The distinction between Claude bound to a project with governing instructions and a fresh chat session explains most reported inconsistencies. What users perceive as personality shifts or safety regressions are actually context window mechanics — when long conversations compress, the model reasons from a different internal state. This isn't a model failure; it's how transformer architecture handles token limits.

The practical fix is a handoff prompt: a structured summary passed forward before session degradation that re-establishes project context, intent, and the framing earlier prompts depended on. Without this pattern, any LLM becomes unreliable across extended interactions. The technique applies universally across providers, not just Anthropic.

Project-level instructions set at the system level produce far more consistent behavior than standalone conversation chains. The documented case — identical prompts yielding different results in new sessions — is precisely what project memory and governing instructions are designed to prevent. Teams treating prompting as a disciplined workflow rather than ad-hoc chat see dramatically better reliability.

The ceiling isn't Claude getting worse — it's users hitting the limits of chat-first mental models. The same rigor applied to evaluating model outputs should apply to prompting habits. When the toolset is used as designed, most complaints dissolve into operator error.