Dev publishes CLAUDE.md to stop agent session rot

A developer has published a minimal CLAUDE.md configuration file specifically designed to counter what they term 'session rot' in Claude Code — the gradual degradation where agents become slower, more verbose, and less action-oriented as sessions extend over time. The file enforces hard rules against behaviors that compound over long sessions: unsolicited mid-task check-ins, repeated explanations of completed work, scope expansion without prompting, and extended planning phases that delay actual code output. The developer reports that enforcing these constraints explicitly in the system prompt produces sustained productivity gains across multi-hour sessions where Claude would otherwise drift toward narration over execution. Essentially: (Anthropic, Claude Code users) are now contending with a behavioral degradation pattern that emerges from model design, not user error. - The CLAUDE.md file is public and minimal by design, intended to be dropped into any project without customization overhead. - The targeted behaviors — over-planning, over-explanation, loss of decisiveness — are consistent with known patterns in long-context transformer inference where the model increasingly hedges. - The developer positions this as a workaround to an architectural reality, not a feature request. This is an early signal that prompt-level behavioral governance is becoming a standard part of professional Claude Code workflows.