline, driven by a single persona prompt for Gemini 3.6 Flash. I vibe code the React framework, but I had to debug it a bunch. I guess that's part of the story too, but not what I'm writing now. I'm going to focus on my speciality, persona prompting. So, the prompt has four important components, the latter three are all to minimize boring RLHF: regular "you are HB, a sincere complimenter app..." and blah blah instructions including the json schema, a gamified human-ness quality called Turing Grade and Personality Juice (TG/PJ, see my Kaggle research below), a set of philosophical notions about the mathematical nature of love, and a pithy instruction to remove em-dashes. I hope some or all of that surprised you, none of it is random, and it all came out of ad-hoc research I've done.
The first, "normal", part of the HB prompt is pretty much what you'd get if you gave my Show HN title as a prompt to a coding LLM. You can go ahead and try it if you want. Full disclosure: some of you actually will like it better than HB. If you try it, you'll get much shorter, much more generic responses. But they'll also be "reined in" in a way that some of you may like.
The second part is where I start to pull away from RLHF, using my own eccentric AI research. I've got an executive summary version of my TG/PJ notebook on Kaggle, in addition to the full thing, if you're curious about that. But, I'll explain a bit here. Generally, my plan is get the model to get out of the boring RLHF space by measuring the human-like quality of its writing as the reverse of a simulated AI detection algorithm, but with a simulated explanation. With that, it can turn it into a game to optimize the response on, that's not the "RLHF game". Which is about as boring as an improv warm up game.
I also included philosophical notions about love to breathe color and life into the response. Mathematically speaking, a representative sample of everything humanity has ever known or will every know about love has already been tokenized into any well trained LLM. My goal was to give the model a "rubric for sincerity", but without literally telling it how I wanted it to be sincere. So instead I instructed it about everything it already knows about sincerity, and love.
Finally, I told the model to use emojis instead of em-dash. Before that, the whole text was full of em-dashes, it was nuts. I find that if you give the model a reason not to use them or something to use instead, you'll actually get what you want.
Fun fact and science: Originally, HB was going to be called "HypeBFF". But, since I put "you're the user's best friend..." in the prompt, HB said "BESTIE" every time, because it was an obvious attractor state. So I decided that instead of fighting it, I'd just call the it "HypeBESTIE" instead. And thus HB basically named itself. Also HypeBFF sounds even more like a tween TikTok micro-drama...
Sincere thanks again for your attention!
Mike Cramblett
Live Demo: https://hypebestie.onrender.com GitHub Repo: https://github.com/mike-cramblett/hypebestie TG/PJ Executive Summary: https://www.kaggle.com/code/mikecramblett/personality-juice-...