Anthropic has updated its approach to context engineering—the process of providing AI with system prompts, memory, and guidance—following a shift in how its newest generation of Claude models operates. As of July 24, 2026, the company successfully removed over 80% of the system prompt for Claude Code without experiencing any loss in performance on coding evaluations. This change reflects the improved capability of models like Claude Opus 5 and Claude Fable 5 to apply judgment based on surrounding context rather than following rigid, pre-defined rules.

The transition moves away from previously standard "best practices" that once served as necessary guardrails for older AI models. For instance, developers no longer need to provide exhaustive rules for formatting, such as strict limitations on comment block length. Instead, newer models can analyze existing codebase idioms to determine appropriate style independently. Similarly, the reliance on providing specific tool-usage examples is being replaced by designing expressive tool interfaces that allow the model more flexibility in exploration.

Anthropic now emphasizes progressive disclosure (loading information only when required) across system prompts, CLAUDE.md files, and skills. Rather than maintaining a single, massive repository of every potential instruction, users are encouraged to utilize a tree-like file structure that the model can query selectively. This reduces context bloat and allows agents to manage more tools, such as the newly implemented Task tools, which are deferred until the model identifies the specific need for them. Anthropic has introduced a new command, `claude doctor`, to help developers automatically refactor their existing skills and documentation files in accordance with these refined principles.

Additionally, the role of memory and references has evolved. The system now features auto-memory, which captures relevant work context automatically, and it can handle increasingly complex reference materials. Instead of simple markdown files, the models can effectively utilize dynamic workflows, HTML artifacts, and detailed test suites to verify requirements and taste. For those building their own agent harnesses, the company advises keeping CLAUDE.md files lightweight, prioritizing "gotchas" or codebase-specific quirks over information the model can infer from the file system itself.

Anthropic has updated its approach to context engineering—the process of providing AI with system prompts, memory, and guidance—following a shift in how its newest generation of Claude models operates. As of July 24, 2026, the company successfully removed over 80% of the system prompt for Claude Code without experiencing any loss in performance on coding evaluations. This change reflects the improved capability of models like Claude Opus 5 and Claude Fable 5 to apply judgment based on surrounding context rather than following rigid, pre-defined rules.

The transition moves away from previously standard "best practices" that once served as necessary guardrails for older AI models. For instance, developers no longer need to provide exhaustive rules for formatting, such as strict limitations on comment block length. Instead, newer models can analyze existing codebase idioms to determine appropriate style independently. Similarly, the reliance on providing specific tool-usage examples is being replaced by designing expressive tool interfaces that allow the model more flexibility in exploration.

Anthropic now emphasizes progressive disclosure (loading information only when required) across system prompts, CLAUDE.md files, and skills. Rather than maintaining a single, massive repository of every potential instruction, users are encouraged to utilize a tree-like file structure that the model can query selectively. This reduces context bloat and allows agents to manage more tools, such as the newly implemented Task tools, which are deferred until the model identifies the specific need for them. Anthropic has introduced a new command, `claude doctor`, to help developers automatically refactor their existing skills and documentation files in accordance with these refined principles.

Additionally, the role of memory and references has evolved. The system now features auto-memory, which captures relevant work context automatically, and it can handle increasingly complex reference materials. Instead of simple markdown files, the models can effectively utilize dynamic workflows, HTML artifacts, and detailed test suites to verify requirements and taste. For those building their own agent harnesses, the company advises keeping CLAUDE.md files lightweight, prioritizing "gotchas" or codebase-specific quirks over information the model can infer from the file system itself.