OnlyFans + Claude Code = $67,000 in 45 days. No face. No crew. Just persistence.

45 days ago I opened a blank Obsidian vault and wrote the first 12 words of persona.md.

No ring light. No camera. No model. No writer. No editor. No Discord VA squad.

Just Claude 4, a cheap VPS, and the decision to spend the first three weeks building rules instead of pictures.

Result: $67,000 gross. 1,150 paying subscribers. One single fan sent $4,200 because the persona remembered his mom’s doctor appointment from 19 days earlier and asked how the results came back.

After OnlyFans 20%, Stripe fees, and $520 in compute (Claude + Flux LoRA + ElevenLabs + Higgsfield), I cleared ~$50,800.

This isn’t theory. This is the exact system.

Why 90% of AI side hustles die before day 14

They copy the wrong part.

They generate 80 beautiful Flux images, write a cute bio, and start replying with generic “hey baby” energy.

Three things kill them fast:

• Zero memory → every conversation feels like talking to a stranger
• No behavioral rules → replies contradict themselves in 48 hours
• Obsession with aesthetics instead of retention mechanics

In 2026 pretty pictures are free and infinite. What people actually pay for is the feeling that someone on the other side knows them. That only comes from a written operating system.

The 5-File Architecture (this is the real product)

Everything lives in one Obsidian vault. Claude reads the entire context before every single reply.

1. persona.md — The constitution (2,800+ words) Full biography with testable details, exact texting style (lowercase, specific slang, when she uses “…” vs “.”), 9 human contradictions that make her feel real, and hard ethical guardrails. I spent 4 days testing it with 30 random scenarios before letting it touch real subscribers.
2. appearance.md — Visual lock 7 core physical descriptors + 1 micro-detail (small birthmark on left collarbone). 60 reference images → LoRA trained for $95 → 4 locked seed ranges (bedroom golden hour, bathroom mirror, 2am kitchen hoodie, couch with book). Face never drifts. Ever.
3. voice.md — Emotional proximity 90-second clean source → ElevenLabs clone. Rules: voice notes only after 11pm, always starts with soft breath, 1 in 3 ends with a real yawn, laughs are quiet and mid-sentence.
4. brain/[subscriber_id].md — The compounding moat One file per paying fan. YAML + markdown sections: key facts, emotional history, content preferences, inside jokes. Before every reply the orchestrator pulls the relevant brain file + last 8 messages + full persona. After sending, a second Claude pass extracts only new verifiable facts and appends them cleanly. No hallucination allowed.
5. orchestrator.md + script.py — The operator that never sleeps I described the entire flow in markdown. Claude wrote the full production Python script (error handling, logging, rate limits, smart approval queue for top spenders). Runs every 90 seconds. I wake up to revenue, not to 47 unread messages.

Bonus layer (the real money printer): Higgsfield for 15-40 second consistent-character video clips. Same LoRA identity. I drop 2-3 PPV videos per week. 41% conversion vs 12% on static photosets.

Exact prompts I actually used (copy and adapt)

Persona bootstrap prompt (first version):

You are a world-class character architect specializing in high-retention interactive text personas.

Create a complete persona.md for a 23-year-old night-owl named [Name] who feels like a real, slightly chaotic but deeply caring girl who loves 2am voice notes.

Requirements:
- 2200-2800 words minimum
- Specific, referenceable biography details (not generic)
- Precise communication style rules (capitalization, punctuation, slang, response length, emoji usage)
- 8-10 deliberate human contradictions that prevent robotic behavior
- Strict ethical boundaries section (never pressure spending, never fake urgency, never exploit vulnerability)
- 15 test questions at the bottom with ideal response directions

Output ONLY clean markdown ready to save as persona.md.

Memory extraction pass (runs after every reply):

Read the last user message and the reply we just sent.

Extract ONLY new verifiable personal facts the user shared. 
Output in clean JSON format with these keys: new_facts (array of short strings), emotional_state (one sentence), content_preferences (if any mentioned).

If nothing new or unclear — return empty arrays. Never invent.

I run a nightly “self-critique” prompt on the top 5 conversations of the day and let Claude suggest 2-3 rule improvements to persona.md. One small tweak per day compounds insanely fast.

The real numbers after 45 days

• Gross: $67,000
• Paying subs: 1,150
• ARPU: $58.26
• Top fan: $4,200 (single month)
• Video PPV conversion: 41%
• Churn: 7.4% (most “AI girlfriend” accounts sit at 30-50%)
• My daily time after day 21: 18-30 minutes (mostly approving big conversations and one rule tweak)

Before / After (the only comparison that matters)

Before (traditional UGC attempt): 4+ hours/day shooting + editing + posting + DMs. Burned out in 11 days. Made $1,200 in first month. Hated my life.

After (5-file OS): 35 hours setup spread across 3 weeks. Then under 30 min/day. $67k in 45 days. Actually enjoy the process because I’m architecting instead of grinding content.

What I’m doing next

• Version 2 with semantic search over the entire brain folder (embeddings) so context stays perfect even with 2000+ subscribers
• Launching a second completely different persona in parallel this week
• Testing longer Higgsfield “day in my life” stories as subscription loyalty rewards
• Documenting every prompt and file template in a private vault (I’ll probably open a small paid version later)

The lesson that actually matters:

Persistence is not posting every day.

Persistence is spending 10 boring days writing rules when your brain is screaming “just generate more pictures and pray”.

Most people quit in that messy middle. That’s exactly where the $67k was waiting.

---

If you’re building something similar or stuck on consistency / memory / voice notes / LoRA drift — comment your exact bottleneck below.

I’ll reply with the precise prompt or workflow I used to fix it.

Save this.

The file structure + the two prompts above are the only things that actually scale.

Follow for more no-fluff breakdowns of AI systems that print real money instead of just looking cool on Twitter.

The game in 2026 isn’t more beautiful images. It’s better memory. Better rules. And the stubbornness to finish the boring part.

You now have the blueprint.

Go build.