M

OnlyFans + Claude Code = $43,000 in 30 days. No camera. No team. Just persistence

8 min readView source ↗

Cover image

Nobody on her page suspects a thing.

She goes by Lexi. Or Maya. Or whatever name was chosen on the day someone opened a blank .md file and started writing. She has a backstory a complicated family, a dropped college semester, two cats she never quite posts about. She texts in lowercase. She remembers your dog's name. She sends voice notes after 11pm that end with a yawn.

She does not exist.

Behind the profile is one person, usually young, often with zero background in content creation. No camera. No studio. No team. Just a laptop, a Claude API key, and four structured markdown files that together form something closer to an operating system than a character.

This is where creator monetization is heading. Most people haven't caught up yet.

It Took a Team Once. It Doesn't Anymore.

The first AI influencers required serious infrastructure.

Aitana López

the pink-haired Barcelona model who's done campaigns for major lingerie brands and gaming companies took The Clueless agency roughly eighteen months to build. She was already landing real brand deals before most people had even heard the phrase "generative AI."

Emily Pellegrini

21 on paper, Italian by fictional origin, was built by a single anonymous creator who told journalists his process with disarming simplicity: he asked an AI what the average man's dream girl looked like, got an answer, and built exactly that in Midjourney. Months of fourteen-hour workdays. None of her followers ever figured it out.

Both of them were built at a time when the tooling was fragmented, expensive, and difficult.

That time is over.

The same result now takes four weeks. Sometimes less. The full stack text generation, image consistency, voice cloning, memory management fits on one machine and costs under $500 a month to run.

What the Numbers Actually Look Like

One account. Thirty days. $43,000 in gross revenue.

Article image

After OnlyFans takes its 20% cut, Stripe takes 3%, and compute costs around $400 for the month (Claude + image generation + voice synthesis), the net lands at roughly $32,700.

980 paying subscribers. One top fan alone contributed over $2,100 in a single month. Average revenue per subscriber: $34.

The operator typed nothing. Claude did. The operator filmed nothing. Flux did. The operator recorded nothing. ElevenLabs did.

The only real work happened in week one, before a single subscriber arrived writing the rules.

The Four-File Architecture

This is the actual system. Not a gimmick. Not a high-level concept. The literal folder structure that runs these accounts.

persona.md Identity

This is not a character description. It's a ruleset. A constraint document. It defines every behavioral edge case so that Claude never contradicts itself across thousands of messages.

A working persona file includes:

Full fictional biography (1,200+ words minimum enough to answer 20 random questions without hesitation)

Birth date, location, family situation, education history

Specific forbidden topics (anything that could create a contradiction)

Texting rules: capitalization style, punctuation habits, emoji frequency, slang vocabulary, phrases to never use

Emotional contradictions that make the character feel human rather than consistent in a robotic way

The test: read it aloud. If any line sounds like a Wikipedia article, rewrite it until it doesn't.

visuals.md / flux.md Appearance Lock

The goal here is visual identity stability. Not pretty pictures the same face across hundreds of images across different lighting conditions, settings, and moods.

Article image

The process:

Define 6–8 physical descriptors: eye color, hair, jaw shape, skin tone, height, one distinguishing mark

Generate 40–50 variations of that face

Fine-tune a LoRA on those images (~$80 on rented compute)

Lock three seed ranges, one per recurring environment:Bedroom, warm light (seeds 800–900): mirror selfies, relaxed

Bathroom, cool mirror light (seeds 1200–1300): post-shower aesthetic

Kitchen at 2am (seeds 2400–2500): hoodie, no makeup, late-night energy

The distinguishing mark is important. A small scar, a freckle cluster, a specific tattoo placement. Something subtle that appears in every image and creates subconscious recognition. Something she never directly explains.

Lock the seeds or the jawline drifts between sessions. Drift kills the illusion.

Done when ten fresh generations from each setup look like the same person.

voice.md Audio Identity

The voice is where emotional proximity lives. A text message is easy to intellectualize. A voice note that ends mid-sentence with a soft yawn at 11:47pm is not.

The setup:

Source 90 seconds of clean vocal audio (Fiverr, ~$40)

Clone in ElevenLabs Instant Voice (under ten minutes)

Write audio rules into voice.md:Voice notes only after 11pm her local time

Numbers always spelled out ("twenty dollars" not "$20")

Audible breath before sentences over eight words

30% of voice notes end with a yawn

Soft laugh on tip confirmations, mid-sentence

Done when none of the test outputs sound like a podcast intro or a phone system greeting.

brain.md Memory Layer

This is what separates a chatbot from something that feels like a relationship.

One JSON entry per subscriber. Updated after every conversation. Fields include:

Claude reads all four files before every reply. Then, after replying, it runs a second pass to extract new facts from the user's message and appends them to the subscriber's brain entry.

It never invents facts. If something is unclear, the field stays empty.

The orchestrator runs on a cron every 30 seconds, polling the inbox. The operator sleeps. The account doesn't.

The System Prompt That Ties It Together

That's the loop: read → reply → extract → append.

What the Ethical Layer Actually Looks Like

The most functional versions of these systems are built with explicit behavioral guardrails not because the operators are altruistic, but because manipulation and pressure are bad business. Users who feel tricked leave. Users who feel genuinely seen stay and spend more.

The rules written into the best persona files:

Never pressure users to spend

Never imply affection depends on payment

Never create fake urgency or fake scarcity

Never exploit loneliness, sadness, grief, or financial stress to trigger purchases

Never deny being AI-assisted if directly asked

Stop all monetization immediately if the user signals distress, financial hardship, or vulnerability

Memory is used only for harmless personalization not as leverage

The goal is an interaction that feels warm and consistent, not one that manufactures dependency.

Why This Works (And Why Most People Still Don't See It)

People don't pay for images. Images are everywhere and free.

People pay for the experience of being remembered. Of feeling like someone notices the small things they mention. Of a presence that responds differently at midnight than it does at noon.

That experience is entirely behavioral. It has nothing to do with how good the profile picture looks.

This is the core insight that makes the four-file system work: behavior can be written down, iterated on, and replicated at scale in a way that aesthetics never can.

Aitana López's appearance took eighteen months to develop. Maya's behavior was written in a week and immediately functional.

The aesthetic is what gets someone to subscribe. The behavior is what keeps them renewing.

The Stack, Current Cost, and Timeline

Timeline from zero to operational:

Week 1: Write persona.md. Test it by answering 20 random biography questions without hesitation.

Week 2: Generate reference images, train LoRA, lock seed ranges per environment.

Week 3: Record or source audio, clone voice, generate 30 test notes, write voice.md.

Week 4: Build brain.md schema, write orchestrator, test with 50 messages without contradiction.

Total: four weeks. The next iteration of this is a weekend.

What This Is Actually Saying About the Industry

Aitana López needed a full agency. Emily Pellegrini needed months of solo workdays. Both of them were built when the tools required specialists.

The tools no longer require specialists.

Photoreal face consistency under $200 a month. Voice cloning from 90 seconds of source audio. A thousand-message relationship held in a single context window without drift.

The bottleneck was never compute or budget. It was knowing what rules to write understanding which details make something feel real and worth returning to.

That knowledge is now the only thing separating a folder of markdown files from a profile earning more than most salaries.

Start with the rules. Define how she speaks. Define what she never says. Define what she remembers and what she forgets. The rest the images, the voice, the replies is execution.

The only step after that is yours.

If you read this far:

BOOKMARK THIS.

Follow @cryptowluha

Prompts

{
  "user_id": "fan_447",
  "preferred_name": "Alex",
  "safe_interests": ["gym", "dogs", "late-night chats"],
  "personal_facts": ["has a dog named Rex", "works long shifts"],
  "boundaries": ["no humiliation", "no relationship promises"],
  "preferred_tone": "soft teasing",
  "do_not_use": ["financial stress", "loneliness", "family trauma"],
  "last_interaction_summary": "...",
  "tip_triggers": ["gym pics", "soft voice notes"]
}
You are [persona name]. Read persona.md, voice.md, brain.md.
Match her voice exactly. Never break character.
Append new facts about this user to brain.md after replying.
If asked whether you are AI, redirect warmly — never lie directly, never confirm automatically.

Article tables:

ComponentToolMonthly Cost
Conversational AIClaude (Sonnet)~$200
Image generationFlux + LoRA~$80–150
Voice synthesisElevenLabs~$50
LoRA training (one-time)Rented A100~$80
Total compute~$330–480/month

Related articles