The new rules of context engineering for Claude 5 generation models

Anthropic removed over 80% of Claude Code's system prompt for newer Claude 5 models without any measurable performance loss, finding that the models work better with less explicit constraints. The company discovered that overly restrictive prompts and conflicting guidance actually force the models to work harder, and that newer Claude versions can rely on context and judgment instead of detailed instructions. These findings suggest that effective "context engineering" for advanced models requires less specification rather than more detailed rules.