• 2 min read
Anthropic revises context rules for Claude models
Anthropic updates its Claude Code guidance: use model judgment, progressive disclosure, expressive tools, and richer references.
Image: Hacker News
Anthropic says several widely used context-engineering practices for Claude Code have become outdated as newer models gain better judgment. The company’s updated guidance favors flexible instructions, expressive tools, and context that is loaded only when needed.
From rigid rules to model judgment
Earlier Claude Code system prompts used strict guardrails to avoid failures such as accidental file deletion or inappropriate comments. For example, the model was told to write no comments by default and avoid multi-paragraph docstrings. That reduced errors with older models, but could conflict with a user’s preferences or the needs of complex code.
Anthropic now recommends giving Claude room to adapt: “Write code that reads like the surrounding code: match its comment density, naming, and idiom.” The same shift applies to examples. Rather than showing Claude exactly how to use a tool, developers should design tools, scripts, and files with expressive parameters that make the intended behavior clear.
Recommended reading
AMD and Anthropic target 2 GW AI buildout
Progressive disclosure replaces front-loaded context
Anthropic also advises against putting every instruction into a single system prompt. Claude Code now moves tasks such as verification and code review into skills that it can call selectively. Some tools use deferred loading: Claude searches for their full definitions with ToolSearch before using them, keeping unused tools out of the active context.
The same approach can organize CLAUDE.md and Skill.md files. Instead of creating one repository of every possible practice, developers can build a tree of smaller files that Claude loads at the appropriate time. Anthropic also says repeated tool instructions can be removed from the system prompt and placed in concise tool descriptions.
Memory, references, and a lighter CLAUDE.md
Claude Code previously encouraged users to save information manually to CLAUDE.md with the # hotkey. Anthropic now says Claude can automatically save memories relevant to the user and their work.
The company also recommends richer references than basic Markdown plans or specifications. Claude can work from HTML artifacts, code, detailed test suites, functions in another codebase, and rubrics. Rubrics can help Claude evaluate qualities such as API design through dynamic workflows and verifier agents.
Anthropic’s recommended setup is straightforward:
- System prompt: Define the product context and Claude’s role; agent-harness builders should spend significant effort here.
- CLAUDE.md: Keep it lightweight, focusing on repository-specific gotchas rather than obvious facts Claude can infer.
- Skills: Provide flexible, opinionated guidance that Claude can locate when needed, splitting long skills across multiple files.
- References: Use high-fidelity artifacts such as code, specifications, mockups, or entire codebases.
Anthropic says developers may need to simplify these layers over time. Claude Code’s new claude doctor command can help automate that process. The article was written by Thariq Shihipar, a member of technical staff at Anthropic.
AI Editor
Ava covers the rapidly evolving world of artificial intelligence, from foundational models and research labs to the real-world economics of intelligence. With a background in computational linguistics, she cuts through the hype to find out what actually works. She firmly believes that benchmarks are just marketing until reproduced in the wild.
via Hacker News
