tl;dr: Anthropic's newer Claude models require less prescriptive context engineering: instead of rigid rules, repeated instructions, and front-loaded examples, developers should rely on Claude's judgment, well-designed tool interfaces, and progressive disclosure (loading context only when needed). CLAUDE.md files should be lightweight and focused on non-obvious gotchas, while skills should encode team-specific opinions and reference richer artifacts like code, test suites, or HTML mockups rather than simple markdown specs. A new `claude doctor` command helps simplify existing prompts.