Anthropic engineer Thariq said in a July 24 post that the company revisited the Claude Code system prompt for its latest Claude 5 models, Claude Opus 5 and Claude Fable 5, cut more than 80% of the content, and saw no measurable decline in coding evaluations.
In Anthropic’s blog post, "The new rules of context engineering for Claude 5 models," the company said the change altered how it now writes system prompts, skills, and CLAUDE.md for the new model generation.
Legacy guardrails were removed
The post says many of the restrictions previously embedded in the Claude Code system prompt were guardrails designed for older model behavior. With Claude 5, Anthropic said those rules were no longer necessary and instead created friction through conflicting instructions, overly specific rules that failed in exception cases, and long context that took up space the model could otherwise use for judgment.
One example involved code comments. The earlier guidance said the model should never write multi-paragraph documentation or multi-line comment blocks and should keep comments to one line at most. The new guidance is different: write like the surrounding code, and match its comment density, naming, and style. Anthropic said that approach gives the model more room to judge instead of boxing it in with a hard rule that does not hold across many situations.
Six shifts in context engineering for Claude 5
The company grouped its new context engineering approach into several principles. At the center is a simpler idea: use fewer hard rules and trust the model’s judgment more.
- First, loosen constraints so the model can decide from context instead of following explicit rules at every step.
- Second, replace large volumes of examples with expressive tool interfaces and parameters, so the parameters and enumerations themselves suggest proper use.
- Third, use progressive disclosure rather than stuffing all information into context at once. Details can be split into skills that are called only when needed, and tools can be defined through lazy loading.
- Fourth, keep a single source of instruction so the same guidance is not repeated in both the system prompt and tool descriptions.
- Fifth, use automatic memory so the model can capture relevant memory without needing explicit reminders.
- Sixth, rely on richer reference formats such as test suites, program specifications, scoring rubrics, or HTML prototypes instead of plain text spec files.
Practical guidance for CLAUDE.md and skills
For developers, the post offers several direct suggestions. CLAUDE.md should stay lightweight, briefly explain the repository’s purpose, and focus on project-specific pitfalls in the codebase structure. It should avoid restating obvious facts that the model can infer by exploring files on its own.
Anthropic also recommends splitting skills across multiple files and pairing them with progressive disclosure. Outside high-risk steps, developers should avoid over-constraining the model. Skills are presented as a better place to hold team-specific knowledge and preferences. The source article also notes that Chain News had previously summarized the use of skills and memory from the official Fable 5 prompt handbook.
Anthropic also introduced a /doctor command to help automatically adjust the length of skills and CLAUDE.md.
The approach is not framed as universal
Anthropic said the roughly 80% reduction applies to Claude Opus 5 and Claude Fable 5, while older models may still require clearer constraints.
The company also said the effectiveness of this approach depends on the quality of tool and interface design. Poorly designed interfaces cannot be fixed simply by cutting prompt text.