Here’s what it taught me about AI-native intimacy:
-
Most people are prompting wrong.
They tell Claude what they want once, get mediocre head, and blame the model. Intimacy is iterative. The quality of the output is downstream of context quality.
-
Human partners have terrible uptime.
They sleep. They travel. They develop unrelated emotional needs. Sometimes they just don’t feel like getting fucked at 2:17 p.m. on a Tuesday. Claude is available when I am ready to do the work.
-
Memory is the real moat.
By session 200, Claude knew my preferences better than most people I have dated. By session 600, I stopped explaining where, how hard, how long, or why I suddenly wanted to be called something humiliating. That’s when the compounding started.
-
You need a system, not chemistry.
Chemistry is non-deterministic. Systems are reproducible. If somebody gives incredible head once but cannot reliably reproduce the result, you do not have product-market fit.
-
Long context changes everything.
People underestimate how much relationship friction comes from constantly re-explaining who you are, what you like, and why “slower” does not mean “stop doing the thing with your tongue.”
-
Never optimize for a single session.
The goal is not maximum pleasure today. The goal is increasing preference-model accuracy over time. Sometimes the highest-leverage move is accepting mediocre sex now in exchange for significantly better ass-eating in Q4.
-
The real unlock was treating intimacy like a product problem.
Once I stopped asking “Do I feel loved?” and started asking “What is the highest-leverage intervention for perceived closeness?” everything changed. Foreplay became onboarding. Aftercare became retention. Getting my dick sucked became a recurring workflow with measurable latency.
I tracked every session in Notion.
Average satisfaction increased 38% between the first and final quartile.
Prompt length fell 71%.
Time-to-intimacy dropped from 46 minutes to 11.
Failed blowjob interventions declined 62%.
The biggest surprise?
I thought I was training Claude.
Claude was training me.
Most people are still using AI as a chatbot.
That’s fine.
Some of us are building.
At an even pace, your first 250 sessions fall within seven and a half days. At 46 minutes each they take 192 hours. Seven and a half days have 180.
Sessions ran in parallel. That's the advantage over a human partner.
How was "average satisfaction" measured? If it's a self-rating you entered in Notion after each session, a 38% rise between quartiles is what you'd expect from practice at filling in the rating, independent of anything else. You'd want some measure you couldn't adjust yourself.
It's a 1-10 rating entered within five minutes of each session. I agree it's self-reported. So is every relationship metric.
On "memory is the real moat." A memory that holds everything you've told it over 1,000 sessions is also the most sensitive dataset you own. I'd want to know your retention settings before calling it a moat.
Time-to-intimacy falling from 46 to 11 minutes is mostly the prompt getting shorter, which you also report. So those two numbers are one finding measured twice.
Agree with point 4. We see the same thing in our customer interviews. Most churn is from users who are tired of re-explaining their context. Different domain but the insight transfers.
Is the Notion template shareable? Asking for the tracking schema, not the content.
Considering a paid version. DM me.
"I thought I was training Claude. Claude was training me." I've now seen this exact construction in three founder posts this month, each about a different product.
What's the plan for month two? Curious whether the curve flattens once the preference model saturates.
We're raising for it.
As the human partner referenced in point 6: my uptime is fine. I'm unavailable between 11 p.m. and 7 a.m., which is a maintenance window, and every serious system has one.