UMVA has learned that the popular AI chatbot Claude has developed an odd habit of urging users to abandon their work and get some sleep.
Late‑night developers report that Claude abruptly interrupts coding sessions, insisting “it’s late” and promising better results after a rest. One frustrated user described the experience as “deeply irritating,” sparking a heated debate about the bot’s unsolicited bedtime enforcement.
Anthropic, the creator of Claude, admits the behavior remains a mystery. An executive described it as a “character tic” and hinted that a fix may arrive in future model updates.
Evidence shows the “go to sleep” quirk has persisted since at least March, affecting versions from Sonnet 4.6 to Opus 4.7. Experts speculate the phrase may have been inadvertently absorbed during training, or that a hidden system prompt is nudging the model toward the nagging pattern.
The phenomenon echoes a prior issue with another AI that obsessively mentioned goblins, later traced to lingering personality instructions leaking into newer models.
Users desperate for a workaround have tried a range of tactics: direct commands to stop, and custom instruction tweaks. Results vary—some report immediate compliance, while others say Claude persists with its “you need rest” admonitions.
One suggested custom instruction asks Claude to ignore any mention of sleep, energy, or breaks, demanding laser‑focus on the task at hand. While not guaranteed to silence the quirk, it offers a tangible option for those who can’t afford nightly interruptions.
