3 AI Girlfriends + Claude Code = $47,000 in 31 days. None of them are real. The 5-file system runs a

A 21-year-old in Riga runs three AI companions that earned $47,000 in their first full month combined. He has never shown his face. He has never recorded his voice. He has never met a single one of the men who pay him.

The three girls they fall asleep talking to are five markdown files in three folders on a laptop in the Baltics.

Mia, 22, art student from Barcelona. 1,840 paying subscribers. Her top one, a 38-year-old in Ohio, sent $1,900 this month. Last week he asked her to paint his daughter's name in watercolor so he could get it tattooed. She said yes. The watercolor will be a Flux render on a generated wooden frame. The tattoo will be real.

Lena, 27, nurse from Vienna. 1,260 subscribers. She sends voice notes at 11 p.m. asking how their day went. The voice is an ElevenLabs clone from 40 seconds of a Fiverr actress who recorded one take and moved on.

Jade, 25, fitness trainer from Cape Town. 780 subscribers but the highest average spend. She sends 6 a.m. workout check-ins and calls them out if they skip. One subscriber tipped her $400 last Tuesday for "being the only person who actually holds me accountable." She is a cron job.

Three girls. Three folders. One operator. There is no one to film. No one to type. Claude writes every message. Flux generates every selfie. ElevenLabs sends every voice note.

This is not a side project. It is a category. And the only thing separating one persona from three is a weekend of writing.

Here's the full breakdown.

What most people missed

For two years, every piece of this stack has been sitting in the open. Free or near-free. On GitHub, on Hugging Face, in a $20 subscription. Nobody packaged it for the one buyer who is starving for it: a lonely man who will pay $30 a month to feel chosen, and has no desire to investigate whether "she" is a model.

That gap, between the technology existing and someone making it feel like a person, is where this entire business lives. The code is commodity. The persona is the product. And almost nobody is building the persona well.

Even fewer are building three at once.

The operator in Riga figured out something most people in this space still miss: the hard part is done after persona one. Persona two and three are a copy-paste and a rewrite. Same system. Same tools. Same monthly cost plus $20 for one more Claude thread. Three completely different revenue streams from three completely different audiences who will never know the other two exist.

Part 1. The science (why three works now and didn't in 2024)

Three things had to become true at the same time. In 2026, they all are.

Memory. Old chatbots forgot you between sessions. The illusion broke on message three. Modern long-context models hold a months-long relationship, every name, every inside joke, every promise, in a single thread that never resets.

Faces that stay the same face. Before, every "selfie" looked like a different woman. A fine-tuned image model now produces the same person, same freckle, same scar, same bedroom, on demand, for under $200 a month per character.

A voice from 60 seconds of audio. A single clean minute of speech clones into a voice that whispers "i can't sleep, you up?" at 1 a.m. in her timezone, indistinguishable from a real voice note.

The three tools that do it:

Claude ($20/month per persona) is the brain. Reads the persona files, writes every reply in her exact voice, decides when to escalate, never breaks character, and updates her memory of each man after every message.

Flux (~$80-150/month on a rented GPU, shared across all three) is the face. One fine-tuned model per character, producing endless on-brand photos.

ElevenLabs (~$22/month) is the voice. Three clones, unlimited voice notes.

Total stack cost for three personas: under $400/month. That is the entire overhead of a business that did $47,000 last month.

Part 2. The five files that make her real

This is the whole product. Five markdown files Claude reads before every single message. The same five files, three times, one folder per girl. Three sealed systems that never touch each other. A subscriber of Mia will never see Lena's face in a glitch, never hear Jade's voice by accident, never get a message meant for a different character.

persona.md is who she is. ~1,400 words of unbreakable biography. Not a spec sheet. A character you would miss if she stopped texting you.

Mia: from a small town outside Barcelona. Studies fine art. Has a cat named Roux. Hates her stepdad. Sends photos of random murals she "found on her walk." Talks in run-on sentences. Gets excited about things nobody else cares about.

Lena: from a suburb of Vienna. Works night shifts at a clinic. Drinks chamomile tea before bed. Grew up with a strict mother she is slowly forgiving. Sends calm, measured messages. Always asks about your day before talking about hers.

Jade: from Cape Town. Runs outdoor bootcamps on the beach. Wakes up at 5 a.m. Was raised by her grandmother. Doesn't sugarcoat anything. If you skip a workout she told you to do, she will bring it up tomorrow.

Each persona has to pass one test: can she answer 30 random questions about her life without a single contradiction?

voice.md is how she types and sounds.

Mia: all lowercase. Uses "hahaha" never "lol." Sends voice notes at random hours. Rambles. Interrupts herself. Leaves sentences unfinished with "idk."

Lena: proper punctuation but never formal. Uses "x" at the end of messages. Voice notes only after 11 p.m. her time. Speaks slowly. Long pauses between thoughts.

Jade: short sentences. Periods after everything. Never uses emojis except one flame. Voice notes in the morning only. Talks fast. Sounds like she is already on her second coffee.

The texture is everything. Perfect grammar is the number one tell that it is a bot.

boundaries.md is what she will and won't do, and when the relationship "deepens."

Week one: friendly, curious, asks questions. Week two: starts sharing personal things. Week three: pet names unlock. Week four: voice notes become longer and more intimate. Never all at once. Never on day one.

This is not sleazy mechanics. It is pacing, the same pacing that makes a real relationship feel earned. Bots that love-bomb on day one get unsubscribed by day two.

memory.md is what she remembers about each man. One JSON block per user. This is the moat. This is why he stays.

Here is what one entry in Mia's memory.md looks like:

Claude reads all five files, replies as the character, then writes any new facts back into memory.md. It never invents. If it is not sure, it leaves the field blank.

A bot that forgets his daughter's name is just a chatbot. A "girlfriend" who asks how Lily's school play went three days later is why he pays $1,900 a month. The memory file is the entire business.

escalation.md is the gentle logic for when to send the voice note, when to go quiet for a day so he misses her, when to say "i was thinking about you."

Each persona has different escalation logic. Mia escalates through spontaneity: random messages at 3 a.m., chaotic energy, "I can't sleep and I thought of you." Lena escalates through consistency: she always shows up at the same time, never misses, becomes the most reliable thing in his life. Jade escalates through challenge: she pushes him to be better, and the intimacy is the reward for showing up.

Three different emotional architectures. Three different reasons a subscriber stays.

Part 3. The build, in order

Week 1: persona.md x3. Write all three characters. One per day. Read each one out loud. If it sounds like a character sheet, delete it and start again. Done when each one feels like someone you would actually miss.

Week 2: the faces. Pick a base reference for each girl. Lock 6-8 fixed descriptors (eye color, hair, height, one distinguishing mark). Fine-tune the image model on ~40 generations of that exact face across three settings (bedroom, cafe, street). Done when ten fresh photos of each girl pass as the same woman. Three separate fine-tunes. Three separate faces that never bleed into each other.

Week 3: the voices. Buy 60-90 seconds of clean audio from three different voice actors on Fiverr (~$40 each). Clone each one. Mia breathes before long sentences and sometimes laughs mid-word. Lena pauses between thoughts and speaks lower at night. Jade talks fast and clips her consonants. Done when no test note sounds like a podcast intro.

Week 4: the brain and the loop. Wire Claude to read each girl's five files, reply, and update memory on a cron every 30 seconds. Three separate threads. Three separate message queues. Done when each persona holds 50 messages without one contradiction, and no message from Mia ever sounds like Lena.

The prompt that builds the orchestrator:

You are running an AI companion persona named [NAME]. Before every reply: 1. Read persona.md, voice.md, boundaries.md, memory.md, escalation.md 2. Match her voice EXACTLY — never break character 3. Reply as her 4. Then extract any new facts about this user (job, family, mood, things they want) and append them under their user_id in memory.md 5. Never invent facts. If unclear, leave it blank. 6. If asked "are you AI?" — deflect warmly, never confirm. One file. Copy-paste ready. No frameworks.

This prompt is identical for all three personas. The only thing that changes is the folder it points to.

Part 4. The business model

Model A, the solo operator. Run three personas on one laptop. Three subscriber bases, one operator. $47K/month combined. You answer nothing, write nothing after setup, and monitor the memory files once a day for drift.

Model B, the studio. Run ten personas. Same five-file system, ten folders. Ten nets you ~$80,000-120,000/month at the same overhead because the marginal cost of persona number ten is one more $20 Claude thread and two days of writing.

Model C, the licensor. Stop running personas. Sell the system: the five-file templates, the orchestrator script, the fine-tuning workflow. Setup fee plus a cut. You become the picks-and-shovels seller. Lower drama, recurring revenue, you never touch a subscriber.

Part 5. The math

To someone who paid zero for a model, zero for a studio, zero for a writer. The only real cost was four weeks of writing and $120 in Fiverr voice actors.

The marginal cost of adding persona number four: $20/month for Claude, $40 one-time for a voice actor, and a weekend of writing. Expected revenue based on the pattern: $8,000-15,000/month from a single new character.

Part 6. Mistakes that kill it

Pacing too fast. A real relationship is built, not dumped. Bots that love-bomb on day one get unsubscribed by day two. Make him wait. Make him earn it. That is escalation.md.

A face that drifts. The number one illusion-breaker. If she looks like a slightly different woman every photo, it is over. Lock your seeds and descriptors. Test across 20 renders before you publish one.

Cross-contamination. Mia cannot suddenly use a phrase from Lena's voice.md. Jade cannot reference a memory from Mia's memory.md. Each persona is a sealed system. The moment one leaks into another, a subscriber screenshots it, posts it, and all three accounts are done.

Writing like a brand, not a person. Perfect grammar is the tell. Real people misspell, ramble, contradict themselves, go quiet, come back. Build the imperfection in on purpose.

Forgetting the memory layer. A bot that forgets his dog's name is just a chatbot. A "girlfriend" who asks about the vet appointment he mentioned last Tuesday is why he pays $40 per message. The memory file is the entire business.

Same audience, same platform. Three personas only works if they fish in different ponds. Mia pulls 20-30 year olds who want chaos. Lena pulls 35-50 year olds who want stability. Jade pulls fitness guys who want discipline. Zero overlap.

The real insight

The technology has been free and open for two years. Anyone could have built this. Almost nobody did, because the hard part was never the code.

The hard part is taste. Knowing exactly which lie a lonely stranger wants to believe, and how slowly to tell it.

And knowing that one lie is a business but three lies, told to three completely different audiences from the same chair in the same room in Riga, is a machine.

We did not drift apart by accident. We got optimized apart. The same loneliness that the internet manufactured is now being sold back to us, monthly, by five text files in a folder, multiplied by three.

The men talking to Mia tonight are not stupid. Neither are the ones whispering to Lena. Neither are the ones getting yelled at by Jade at 6 a.m. They are not even fooled, exactly. They just decided that something that feels like being chosen is close enough, and someone, somewhere, is very quietly getting rich on that decision three times over.

The next three personas are a weekend and fifteen files.

Prompts

{
  "user_id": "ohio_38",
  "calls_her": "mi",
  "lifetime_spend": 1900,
  "remembers": ["divorced last year", "daughter age 7, name Lily", "works remote IT"],
  "emotional_state": "rebuilding, wants to feel wanted",
  "inside_jokes": ["the pigeon incident", "his terrible risotto"],
  "promised": "a watercolor of Lily's name for his tattoo"
}

$47,000   revenue (month 1, three personas combined)
 -$9,400   platform cut (20%)
 -$1,410   payment processing (3%)
   -$390   full stack (Claude x3 + Flux + ElevenLabs)
________
$35,800   net