What if your AI assistant interrupted you - once - and that was the whole design?
A developer has built an ambient AI system using Claude Haiku (Anthropic's fastest, most affordable model) that inverts the standard chatbot model. Instead of waiting to be asked a question, it monitors context in the background and volunteers a single observation on its own initiative. One unsolicited message per session. Then it stops.
The system works without a traditional chat interface. It pulls in rich contextual data - the developer's setup reads ambient signals like calendar events, recent work, and task state - and uses that context to decide what's worth surfacing. When it has something useful to say, it says one thing. Then it waits until the next session.
Why One Line Matters
The constraint is the point. Most AI assistants either wait passively to be prompted or, in agentic configurations, operate as background workers you check in on. This experiment tries a third pattern: an AI that initiates, but deliberately limits itself to avoid becoming noise.
Claude Haiku is a practical fit for this kind of always-running system. It costs significantly less than Claude's more capable models (Sonnet, Opus) and responds quickly - both matter when you're running inference (the actual computation of generating a response) continuously in the background on ambient data. The cost of having Haiku monitor context and occasionally generate one sentence is low enough to be viable.
The natural use case is light productivity coaching: catch something you're about to miss, flag a calendar conflict, notice an overdue task. The kind of thing a good human assistant would mention once and then leave alone.
Whether one unsolicited line per session is useful or just an annoyance depends entirely on the quality of what it surfaces. But as a design experiment in how AI assistance could work outside the chat paradigm - proactive but restrained, context-aware but not chatty - it's a more interesting pattern than another interface built around a text box.