01conversation · tone · 18+

How to talk to an AI girlfriend past the first ten minutes

Most conversations stall for the same three reasons, and none of them is the model. Here is what to open with, when to move the register, and how to tell a good thread from a polite one.

Fictional AI characters only. Nobody here is a real person.

In conversationThread open
Badge photograph on the talk to AI girlfriend page
On shiftregister: dry

Tone changed mid-thread

Switch to someone else

Round portrait of a fictional AI character in warm lamplight
Round portrait of a fictional AI character against a pale wall
Round portrait of a fictional AI character with her hair pinned up
Round portrait of a fictional AI character working the night shift
+2
Tone change
next message
New thread needed
no

Fictional AI characters only. Nobody here is a real person.

  • Openers
  • Tone control
  • Long threads
  • 18+

02The detail

Why a conversation stalls, and what actually restarts it

A thread dies from interrogation, from politeness, or from a mismatch you never corrected. All three are fixable in one message.

The first failure is interrogation. If every message you send is a question, you will get answers and nothing else — and answers are the least interesting thing a written character produces. People who talk to an AI girlfriend well tend to state things instead: an opinion, a complaint, something that happened. That gives her something to disagree with, which is where a conversation actually lives.

The second is politeness on both sides. A character written as blunt will still open carefully, because openings are careful. Push once — contradict her, or answer a question sideways — and the temperament shows up. If it does not survive two pushes, that is worth knowing about the product, not about you.

The third is a register you never corrected. Ask for drier, softer or more direct in plain words, in the middle of the thread. The next message answers in it. Most people never try this, then conclude the character is flat, when what they had was a default nobody adjusted.

What works well

  • Statements open a conversation wider than questions do
  • A single push tells you whether the temperament is real
  • The register can be changed mid-thread, in plain language
  • Six written personalities means a mismatch is fixable by switching
  • Everything here is testable free, in one evening

Worth knowing first

  • Two of the six characters read more alike than their descriptions suggest
  • A very long session can drift; a fresh subject brings it back
  • Memory of fine detail from weeks ago is patchy
  • Every character is fiction, written as an adult, and never a real person

03On this page

Three conversations that went somewhere

Not scripts to copy — the shape of what tends to work.

Fictional AI character listening, photographed in daylight

Open with something that happened, not with a question.

Fictional AI character at a desk, halfway through a long answer

Contradict her once and see whether the temperament holds.

Fictional AI character on a balcony, still talking after midnight

Ask for a different register in plain words, mid-thread.

04In practice

What a good second session looks like

Come back a day later and say nothing about the last conversation. What she raises first tells you how much of it survived, and whether the thread is worth continuing. This is also the only real test of memory, which cannot be judged inside a single sitting.

If she opens with the specific thing you were arguing about, the file is working. If it takes a nudge, that is the honest ceiling of the memory, and it is better to find it in week one than in month three.

05At the counter

Talking to her — what people get wrong first

01

What should I say first to an AI girlfriend?

Something that happened, rather than a question. A statement gives her a position to agree or disagree with, and disagreement is where a conversation gets interesting. Questions produce answers, which close a thread rather than open one.
02

Why do the replies feel generic at the start?

Openings are cautious by design, in the way first conversations are. Push once — contradict something, or answer sideways — and the written temperament shows up. If two pushes change nothing, the character is thinner than advertised.
03

Can I ask her to be less polite?

Yes, in plain words, in the middle of the conversation. The register changes from the next message rather than the next thread, and it stays where you set it until you move it again.
04

How do I know if I picked the wrong character?

If two pushes and a register change all land flat, the mismatch is in the pairing rather than in the app. Switching is free and the roster is short enough to try three in an evening.
05

Does talking more make her better at talking to me?

Within a thread, noticeably: names, positions and running jokes accumulate and get reused. Across months the fine detail blurs, so the improvement is real but not unlimited.

07Check in

Open a thread and push once

Say something that happened, contradict the first reply, then ask for a drier register. Three messages tell you more than an hour of small talk. Free to try.

Wide night shot of a lit reception counter with the chat open on a tablet

The desk is open. Free to start, nothing to install.

Check in free