Anthropic rethinks context engineering for Claude 5: simplifying instructions improves results

Anthropic has discovered that over-constraining Claude with verbose system prompts, CLAUDE.md files, and skills often backfires, as conflicting instructions force the model to spend extra reasoning effort. By removing more than 80% of the Claude Code system prompt for Claude Opus 5 and Claude Fable 5, the company saw no measurable drop in programming test results. The new context engineering philosophy shifts from giving rigid rules to trusting the model's judgment, from providing examples to designing clean interfaces and tool schemas, and from placing all context upfront to using progressive disclosure via skills and lazy-loaded tools. Instead of repeating instructions in both system prompt and tool descriptions, Anthropic now puts simple descriptions solely in the tool definitions. Memory is no longer stored manually in CLAUDE.md but handled automatically by Claude. For specifications, the company advocates using rich references such as code, HTML artifacts, and evaluation rubrics rather than simple Markdown plans. Users are advised to keep CLAUDE.md focused on non-obvious codebase details, use skills as compact guides, and leverage progressive disclosure to load context only when needed. The new `claude doctor` command helps automate this optimization.