Anthropic Overhauls Claude 5 Context Engineering

Anthropic recently stripped over 80% of the system prompt content from Claude Code for its fifth-generation models, including Claude Opus 5 and Fable 5, with no measurable drop in coding evaluations. The move signals a major shift in context engineering: from a heavy reliance on explicit rules to trusting the more sophisticated judgment of the newer models.

Why Drastically Cut System Prompts?

In a blog post by Thariq Shihipar at Anthropic, the team observed that Claude Code frequently became overconstrained due to conflicting instructions. For instance, the old system prompt sometimes said "leave documentation as appropriate," while skills or the CLAUDE.md file outright banned adding comments. "We saw several contradictory messages in a single request," Shihipar noted. For earlier models, rigid rules were needed to prevent catastrophic mistakes like file deletion. But the new models' improved judgment allows them to navigate nuance without strict constraints, deciding autonomously when to add comments or documentation. Consequently, 80% of the system prompt content was jettisoned with zero impact on internal coding benchmarks.

Shift from Myths to New Practices

Anthropic identified several old practices that are now completely inverted:

• Old: Give Claude rules. New: Let Claude use its judgment. Instead of banning comments, the prompt now says: "Write code that blends in: match the comment density, naming, and conventions."
• Old: Provide tool usage examples. New: Design tool interfaces with expressive parameters. Guide via status enumerations without explicit examples that restrict exploration.
• Old: Front-load all information. New: Use progressive disclosure. Verification steps or code review instructions are no longer included in full; they're split into skills that load only when needed. Some tool definitions even delay loading until absolutely required.
• Old: Repeat instructions everywhere. New: Describe each tool once, in its own place; no need to echo in the system prompt.
• Old: Manually store memory in CLAUDE.md. New: Claude automatically saves memories relevant to the task and user preferences.
• Old: Specify in simple markdown. New: Provide rich references such as HTML artifacts, test suites, or rubrics that check your design taste via dynamic verification agents.

These changes not only cut token usage but also speed up responses and reduce conflicting instructions.

Practical Guide for Developers

For developers building agents with Claude, here is a summary of advice from the Anthropic team:

• System prompt: Closely tied to product context. If you're building your own agent, invest heavily here, not in configuration files.
• CLAUDE.md: Keep it extremely concise. Only describe repo-specific quirks not evident from the file structure. Use progressive disclosure: if you have lengthy instructions (e.g., verification steps), create a separate skill and simply reference it from CLAUDE.md.
• Skills: Lightweight guidance carrying team opinions or best practices. For long skills, split them into multiple files so they can load incrementally.
• References: Include code files or HTML mockups instead of long descriptions. Rubrics or interactive artifacts can now exemplify the design taste you want.

Progressive disclosure prevents the context from being flooded with irrelevant information, letting the model focus on the task at hand. To streamline this, Anthropic released the /doctor command in Claude Code, which automatically evaluates and suggests simplifications for your prompts, skills, and CLAUDE.md.

Implications for Developers Worldwide

Developers worldwide can immediately adopt this approach. By simplifying prompts, embracing progressive disclosure, and relying on automatic memory, the development process becomes more efficient and yields more context-aware code. This means developers spend less time crafting rules and more time designing solutions. The /doctor command makes it easy to review existing configurations. The shift underscores that as AI models grow more sophisticated, they require fewer explicit instructions; trusting the model's judgment unlocks higher productivity.