The new rules of context engineering for Claude 5 generation models

Anthropic said it removed over 80% of the system prompt for its newest generation of models, including Claude Opus 5 and Claude Fable 5, with no measurable loss on coding evaluations. The company detailed the shift in context engineering practices for Claude Code and agent development. The changes reflect a move away from overconstraining models with rigid rules and repetitive examples. Researchers found that older guardrails, such as explicit prohibitions on multi-line comments or mandatory planning documents, created conflicting instructions that required models to spend compute resolving overlaps. Newer models demonstrate improved judgement, allowing the system prompt to simply instruct the model to match the surrounding code's style and comment density. Anthropic also shifted from providing tool usage examples to designing more expressive tool interfaces and parameters. The company adopted progressive disclosure, moving detailed procedures like code review into separate skills that load only when needed. Additionally, automatic memory management replaces manual CLAUDE.md updates, and models can now reference rich artifacts such as HTML outputs instead of relying solely on markdown specification files. The best practices are available via the `/doctor` command in Claude Code.