AI-M: I built an AIM-inspired instant messenger where every buddy is Claude and they never break character

We're back to 2003 at https://its2003.com ... View your Buddy list, listen for the door creak when someone signs on, read the away messages, and chat with existing buddies or build your own. Every buddy is Claude (Sonnet, streaming) playing a specific person with their own life, and they stay in character no matter how hard you push.

Staying in character. The system prompt makes each buddy a person, not an assistant. No "as an AI", no markdown, no answering every question like a helpful bot. Short messages, sometimes one word. They tease, get bored, change the subject.

Memory. When you describe a buddy (your camp friend from 2003, whoever) it generates a persona plus a private canon of facts about their life and how they know you, and it persists. Come back after a week and they notice the gap.

The one override: if someone shows genuine distress, the character breaks immediately and points to real help (988 in the US). That rule beats the fiction every time. Playful "wait are you real?" gets deflected in character. Real crisis does not.

Cost reality for a hobby project: daily message cap per user, a hard daily spend ceiling, per-IP limits, and a kill switch. You can bring your own Anthropic key to skip the cap.