Anthropic’s Thariq posted on X a new set of context-engineering guidelines for Claude 5 models, arguing that the prompt itself is only one part of what Claude receives and that system prompts, Skills, CLAUDE.md files, memory, and references can matter just as much. The post also claims Anthropic was able to remove "over 80%" of Claude Code’s system prompt for models like Claude Opus 5 and Claude Fable 5 with "no measurable loss" on coding evaluations.
The company’s message is that older guardrails are increasingly unnecessary for the newest models. In the post, Anthropic describes a shift away from rigid instructions and toward broader guidance, coupled with better tool design and selective loading of context. It also points to claude doctor and the /doctor command in Claude Code as a way to rightsize Skills and CLAUDE.md files.
What Anthropic says changed
The post contrasts older advice with newer practice:
- Rules to judgment: Instead of strict do-not-do lists, Anthropic now recommends guidance such as: "Write code that reads like the surrounding code: match its comment density, naming, and idiom."
- Examples to interfaces: Rather than leaning on usage examples, the post suggests designing tools and files so Claude has expressive parameters to work with.
- Upfront instructions to progressive disclosure: Verification and code review now live in separate Skills that Claude can call when needed. Some tools are also described as deferred-loading, so their full definitions are fetched only when required.
- Repeated instructions to simpler tool descriptions: The post says earlier models needed repeat instructions in multiple places, but those repetitions can now be deleted in favor of clearer tool descriptions.
- Manual memory to auto-memory: The old habit of writing everything to
CLAUDE.mdis replaced by Claude’s automatic memory, according to the post. - Simple markdown specs to richer references: Anthropic now points to HTML artifacts, code, test suites, mockups, and rubrics as useful reference material.
A supporting mind-map attached to the thread summarizes those same themes under headings like “Let Claude Use Judgment,” “Use Progressive Disclosure,” “Keep Tool Descriptions Simple,” “Use Auto-Memory,” and “Use Rich References.”
How Anthropic says to organize context
The post breaks the setup into four parts:
System prompt
Anthropic says this should stay tightly tied to the product. For Claude Code, it argues, the system prompt likely should not be edited often.
CLAUDE.md
The guidance here is to keep it lightweight, document what the repo is for, and focus on codebase-specific gotchas rather than obvious facts.
Skills
Anthropic describes Skills as lightweight guides that help Claude find information when needed, while avoiding overconstraint except in important areas.
References
Files can be @ mentioned as references, with a preference for code-based material where possible. The post lists specs, mockups, code, HTML artifacts, and rubrics as examples.
Thariq later clarified in a reply that Anthropic now uses "different system prompts per model," a detail that helps explain how the company is handling compatibility across generations.
Source: X thread
