Anthropic trims Claude Code system prompt by 80% as context engineering rules shift

Anthropic has published updated context engineering guidelines for its Claude 5 generation models, including Claude Opus 5 and Claude Fable 5, marking a significant departure from previous development practices. The company revealed that it removed over 80% of the system prompt for Claude Code without experiencing any measurable loss in performance on coding evaluations. This reduction reflects the models' improved judgement capabilities, which have reduced the need for the strict guardrails and overconstraining instructions that were previously necessary to prevent errors.

The update advises developers to reduce overconstraining instructions, noting that newer models have better judgement and can handle decisions without explicit rules that were previously necessary to avoid worst-case scenarios. Anthropic recommends avoiding the provision of usage examples for tools, as this can constrain the model’s exploration space; instead, developers should focus on tool design and parameters. The guidelines emphasise the use of progressive disclosure for tools and skills, suggesting that context should be loaded only when needed rather than included in a central repository.

Specific best practices include keeping CLAUDE.md files lightweight, using HTML artifacts or code-based references such as test suites over markdown files for specs, and utilising deferred loading for tool definitions. A new command, `claude doctor`, has been introduced within Claude Code to help users automatically "rightsize" their skills and CLAUDE.md files, assisting with the transition to these new standards.

Context engineering involves assembling context from system prompts, Skills, CLAUDE.md files, memory, and other sources, which impacts results across many requests rather than just single prompts. Previous versions of Claude Code relied heavily on CLAUDE.md for memory, information, and guidance, whereas newer versions utilise memory, artifacts, and skills for more dynamic context management. Earlier models required repeated instructions and strong guardrails, such as strict rules on comment density, to prevent errors, a practice that has become less necessary with the evolution of Claude’s capabilities.

The concept of "progressive disclosure" allows agents to search for full tool definitions using ToolSearch before use, enabling the inclusion of more tools without consuming excessive context. The article was authored by Thariq Shihipar, a member of technical staff at Anthropic, who outlined how these changes allow for more flexible and efficient agent harnesses. The shift signals a move away from treating context as a static repository towards a more dynamic, on-demand information retrieval system.