Rethinking Context Engineering for Claude 5: How Instruction Simplification Improves Results
Anthropic
Anthropic has revisited its approach to context engineering for new Claude models. Removing more than 80 percent of the system prompt from Claude Code for Claude Opus 5 and Claude Fable 5 did not degrade programming tests. Instead of rigid rules, the recommendation is to trust the model's judgment, use progressive context disclosure, and leverage rich references such as HTML artifacts and code.
Anthropic's developers have shared new context engineering guidelines for fifth-generation Claude models. The main change is the removal of excessive restrictions: for Claude Opus 5 and Claude Fable 5, over 80% of the Claude Code system prompt was eliminated without any loss of quality on programming benchmarks. Instead of rigid rules, models now rely on their own judgment, and examples are replaced with well-designed tool interfaces. It is recommended to use progressive context disclosure, where instructions and skills are loaded only when needed, rather than all at once. Memory is now automatic, rather than stored in CLAUDE.md. Instead of simple Markdown specifications, rich references should be used—HTML artifacts, code, evaluation rubrics. To optimize their context, users can run the /doctor command in Claude Code. This material was prepared by the 'AI for Devs' community.
Source: Habr — хаб ИИ —
original