The new rules of context engineering for Claude 5 generation models
5 hours ago
- Removed over 80% of Claude Code's system prompt for advanced models (Claude Opus 5, Claude Fable 5) without loss in coding evaluations.
- Context engineering involves assembling system prompt, skills, CLAUDE.md, memory, and other sources to impact results.
- Overconstraining Claude with conflicting instructions is harmful; let Claude use judgement instead of rigid rules.
- Replace examples with well-designed interfaces (e.g., enums for tool parameters) to avoid constraining exploration.
- Use progressive disclosure: load context only when needed, not all upfront.
- Simplify tool descriptions and avoid repeating instructions across system prompt and tool descriptions.
- Auto-memory: Claude automatically saves relevant memories instead of manual CLAUDE.md updates.
- Use rich references like code, HTML artifacts, test suites, and rubrics, not just simple markdown specs.
- Apply to own context: keep system prompt for product, CLAUDE.md lightweight with gotchas, skills as lightweight guides, use @-mentions for references, and simplify with 'claude doctor'.