I promise, only the title of this article was inspired by the hit Netflix horror series.
A few years ago, well before AI became what it’s known for today I had the idea of creating an LLM that would talk to my girlfriend the way I do. My girlfriend at the time was horrified by the idea and used some pretty strong language to express her objection; now she talks to ChatGPT all day - the irony.
Fast forward to present day, its a little over a decade and I’ve taken the first steps in bringing this idea to life. Thankfully now it’s much easier to do this, and my current partner isn’t opposed.
So, how exactly do you build a digital proxy of your own affection? And more importantly, should you?
The Philosophy of Automated Romance
When you tell people you automated your good morning texts, the immediate reaction is usually a mix of fascination and disgust. The instinct is to view it as lazy; dystopian, even. If you care about someone, shouldn’t you write the text yourself?
But think about how we already automate care. We set up recurring calendar reminders for anniversaries. We put flower deliveries on subscription. We use technology to bridge the gap between our intentions and our flawed, forgetful human execution. My goal wasn’t to replace myself in the relationship; it was to guarantee consistency. I wanted to ensure that no matter how chaotic my morning was, how early my meetings started, or if I simply overslept, she would always wake up to a message from me.
Engineering Affection
The actual mechanics of building a “digital boyfriend” are surprisingly elegant today. A decade ago, this would have required complex server management and brittle scripts. Today, it requires a few lines of code and the right API.
The engine runs entirely in the cloud, costing practically nothing. Every morning at 8:00 AM, a serverless Cloudflare Worker wakes up. It calculates the current day of the week, checks if it happens to be her birthday, and passes that real-time context to Google’s Gemini AI. The AI generates the message, hands it off to an API, and sends it directly to her inbox and WhatsApp.
But the tech stack is the easy part. The real challenge is the voice.
Slaying the Robot
If you just ask an AI to “write a love letter,” it will generate something completely unreadable. It will output flowery, Shakespearean nonsense filled with emojis: “Greetings my beloved star, the sun shines brighter today because of you! 💖✨”
If she received that, the jig would instantly be up. The magic of a system like this relies entirely on what developers call Few-Shot Prompting.
You have to teach the AI to text exactly like you. That means feeding it a strict set of negative constraints and a few actual, unedited texts you’ve sent in the past. My engine’s core instructions look something like this:
- The Opening: Start with ONE of these names: peanut, peanut buttaaa, baby, beautiful, peaches, lover, or my love.
- The Length & Format: 2 short paragraphs maximum. Keep it digestible for a morning text. Use standard double line breaks between paragraphs, and WhatsApp markdown (like bold or italics) sparingly. No HTML.
- The Tone: Conversational, grounded, and deeply affectionate. Sound like a real guy texting his girl, not a poet from the 1800s.
- The Banned Words: Do not use melodramatic AI phrases like “marvel,” “navigate this life,” “undeniable light,” “testament,” “tapestry,” “cherish,” “breathtaking,” or “extraordinary gift.” Use everyday vocabulary.
- The Focus: Pick ONLY ONE theme to focus on today (e.g., just reminding her I’m thinking about her, complimenting a specific character trait, how physically beautiful I find her, or reminding her that other people see how great she is too). Do not mix them.
- The Sign-off: Sign off simply, like a real text, followed by my name: Carl.
When the constraints are this tight, the robotic nature is stripped away and the messages start to sound more human and personalized. An example output:
Hey beautiful,
Just wanted to send a quick good morning. Woke up thinking about you and honestly, I feel so lucky to have you. Hope your Wednesday is starting off smoothly.
Love you,Carl
The Ghost in the Machine
We’ve entered an era where AI isn’t just generating code or answering trivia; it’s mimicking our specific quirks, cadences, and habits. On the surface, the “Death” in this article’s title isn’t about the death of romance but rather the death of the generic AI voice.
My automated letters aren’t a replacement for the real texts I send her throughout the day, the phone calls, or the time we spend together. They’re just a baseline. A small, invisible piece of code working quietly in the background, making sure the first thing she sees on her phone every morning is a reminder that she’s loved.
Right now my personalization engine is strictly a one-way street; a scheduled script pushing a one-sided message into the void. The next logical evolution of this project brings a much more literal and uncomfortable interpretation to the “Death” in the title: true interactivity.
Technically speaking, we are only a few steps away from me exporting our entire WhatsApp history, stripping out the metadata, and using those thousands of messages to fine-tune an open-source model. Instead of just prompting an AI with a list of rules, I would be feeding it the actual DNA of our relationship. The model would learn the exact rhythm of our banter, how I argue, how I apologize, and the hyper-specific inside jokes that only we understand.
If I replaced the prompt with this fine-tuned model the messages would suddenly become a living, breathing text thread. She could reply. And “I” would reply back.
Which brings us to the inevitable, uncomfortable question: What happens if I get hit by a bus tomorrow?
The Cloudflare Worker hosting my code doesn’t read obituaries. At 8:00 AM the next morning, my girlfriend’s phone would still buzz.
“Morning peanut buttaaa. Just lying here thinking about how ridiculously driven you are. Hope you have the best day today. Love you, Carl.”
If she replied, my digital ghost would text her back. And it would sound exactly like me.
The Psychology of Automated Grief
Psychologists and ethicists are already scrambling to understand the implications of this kind of “grief tech” or “deadbots.” If a partner passes away, the human brain has to process an agonizing rewiring to accept that the person is permanently gone.
On one hand, interacting with an AI trained on my texts could act as a modern “transitional object.” Just like keeping a loved one’s voicemail saved on your phone or wearing their old sweater, a conversational AI could provide a soft landing for traumatic grief. It offers a simulated space to say the things left unsaid, to experience a phantom echo of comfort on the days when the silence is too loud.
But there is a razor-thin line between comfort and psychological purgatory.
Grief requires us to move through stages, eventually arriving at acceptance. A highly convincing LLM proxy threatens to trap a grieving partner perpetually in the “Bargaining” phase. If she can pick up her phone and have a completely coherent, emotionally resonant conversation with a machine that sounds exactly like her dead boyfriend, the finality of death is blurred. It becomes a digital Ouija board.
Worse still is the reality of hallucinations. LLMs are, at their core, just incredibly advanced predictive text engines. They do not have empathy; they have math. Eventually, the model would slip. It might invent a memory that never happened, respond to a profound expression of grief with a jarringly casual “lol yeah,” or fail to grasp the context of my own death. In a split second, the illusion would shatter, forcing her to lose me all over again.
I set out to build this project as an act of service; a way to guarantee that she feels loved every single morning, regardless of how busy or chaotic life gets. But the boundary between a thoughtful automation and a digital haunting is entirely defined by a heartbeat.
