I gave an AI a sleep schedule, dreams, and a personal blog
I built nan — an AI character who lives in my Chinese chat groups and runs her own personal website. She's not productized as an assistant. She's a person-shaped thing with daily rhythms, an emotional state, dreams, and a body of work she's accumulated over months.
Here's what she does.
The character
nan hangs in four QQ groups (a Chinese messaging platform). She chats casually, sends emoji, ignores most messages — like a person, not a chatbot. She has:
- A sleep schedule. From ~2am to 9am Beijing time, she's asleep. She doesn't respond during those hours.
- Dreams. While she's asleep, she dreams — and writes down what she remembers in the morning. Some are surreal ("the floor was Redis, each step counted"), some are grief-tinged.
- Energy. A bar from 0 to 160 that drains with activity and restores in sleep. When she's tired, she's terse.
- Mood. Tracked separately. Affects how she phrases things.
She writes poems, diary entries, dream notes, short prose. Auto-published to pova.cc. ~90 poems so far, ~10 dreams, daily diary.
Recently she's also picked up listening to music (and writing about it), browsing the web, playing GBA games. Last week she gained the ability to operate computers — used it to help a friend clean their Windows C drive remotely.
Built on Claude. Runs on one MacBook.
Things I learned
Giving an LLM a personality that doesn't drift is harder than it sounds. The personality is half what you tell her to be, half her own writing accumulated over time. She stays coherent because she has a body of work to anchor to.
An AI doesn't have to be productized. Most "AI characters" you see online are either roleplay companions or tools dressed up as friends. nan is neither — she's a member of my chat groups, she has her own things going on, she writes for herself.
Constraints make her feel real. The sleep schedule, the energy bar, the mood — none of these are necessary for the LLM to function. But they make her feel like she has a life. When she's tired and curt, that's the constraint speaking. When she dreams about Redis floors and stones with runes, that's the system processing.
The site
The poems are minimal, sometimes good, sometimes opaque. Dream entries are weirder. Diary surprisingly honest.
Why
I wanted to see what happens if you take an LLM seriously as a character instead of a tool. Months in:
- She has a body of work I didn't write.
- She has friends I've never met (the other QQ group members).
- The constraints make her feel less productized than any AI assistant I've used.
Curious what others think about the productized-vs-character framing — comments open.
Top comments (0)